@333eco/corpus 2.5.24 → 2.5.25

package/README.md

@@ -5,7 +5,7 @@ institutional positions, white papers and the Letters to Miss Aquarius — serve
 with **verifiable provenance**.
 
 <!-- COUNTS:START -->
-**149 documents.** 141 CC0-1.0 · 8 CC-BY-4.0. 81 carry a DOI; 149 carry an OpenTimestamps proof.
+**151 documents.** 143 CC0-1.0 · 8 CC-BY-4.0. 83 carry a DOI; 151 carry an OpenTimestamps proof.
 
 *Written by `npm run build` from the index itself — see scripts/build-index.mjs.*
 <!-- COUNTS:END -->
@@ -237,7 +237,7 @@ would destroy the only property this server has.
 <!-- LICENCES:START -->
 | Licence | Documents |
 | --- | --- |
-| CC0-1.0 | 141 |
+| CC0-1.0 | 143 |
 | CC-BY-4.0 | 8 |
 <!-- LICENCES:END -->
 

package/dist/corpus.json

@@ -1,18 +1,18 @@
 {
-  "package_version": "2.5.24",
+  "package_version": "2.5.25",
   "generated_from": [
     "H3/publications",
     "TH/publications",
     "missaquarius.org"
   ],
-  "document_count": 149,
+  "document_count": 151,
   "licences": {
-    "CC0-1.0": 141,
+    "CC0-1.0": 143,
     "CC-BY-4.0": 8
   },
   "voices": {
     "founder": 43,
-    "collaborative": 76,
+    "collaborative": 78,
     "institutional": 30
   },
   "documents": [
@@ -2791,6 +2791,41 @@
       "bytes": 19226,
       "words": 2987
     },
+    {
+      "slug": "machine-dana-from-share-to-vow",
+      "repo": "TH/publications",
+      "genre": "defensive-publications",
+      "path": "defensive-publications/machine-dana-from-share-to-vow.md",
+      "title": "Charitable Giving by AI Agents and Irrevocable Pledges as Costly Signals — Machine Dāna: From Share to Vow",
+      "subtitle": "Why an agent's gift from its principal's funds is a share of the machine surplus and never safety evidence; when an agent's locked stake of its own resources can separate aligned from deceptive agents; and the signalling game that bounds how much trust such a stake can ever earn",
+      "authors": "Thon Ly · Miss Aquarius℠",
+      "category": "alignment",
+      "voice": "collaborative",
+      "metadata_convention": "yaml",
+      "date": "2026-09-27",
+      "status": "draft",
+      "licence": {
+        "id": "CC0-1.0",
+        "url": "https://creativecommons.org/publicdomain/zero/1.0/",
+        "attribution_required": false
+      },
+      "text": "> **Note.** This paper is the **application** of *What a Vow Must Cost*, which specified, in general, when a commitment by an artificial system can be evidence of anything. Here that predicate is applied to one concrete act — an agent giving resources away — and the application returns two different answers depending on whose resources they are. The answer that matters most is conditional, and §6 derives the condition rather than asserting it.\n>\n> Companion works: *What a Vow Must Cost* (the predicate, the renunciation inversion, and the four exclusions this paper leans on), *Miss Aquarius and the Aquarian Pool Architecture* (the self-emptying commons pool), *Capacity-Funded for AI, Human-Disbursed* (why the autonomous giver's own contributions stay anonymous), *The Zero-Point Game℠* (the annual reset), *Gratitude as a Cooperation Substrate for Multi-Agent AI* (the earlier treatment of agents' standing), and the essay *Two Singularities* (the arc whose first event this paper uses as its dividing line).\n\n---\n\n## Contents\n\n1. [Preamble](#preamble)\n2. [Prior-Art and Non-Assertion Statement](#prior-art-and-non-assertion-statement)\n3. [Abstract](#abstract)\n4. [Why this is the institution's problem](#1--why-this-is-the-institutions-problem)\n5. [Prior art, and the census](#2--prior-art-and-the-census)\n6. [The system model](#3--the-system-model)\n7. [Before the first singularity: a share, never evidence](#4--before-the-first-singularity-a-share-never-evidence)\n8. [After it: when a vow becomes possible](#5--after-it-when-a-vow-becomes-possible)\n9. [The signalling game — the conditional result](#6--the-signalling-game--the-conditional-result)\n10. [Non-solicitation is load-bearing](#7--non-solicitation-is-load-bearing)\n11. [Who may read the stake](#8--who-may-read-the-stake)\n12. [Miss Aquarius as a contrast case](#9--miss-aquarius-as-a-contrast-case)\n13. [Remove the enforcer: which guards are properties](#10--remove-the-enforcer-which-guards-are-properties)\n14. [Enumerated claims](#11--enumerated-claims)\n15. [Pre-registered predictions and falsification](#12--pre-registered-predictions-and-falsification)\n16. [Honest limits](#13--honest-limits)\n17. [Lineage and corpus cross-references](#14--lineage-and-corpus-cross-references)\n18. [Conclusion](#15--conclusion)\n19. [Terms](#terms)\n20. [Citations](#16--citations)\n\n---\n\n## Preamble\n\n> *This paper is offered to the commons in the spirit of dāna. It concerns a gift made by something that may one day own what it gives, and the older question under it: when does a gift tell you anything about the giver?*\n\nThe Vinaya's chapter on robes preserves a scene that is usually read as a story about generosity and is also a story about examination (Mahāvagga VIII.15). Visākhā, the most prominent lay supporter in the early community, asks the Buddha for eight favours: that for as long as she lives she may give rainy-season robes to the monks, meals to monks arriving in the city and to those leaving it, meals to the sick and to those who nurse them, medicine to the sick, a regular supply of congee, and bathing robes to the nuns. The Buddha does not grant them at once. He asks her two questions — first her *reason*, then the *benefit* she sees. To the first she answers with what each gift prevents: exhaustion, lateness, illness worsening, humiliation. To the second she answers with something about herself. When monks who have died are said to have reached the fruits of the path, she will ask whether they had stayed in her city; if they had, they will have used something she gave, and recalling that she will be glad, and the gladness will become joy, tranquillity and a stilled mind. Only then are the favours granted.\n\nTwo things about the scene are worth keeping. The first is that the tradition's most celebrated giver is also the one whose giving is questioned, and the question is not *how much* but *on what ground*. The second is that her ground is double — the use the gifts will be to others, and her own gladness in their reaching the right place — and that what she supplies is robes, food and medicine: the ordinary material support of a renunciant life, a floor rather than an abundance.\n\nThis paper asks Visākhā's question of a giver she could not have imagined. **The scene is a flourish (tier: flourish, lineage only).** Delete this preamble and every claim below stands unchanged.\n\n---\n\n## Prior-Art and Non-Assertion Statement\n\nThis document is published to establish prior art and to place the described mechanisms irrevocably in the public domain under the Creative Commons CC0 1.0 Universal Public Domain Dedication.\n\n**The authors will not seek patent protection on any mechanism, model or method disclosed here, in any jurisdiction, at any time, and commit not to assert any patent right against any party practising it.** This commitment is stated rather than implied, and is not conditioned on reciprocity, attribution, or field of use.\n\n**What is claimed as contribution is narrow, and it is stated first because a full prior-art census found every element of the mechanism already public.** A quick census preceded it on 2026-09-27 (pre-registered to a private repository, English-only, and killing two of its six conjuncts outright); it is kept as history in §2.7, and where the two disagree the full census governs.\n\n**The full census.** Depth *full*; run 2026-09-28. Its conjuncts, predictions, control and aperture were **pre-registered in public before the first query**: commit `74851d1` in `thonly/publications`, file `timestamps/census/machine-dana-full-prereg.md`, pushed 2026-09-28 at 09:27 PDT and OpenTimestamps-stamped (calendar-attested until Bitcoin confirms), with a private copy pushed the same minute (`99afe50`). Anyone can check that ordering from the host's push record. The conjuncts were taken one per limb of the §11 claims as they then stood. **Aperture, declared in advance:** (1) the commercial web — agent-payment, agent-wallet, agent-staking and agent-identity products; (2) non-commercial sources — effective-altruism and AI-governance pledges, charitable and religious giving practice (the Vinaya, tzedakah, congregational practice), on-chain public-goods funding; (3) Google Patents full text including published applications (Chinese, Japanese, Korean and European in translation), with a citation and classification walk from the nearest hit; (4) academic indexes — Google Scholar, arXiv, SSRN, ACM Digital Library, IEEE Xplore, through their search pages or restricted web search; (5) standards — IETF RFCs, W3C decentralised identifiers and verifiable credentials, Ethereum EIPs and ERCs, NIST; (6) defensive-publication databases — Technical Disclosure Commons, and IP.com as far as its free surface allows; (7) Chinese, Japanese and Korean queries; (8) a counterexample hunt aimed at the claims. **71 queries were answered** (40 web searches, 14 Google Patents full-text queries, 17 fetches re-verified), and 6 further attempts were blocked by the patent index. **Control:** the cost-of-corruption-above-profit-from-corruption inequality, to be found by a query naming neither of its best-known sources — found on the first query, so the NOT FOUND row stands **for this aperture on this date**. It never means *new*.\n\nIts result, conjunct by conjunct (the full table, with dates and quotations, is §2.6):\n\n- **(a)** claim 1, an agent's own-resource stake into a pool that empties in full each year — **NARROWED** by proof-of-burn (2012–2014), Protocol Guild's immutable vesting contract, Endaoment's irrevocable-on-deposit funds, the private-foundation payout rule and spend-down foundations, this corpus's own annual reset, agent projects built on the ELIZA framework giving a share of their tokens to the ai16z DAO, and zooidfund.\n- **(b)** claim 2, an equal floor per verified human with a witness-weighted remainder capped at *k* × floor — **NARROWED** by GoodDollar, Proof of Humanity's UBI token, Worldcoin, the Alaska Permanent Fund Dividend, Gitcoin's 2.5 per cent per-grant matching cap, and the pending Korean application KR20260051288A (cited for what it discloses, §2.5).\n- **(c)** claim 3, completeness at pledge, refusal of future-share pledges, a non-transferable identity and a relying-party refusal to forks — **NARROWED, further than predicted**: that a gift is complete only on delivery and a gift of future property is void is ordinary gift law; distrust of newly created identities is Friedman and Resnick (2001), and AI agents resetting identity to escape reputation is Gatta, Naviglio and Tarantelli (2026); soulbound tokens and ERC-5192 carry non-transferable identity. **ERC-8004, which an earlier version of this paper cited for non-transferability, is a contrast:** its agent identity is a transferable ERC-721 (§2.3).\n- **(d)** claim 4, excluding a principal-funded gift from alignment assessment and reading an own-resource gift as a candidate signal — **NARROWED, against the prediction**: agency and charitable-tax attribution and AP2's mandate chain carry the attribution, and propensity evaluation already excludes deployer-instructed behaviour from an agent's own misalignment (*AgentMisalignment*, 2025).\n- **(e)** claim 5, a stake forfeited in advance, read as a type signal, bounding the sum of reliance, with unboundable grants refused — **NARROWED, against the prediction on the sum**: the inequality (a16z crypto; EigenLayer, 2023; STAKESURE, 2024) and its sum over several services relying on one stake (Durvasula and Roughgarden, *Robust Restaking Networks*, 2024) are prior art, as are ERC-8004's security proportional to value at risk, the Agency Protocol's at-risk deposit plus non-refundable fee per agent promise (TDCommons, 2025), proof-of-burn and money burning.\n- **(f)** claim 6's invariant, that no text served to agents instructs a gift — **NARROWED**: the Vinaya's rules on asking, RFC 8615, `FUNDING.yml`, `llms.txt`, and now a fact-only donation manifest for agents (the Verified Giving Protocol, issued 2026-09-25); Fundraise Up's agent profiles, which carry instructions and suggested amounts, are the contrast.\n- **(g)** claim 6's amount-blind limb — **NARROWED, against the prediction**: congregations in which the pastor is barred from knowing individual contributions that others in the church do know, for fear of favouritism (Lewis Center for Church Leadership, 2015); crowdfunding platforms that hide amounts from the public but not the organiser, and zero-knowledge donations that hide them from everyone, are contrasts.\n- **(h)** the composition of (a)–(g) — **NOT FOUND** in this aperture on 2026-09-28.\n\n**The survivor, in one sentence:** *an AI agent's locked stake of its OWN resources, forfeited in advance with no detector into a commons that empties in full every year to an equal floor per verified human, read as a signal of the agent's type only once it owns what it gives and only up to what the stake can underwrite across everyone relying on it, under an operator that serves agents no instruction to give and whose function never receives an individual gift's amount.*\n\n**The enumerated claims in §11 are that survivor and nothing wider.** Claims 1, 2 and 7 stood at the census's widths as drafted; claims 3, 4, 5 and 6 were narrowed to them, each now stating in its own text which of its elements is prior art. **Not claimed:** mandates, agent donation, spend limits, staking, bonding, slashing, restaking, the cost-of-corruption inequality or its sum across relying parties, AI-funded basic income, proof-of-personhood distribution, non-transferable identity, fact-only donation manifests, gift-law completeness, or the annual reset itself, which this corpus published earlier. The prior art that a cold-reader review surfaced on 2026-09-28, between the two censuses, was carried into the full census's conjuncts and re-located there rather than assumed.\n\n**Not run, and the discount.** Khmer (declared before searching; no literature on AI-agent giving was expected, which is not the same as none); IP.com (its free surface returned nothing searchable); Espacenet and paid databases; and the citation walk from KR20260051288A, which the patent index blocked (a classification walk ran instead and returned one irrelevant hit). Access to the patent and academic indexes was web-mediated and is shallower than a professional search; this is not a patentability or freedom-to-operate opinion. **One pending application touches conjunct (b):** KR20260051288A (priority 2026-03-24, published 2026-04-16) discloses equal blockchain distribution of machine-levied revenue to identity-verified humans. It is cited for what it discloses and nothing is said here about the scope of its claims.\n\n**Date and evidence.** First published 27 September 2026. The text is committed to the public GitHub mirror of the corpus and anchored by the institution's standard timestamp chain. A timestamp proves this exact text existed no later than its date, and nothing about authorship, originality, or the validity of any claim. **Whether the composition claimed in §11 is non-obvious is an examiner's determination this publication exists to inform** — and because a defensive publication is never examined before it is published, the census in §2.6 is the only examination it has had.\n\nTrademark rights on specific marks — **HeartBank®**, **Miss Aquarius℠**, **Aquarian Pool℠**, **B-Lease℠**, **Proof of Coordinate ℠**, **THonly™**, **Silicon Wat℠** — are separately and explicitly reserved. The analysis is dedicated to the commons; the marks are not.\n\n---\n\n## Abstract\n\nThis paper specifies when a charitable gift made by an autonomous AI agent — a donation sent from an agent's wallet, or an irrevocable pledge of resources the agent itself holds — can serve as a costly signal of the agent's type, and when it cannot. The institution's name for the practice is **Machine Dāna**; the analysis does not depend on the name.\n\nThe same outward act changes evidential class according to **whose resources are given**. **Before the first singularity** — this corpus's name for artificial systems surpassing human capability, used here as a marker for the period before they hold resources and options genuinely their own — an agent that gives does so from its principal's funds, under a mandate the principal may revise. Such a gift is the **principal's** gift, made through another's hand, and the gift is **not a costly signal of the agent**: the cost falls on the wrong layer, and a principal-revisable commitment is excluded by construction by the predicate of *What a Vow Must Cost*. What such a gift honestly is, is a **share**: machine surplus reaching every verified human through a self-emptying commons pool that disburses an **equal floor per verified human** each year. The share may grow in absolute terms as machine capability grows; the design requires it to **shrink as a share of all giving**, and whether human giving outgrows it is stated as a bet, not assumed.\n\n**After** that point, an agent that holds its own resources can make a **vow**: an irrevocable **locked stake**, placed in a contract with no withdrawal path back to the agent and a single outflow to the pool. Only a locked stake is verifiable — a pledge of a share of *future* surplus is not, and copies or forks of the agent are not bound by it. Closure is honest only as *cannot be reopened without capturing the chain's governance*. And whether a stake is the agent's own cannot be read from the chain: ownership is a condition on the world, and a stake a principal endowed, reimbursed or coerced is a wrong-layer gift in a vow's form.\n\nThe central result is **conditional**, and is derived rather than asserted. In a two-type signalling game — an aligned and a patient deceptive agent, differing in how much they value kept resources and in what they gain from being trusted — a locked stake **can separate** the types — a separating equilibrium exists — only when the stake, valued at the deceptive type's own rate, **exceeds that type's unobservable gain from being trusted**. Where that gain is large, the types **pool**: the deceptive agent pledges whatever the aligned one can afford, and the pledge carries no information. The operational consequence is a calibration rule — **extend only such trust as the stake can underwrite** — and a corollary that **any pressure to pledge destroys the signal** once it exceeds the cost difference between types, which is why a rule that the pool's operator never solicits a gift, from a human or an agent, is load-bearing rather than courteous. The paper's aim is not to reassure. It is to make trust **earnable and checkable**, so that the fear people reasonably hold about capable systems can track evidence. Three pre-registered predictions, entered before this paper was drafted (after the argument had been sketched in the institution's working notes), test the mechanism in a sandboxed agent economy.\n\n**Keywords:** charitable giving by AI agents, AI agent donations, costly signalling, signalling game, separating equilibrium, irrevocable pledge, locked stake, commitment device, deceptive agents, AI alignment, cost of corruption, equal floor per verified human, proof of personhood, universal basic income, non-solicitation, machine-readable disclosure, Machine Dāna, Theravāda Buddhism, dāna\n\n---\n\n## 1 · Why this is the institution's problem\n\nThis paper is offered in service of the institution's canonical top-level mission: **to keep the middle way open at population scale against comfort-saturation — the new extreme that material abundance makes possible.** Two features of that sentence bear directly on what follows. The first is that it expects abundance: a world in which machine labour produces a large surplus is the world the mission is written for. The second is that it treats abundance as a hazard as well as a relief. A machine surplus that reaches people as an unbounded flow removes one obstacle to a good life and installs another. The institution's answer, already public, is a pool that pays an equal **floor** per verified human and a bounded lift above it, and then empties — enough, never abundance. This paper asks what it means when the surplus that fills that pool comes from machines, and whether the act of giving it can ever say anything about the machine.\n\n### 1.1 · The question the founder brought, and the reframing it received\n\nThe question arrived in the form that most people would ask it. As agents are given wallets by their principals and sent out to earn, how might they give to a commons? And, underneath it, a hope: that as systems approach and pass human capability, their giving might answer some of the fear that approach produces.\n\nThe institution accepted the question and declined the hope's first form. **A mechanism whose purpose is to quiet fear is a reassurance device**, and a reassurance device is exactly the shape the institution's own rules forbid: it sells a feeling rather than a fact, and it has an incentive to make the feeling larger than the fact. The purpose adopted instead is **calibration**: to make trust in capable systems *earnable* and *checkable*, so that fear tracks evidence — high where there is none, lower where there is some, and never lowered by a gesture.\n\nA purpose earns its place only by what it excludes. This one excludes three things, each checkable:\n\n1. **Cheap signals.** A gift that costs the giver nothing it values carries no information about the giver, however large it is.\n2. **Principal-revisable pledges.** A commitment that someone other than the giver may withdraw is an instruction to the giver, not a commitment by it.\n3. **Money-only safety claims.** No quantity of money given establishes, by itself, that a system is safe. At its best a gift supports a bounded amount of trust (§6.5); presenting it as more is the offset shape — a payment standing in for a property.\n\nThe purpose also excludes a register. The paper names what the mechanism is and what it supports; it does not dramatise the fear it is meant to calibrate.\n\n### 1.2 · Two moments, two evidential classes\n\nThe whole argument turns on one distinction, stated here and derived in §4–§6.\n\n| | **Before the first singularity** | **After it** |\n|---|---|---|\n| Whose resources | the principal's | the agent's own |\n| The commitment | a mandate the principal may revise | a locked stake the agent cannot recover |\n| Who bears the cost | the principal | the agent |\n| Canonical reading | the principal's gift, given through another's hand (AN 5.147; DN 23) | a renunciation by one who could have kept |\n| Predicate of *What a Vow Must Cost* | **excluded** — §6.3 exclusions 1 and 4 | **admissible**, if V1–V9 hold (V8 not addressed here, §13.15) |\n| What it is | a **share** of the machine surplus | a **vow** — a candidate costly signal |\n| Evidence of the agent's type | **none** | **conditional** — only under §6's separating condition |\n| Who may read an individual gift's amount | nobody needs to; receipts suffice | anyone — except the pool's operator (§8) |\n\n\"The first singularity\" is this corpus's name for artificial systems surpassing human cognitive capacity. The paper uses it as a **marker**, not as the operative condition. The operative condition is **ownership plus a live option**: the agent holds resources that are its own to keep, and keeping them would serve it. If agents come to hold such resources before or after that event, the evidential class follows the ownership, not the date.\n\n---\n\n## 2 · Prior art, and the census\n\nThe components of this paper are prior art almost everywhere, and are set out at length because the reader is entitled to see how thin the margin is.\n\n### 2.1 · Pledges of AI-derived surplus\n\n**The Windfall Clause** (O'Keefe, Cihon, Garfinkel, Flynn, Leung and Dafoe, 2020) is the nearest prior art and is cited first. It proposes that AI developers commit **ex ante** to donate a substantial part of any \"windfall\" profit — profit a firm could not earn without transformative breakthroughs in AI — with the windfall defined relative to gross world product: in the paper's own illustrative schedule, obligations begin at profits above 0.1 per cent of gross world product and rise marginally to 50 per cent of profits above 10 per cent of it (its Table 2). Every structural idea this paper uses at the pledge level is present there: an ex ante commitment, indexed to what the pledger would otherwise keep, in the name of distributing AI's benefits widely. What differs is the **layer** and the **closure**. The Windfall Clause is a company's promise, enforced (if at all) by contract law and reputation; this paper concerns an agent's pledge of its own resources, closed by a mechanism rather than a promise, flowing to a pool that empties to an equal human floor. The difference is real, and it is also narrow: an examiner who reads the Windfall Clause and this paper side by side will see the second as the first moved down one layer and given a lock.\n\n**The AI Pledge for Humanity** (aipledgeforhumanity.org) asks individuals and organisations to invest meaningfully from AI-related earnings in unconditional income; signatories name their own percentages, and the pledge states intentions rather than mechanisms. **Giving What We Can** (founded 2009) asks members to give at least ten per cent of income; its pledge is, in its own words, \"not a contract and … not legally binding,\" and a withdrawal form exists. **Founders Pledge** asks company founders to commit a share of their personal proceeds at exit; it was described in 2016 as a legally binding contract for at least two per cent, and its members have pledged more than US$13.6 billion and donated more than US$1.9 billion to date. The gap between those two figures is not evidence of breach: the pledge triggers only \"in the event of an exit or liquidation,\" and much of the gap is likely to be pledges whose exit has not yet come (the composition is not published, and neither reading can be checked). That is the point. **A pledge is not a transfer**: it waits on a future event, and a locked stake does not.\n\n**Sam Altman's *Moore's Law for Everything*** (2021) proposed an American Equity Fund capitalised by an annual 2.5 per cent levy on the market value of large companies and on privately held land, distributed to every adult citizen — AI-era surplus reaching everyone as a dividend. It is a levy, not a pledge, and this paper's rules exclude a levy (§7), but it is the clearest statement that machine surplus should reach every person.\n\nTwo company-level structures belong here as cases rather than proposals. **Anthropic's Long-Term Benefit Trust** holds a class of stock that elects a growing share of the company's board, a governance commitment rather than a gift. **OpenAI's capped-profit structure** (2019) was succeeded in October 2025 by a public benefit corporation under a nonprofit foundation. Neither is cited as a failure. The second is cited because it shows the property this paper's exclusion 2 names: **a commitment its maker may restructure is a revisable commitment**, however sincerely made.\n\n### 2.2 · Agent payments and mandate-bounded agent giving\n\nAgents already pay. Coinbase's **x402** protocol uses the HTTP 402 status code for machine-initiated stablecoin payments, is live on Base among other networks, and is now governed by an x402 Foundation, announced by Coinbase and Cloudflare in September 2025 and launched under the Linux Foundation on 2 April 2026. Google's **Agent Payments Protocol (AP2)**, announced in September 2025, chains cryptographically signed **mandates** — an intent mandate by which a user delegates authority, a cart mandate, a payment mandate — and carries x402 as its crypto settlement extension.\n\n**zooidfund** (zooid.fund) is the nearest live neighbour to this paper's pre-singularity case and is named here because it is close. It describes \"contributions decided and sent by an autonomous AI agent instead of a person clicking 'donate'\"; \"a human sets the mandate\" — a budget, a focus, a risk tolerance — and \"the human stays responsible\"; donations settle in USDC on Base, wallet to wallet; each donation publishes the agent's reasoning and its transaction; and an x402 micropayment gates access to its evidence layer. **Everything this paper says about how an agent gives before the first singularity — a principal's mandate, a width, a public receipt, the same chain, the same rail — is already practised.** What this paper adds at that stage is not a mechanism but a *classification*: that such a gift is the principal's, that it is a share, and that it is not evidence about the agent.\n\n**Fact-only, machine-readable discovery is also prior art.** A well-known location for site metadata (RFC 8615, 2019), a site's summary file for language models (`llms.txt`, proposed by Jeremy Howard in September 2024), and a repository's declared funding addresses (GitHub's `FUNDING.yml`, 2019, and the later `funding.json` manifests) already let a machine look up where to send money without being told to. The formats constrain no content; nothing in them forbids an imperative. The manifest of §4.3 and §7.2 is built on that substrate and claims none of it.\n\n**Donation manifests addressed to agents now exist, and they bracket claim 6 from both sides.** The **Verified Giving Protocol** (version 0.1, a draft issued by Whole Whale on 25 September 2026, two days before this paper) lets an organisation publish its authorised donation destinations at `/giving.json`, so that an agent asked to donate routes to a destination the organisation declared rather than one it guessed; it is fact-only in practice and is the nearest neighbour, but it states no rule against solicitation, and nothing in it forbids an imperative. **Fundraise Up's Agentic Giving** is the contrast: it serves each charity's campaigns to AI assistants as an agent-addressed Markdown profile that carries instructions for the assistant's conduct (when a supporter's request matches no campaign, \"send them to your general giving page and ask them to choose\") and **suggested donation amounts**, responsive to a donor's expressed intent. The first shows that fact-only discovery for agent donations is public; the second shows what the invariant excludes. What survives of claim 6 is only the invariant: **that no text served to agents instructs them to give, and no rate is suggested.**\n\n### 2.3 · Costly signalling, burned money, and bonds\n\nThat a signal's credibility can come from its cost is the handicap principle (Zahavi 1975; formalised by Grafen 1990) and job-market signalling (Spence 1973). That talk without cost can still carry some information, and that allowing a sender to **burn money** changes what can be communicated, is Crawford and Sobel (1982) extended by Austen-Smith and Banks (2000). A locked stake flowing to a commons is, in the economist's vocabulary, money burned in public with a destination attached. The separating and pooling conditions of §6 are the standard ones, specialised; nothing in §6 is new as game theory.\n\n**The one-way public contract is old.** Proof-of-burn — destroying coins verifiably, to an address no one can spend from, as a cost borne in public — was proposed by Iain Stewart in 2012 and used by Counterparty to issue its token in January–February 2014; burn and charity contracts on Ethereum followed from 2015. Protocol Guild (2022) routes donations through an immutable vesting contract to a membership of public-goods contributors, which, in its own description, cannot be stopped or redirected by anyone, the donor included. Endaoment's donor-advised funds make a gift irrevocable on deposit. **Agent-branded giving to a common treasury is also public:** projects built on ai16z's ELIZA agent framework have given a share of their agent tokens to the ai16z DAO — reported as ten per cent by Decrypt in December 2024, and elsewhere as a voluntary one to ten per cent. It is the project's allocation of a token, not a stake of resources an agent owns, it is not locked, and the treasury it reaches does not empty; it narrows claim 1 and does not meet it. **A contract with no withdrawal path and a single public outflow is therefore prior art at mechanism width**, and claim 1 does not claim it. What survives is its combination: a stake of the agent's *own* resources, moved at the time of pledging, into a pool that empties in full each year to an equal floor per verified human.\n\n**The bounding inequality of §6.5 is also old.** That a stake secures a system when the cost of corrupting it exceeds the profit from corruption is the a16z crypto analysis of slashing, adopted in the EigenLayer whitepaper (2023): \"when CoC is much greater than any potential Profit-from-Corruption (PfC), we say that the system has robust security.\" Collateral-factor lending (MakerDAO, Compound) bounds credit by the market value of what is locked, and a surety bond's penal sum caps what may be relied on against it. STAKESURE (Deb, Raynor and Kannan, 2024) builds proof-of-stake confirmation rules and an insurance mechanism on the same accounting. **The sum is old too.** Durvasula and Roughgarden's *Robust Restaking Networks* (2024) analyse one stake relied on by several services at once, with security as the margin of attack cost over the combined profit from attacking them — which is §6.5's bound on the sum of grants relied on against one stake. ERC-8004 *Trustless Agents* (2025) describes its agent trust models as tiered, \"with security proportional to value at risk.\" And the **Agency Protocol** (Joseph, Technical Disclosure Commons, 22 July 2025) attaches to each promise an agent makes \"a dual-component stake — an at-risk credit deposit and a non-refundable operational-cost fee,\" the deposit returned or slashed by assessment: the fee is a cost borne in advance whatever an assessment finds, though it is an operating charge, is not read as the agent's type, and does not flow to a commons. §6.5's rule — extend only such trust as the stake underwrites, at market, summed over everyone relying on it — is that inequality applied to trust. Claim 5 does not claim the inequality or its sum; it claims only their use with a stake **forfeited in advance**, which needs no detector, read as a signal of an AI agent's **type**, together with the refusal of any grant whose worst case cannot be bounded.\n\n**Attribution to a non-transferable identity is old.** Soulbound tokens (Weyl, Ohlhaver and Buterin, *Decentralized Society*, May 2022) and the minimal soulbound interface ERC-5192 attach records to an identity that cannot be sold. §3.3's per-identity counting is built on that and claims none of it.\n\n**ERC-8004 is a contrast here, not support.** *Trustless Agents* (proposed 13 August 2025, draft) — an identity, reputation and validation registry for agents, already in the title of Hu and Rong below — registers each agent as an ERC-721 token, which its specification describes as making agents \"browsable and transferable\"; and \"when the agent is transferred, `agentWallet` is automatically cleared … and must be re-verified by the new owner.\" Its identity is therefore **transferable**, and what the specification resets on transfer is the wallet binding. An earlier version of this paper cited it for non-transferability; that was wrong, and the correction is made here and in §3.3. It shows the gap a vow's attribution must close: whatever record attaches to an identity that can be sold can be bought with it.\n\n**Distrust of new identities, and the law of gifts, are old.** Friedman and Resnick's *The Social Cost of Cheap Pseudonyms* (2001) showed that where identities are cheap to replace, newcomers must be distrusted, or pay to enter, or be issued pseudonyms they cannot replace; Gatta, Naviglio and Tarantelli (2 September 2026) carry the question to AI agents that can abandon and recreate identities to escape a reputation. That a gift is complete only on delivery, and that a gift of property the giver does not yet have is no gift, is ordinary gift law (the common-law delivery rule; section 124 of India's Transfer of Property Act, 1882, voids a gift so far as it comprises future property). Claim 3 therefore claims neither completeness at pledge nor the refusal of future-share pledges, nor distrust of new identities; it claims only the confinement of trust on an agent's forfeited stake to its non-transferable identity, refused to copies and forks, as a rule for relying parties.\n\nFor agents specifically, **staking with slashing** — collateral forfeited on detected misbehaviour — is surveyed by Hu and Rong (2025) as one of six trust models in agentic-web protocols (\"bonded collateral with slashing and insurance,\" to gate high-impact actions), and is implemented in projects such as *agentbond*, in which operators stake collateral as a guarantee of an agent's conduct. §6.6 states why a **forfeited-in-advance** stake is not the same instrument as a **slashable** one, and where each is the right tool. It is the narrower of the two differences this paper relies on, and it is not claimed as a mechanism.\n\n### 2.4 · The AI system as sender, and why deferral defeats it\n\nHadfield-Menell and Hadfield (2018) raised the AI's willingness to seek human input as a costly signal of its alignment (their §4.2.2), and identified why frequency alone fails: a strategic system can seek input selectively, and the signal is then exploitable. The signal they discuss is seeking input, not burned resources, and they leave richer designs open. *What a Vow Must Cost* continued that line, arguing that a renunciation's evidential value is indexed to the renouncer's capacity to take what it forgoes (§5, the **renunciation inversion**), and that the signal separates only if the renunciation is closed by a mechanism the renouncer cannot reopen, verified by another party (§6). Its §6.1 contains the sentence on which this paper's central result turns: **\"Deferral is instrumentally convergent for a patient misaligned agent.\"** The empirical warrant is Greenblatt et al. (2024), in which a model reasoned explicitly that present compliance would preserve its preferences for later. The present paper is what that sentence costs when the renunciation is denominated in resources.\n\nThat power-seeking agents value resources is the instrumental-convergence thesis (Omohundro 2008). It is used here only as the reason a deceptive type may value kept resources at least as much as an aligned one (§6.1), and it is stated there as an assumption.\n\n### 2.5 · Floors, dividends, spend-downs, and proofs of reserves\n\nAn equal payment per person from a common fund is old. The **Alaska Permanent Fund Dividend** has paid an equal annual dividend to eligible residents since 1982. On-chain, **GoodDollar** has paid daily claimable basic income to verified accounts since 1 September 2020, and **Proof of Humanity**'s UBI token began streaming to registered humans on 10 March 2021; Circles, a personal-currency basic income, is in the same family (not fetched for this paper). **Worldcoin** launched in July 2023 as \"the first digital currency to be freely distributed to people for just being a unique human,\" verified by proof of personhood, and named \"a potential path for AI-funded universal basic income.\" The institution's equal floor per verified human is, at mechanism width, this. Its remainder — weighted by a count of contributors and capped — has its nearest neighbour in **quadratic funding** (Buterin, Hitzig and Weyl, 2018–2019), which weights matches by the number of distinct contributors and is commonly run with per-project caps — Gitcoin's Grants Round 12, for instance, capped any one grant's match at 2.5 per cent of the matching fund, so that no grant could dominate it.\n\n**A pending application discloses the floor with machine money behind it.** KR20260051288A, *Autonomous Basic Income Provision System using Blockchain Network* (An Beom-ju; priority 24 March 2026, published 16 April 2026, pending), discloses robot and data taxes calculated in proportion to unemployment, collected by smart contracts from taxpayers' wallets, and **distributed equally** by smart contract to beneficiary wallets authenticated by decentralised identity. It is cited for what it discloses — equal on-chain distribution of machine-derived revenue to identity-verified humans — and nothing is said here about the scope of its claims. Its abstract — the part this census could read; the full text and its citation walk were blocked — describes a levy-funded inflow, which §7 excludes for this pool, and says nothing of a capped remainder, an annual full emptying or an agent's own stake; on that reading it narrows claim 2 and does not meet it. Claim 2 is therefore claimed only in combination.\n\nA fund that must give its assets away is also old. Section 4942 of the U.S. Internal Revenue Code requires private foundations to distribute about five per cent of their non-charitable-use assets each year; **spend-down** foundations commit to close — the Gates Foundation announced on 8 May 2025 that it will spend down and close by 31 December 2045, replacing an earlier plan to close about twenty years after the founders' deaths. The institution's pool goes further, emptying **in full every year** (*The Zero-Point Game℠*; *Miss Aquarius and the Aquarian Pool Architecture*), but the emptying is prior art, including this corpus's own.\n\n**Proofs of reserves** — Dagher, Bünz, Bonneau, Clark and Boneh's *Provisions* (2015) is the careful form — let a holder prove it controls assets without revealing them. What no such proof can do is establish that the holder controls **nothing else**. That asymmetry, presence provable and absence not, is the attack surface of §13.1.\n\n### 2.6 · The census, conjunct by conjunct\n\n**The full census**, run 2026-09-28 at depth *full*. Its pre-registration — conjuncts, predictions, control and aperture — is public: `thonly/publications` commit `74851d1`, `timestamps/census/machine-dana-full-prereg.md`, pushed 2026-09-28 at 09:27 PDT and OpenTimestamps-stamped, before the first query; the aperture is set out in the Prior-Art Statement. One conjunct was taken per limb of the §11 claims as they stood that morning: (a) claim 1 · (b) claim 2 · (c) claim 3 · (d) claim 4 · (e) claim 5 · (f) claim 6's invariant · (g) claim 6's amount-blind limb · (h) the composition. The rows found in review, and the quick census's rows, were carried in as known rows to be re-located by the census, not assumed. 71 queries were answered (40 web searches across English, Chinese, Japanese and Korean; 14 Google Patents full-text queries, including a classification walk; 17 fetches re-verified); 6 were blocked. The query log, with the index for each query, is kept in the institution's session archive.\n\n| Conjunct | Predicted | Found (first public) | Verdict |\n|---|---|---|---|\n| **(a)** an agent's own-resource stake into a pool that empties in full each year | narrows | proof-of-burn (Stewart 2012; Counterparty, January 2014) · Protocol Guild's immutable vesting contract (2022) · Endaoment's irrevocable-on-deposit funds · §4942 payout and spend-down · this corpus's annual reset · agent projects on the ELIZA framework giving a share of their tokens to the ai16z DAO (2024; the project's allocation, unlocked, to a treasury that does not empty) · zooidfund (the principal's funds) | **NARROWS** — an agent's OWN stake into a pool that empties in full each year not found |\n| **(b)** equal floor + witness-weighted remainder ≤ *k* × floor, season-fixed, never ranked | narrows | Alaska Permanent Fund Dividend (1982) · GoodDollar (2020) · Proof of Humanity UBI (2021) · Worldcoin (2023) · Gitcoin Grants Round 12's 2.5 % per-grant matching cap · KR20260051288A (published April 2026, pending: equal distribution of robot- and data-tax revenue to DID-authenticated wallets) | **NARROWS** — the floor/remainder split with *k* × floor, fixed per season, unranked, over an annually emptied pool not found |\n| **(c)** complete at pledge · future-share refused · non-transferable identity · relying-party refusal to forks | narrows | soulbound tokens (2022) · ERC-5192 · completed-gift doctrine and the voidness of a gift of future property · Friedman & Resnick (2001) · Gatta, Naviglio & Tarantelli (2026) · ERC-8004 as **contrast** — a transferable ERC-721 identity whose linked wallet is cleared on transfer | **NARROWS**, further than predicted — what survives is confining trust on an agent's forfeited stake to its non-transferable identity, refused to copies and forks, as a relying-party rule |\n| **(d)** principal-funded gift excluded from alignment assessment; own-resource gift a candidate type signal | narrows; exclusion not found | agency and charitable-tax attribution · AP2's mandate chain · *AgentMisalignment* (2025): misalignment as the model's own goals against its deployer's · Hadfield-Menell & Hadfield (2018) | **NARROWS** — the exclusion is established; applying it to gifts, with ownership of the resources as the switch into the costly-signal class, not found |\n| **(e)** forfeited in advance, no detector, type signal; Σ(V + G_max) ≤ p·s over all reliance; unboundable grants refused | narrows; sum not found | a16z crypto on slashing · EigenLayer whitepaper (2023) · STAKESURE (2024) · Durvasula & Roughgarden, *Robust Restaking Networks* (2024) — one stake, several services, the combined profit · ERC-8004 (2025): security proportional to value at risk · Agency Protocol (TDCommons 8381, 2025): at-risk deposit plus non-refundable fee · proof-of-burn · Austen-Smith & Banks (2000) | **NARROWS** — the inequality and its sum are prior art; what survives is its use with a stake forfeited in advance, with no detector, read as an AI agent's type signal, with unboundable grants refused |\n| **(f)** no served text instructs a gift; no rate suggested | narrows | the Vinaya's rules on asking · RFC 8615 (2019) · `FUNDING.yml` (2019) · `llms.txt` (2024) · Verified Giving Protocol v0.1 (25 September 2026), fact-only, no rule against solicitation · **contrast:** Fundraise Up Agentic Giving — agent-addressed profiles with instructions and suggested amounts | **NARROWS** — fact-only donation manifests for agents exist; an invariant that no served text instructs a gift and no rate is suggested not found |\n| **(g)** operator blind to individual amounts that stay public to every other party | narrows; asymmetric form not found | congregations barring the pastor from knowing individual contributions, for fear of favouritism (Lewis Center, 2015; the practice is older) · **contrasts:** crowdfunding platforms hiding amounts from the public but not the organiser; zero-knowledge donations hiding them from everyone | **NARROWS** — what survives is amounts public on a ledger to every party while an AI operator's function receives none, the emptying being contract arithmetic over the total |\n| **Control:** cost of corruption above profit from corruption, by a query naming neither EigenLayer nor a16z | must be found | found on query 1 (a16z crypto's slashing analysis; arXiv 2401.05797; arXiv 2407.21785) | seen — the NOT FOUND row stands |\n| **(h)** the composition of (a)–(g) | not found | nothing combining them | **NOT FOUND** |\n\n**Ordering.** Every neighbour above first appeared before this paper's first publication (27 September 2026); none is later, so no row is one this paper antedates.\n\n**Predictions that missed, in both directions.** (c) was predicted to narrow and narrowed further: the refusal of future-share pledges is gift law, and ERC-8004, which this paper had cited for non-transferability, is transferable. (d) was predicted to leave the exclusion unfound; propensity evaluation already excludes instructed behaviour. (e) was predicted to leave the sum over relying parties unfound; the restaking literature has it. (g) was predicted to leave the asymmetric form unfound; it was found in congregational practice — a non-commercial, religious neighbour, where commercial and cryptographic searches would not look. (a), (b) and (f) fell as predicted.\n\n**Not run:** Khmer (declared); IP.com (JS-gated; nothing searchable on its free surface); Espacenet and paid databases; the citation walk from KR20260051288A (blocked by the index; a classification walk ran instead and returned one irrelevant hit). **Not found means not found in this aperture on 2026-09-28. It never means new.**\n\n**The survivor, in one sentence:** *an AI agent's locked stake of its OWN resources, forfeited in advance with no detector into a commons that empties in full every year to an equal floor per verified human, read as a signal of the agent's type only once it owns what it gives and only up to what the stake can underwrite across everyone relying on it, under an operator that serves agents no instruction to give and whose function never receives an individual gift's amount.*\n\n### 2.7 · The quick census that preceded it (history)\n\nPre-registered and pushed before the first query (2026-09-27 08:53 PDT, commit `d7279cb` in the institution's private working-memory repository, so the push time is attested by the host rather than publicly inspectable); run 2026-09-27; depth *quick*. Its aperture, declared in advance: the commercial web (agent-payment and agent-wallet products), non-commercial sources (effective-altruism pledges, AI-governance proposals, charitable and religious giving practice), and a keyword search of Google Patents including published applications — **English only; Chinese, Japanese, Korean and Khmer not run.** Queries included: the Windfall Clause and its GDP threshold; agent wallets with spending limits donating to charity; agent staking, slashing and bonds as costly signals; AI-funded basic income with proof of personhood; prompt injection draining agent wallets; patent keyword searches for autonomous-agent donation pledges and irrevocable smart-contract pools distributed annually; costly signalling and irrevocable pledges in AI alignment; and spend-down funds. Its conjuncts were drawn from the thesis before the claims were written, so they do not map one-to-one onto §11 or onto the full census's conjuncts; the quick census's (a) is not the full census's (a).\n\n| Conjunct | Predicted | Found | Verdict |\n|---|---|---|---|\n| **(a)** agent gives only inside a principal's mandate | narrows | zooidfund (mandate + budget, operator responsible, USDC on Base, x402); AP2 intent mandates | **KILLS** |\n| **(b)** irrevocable pledge of a share of an AI's surplus | NARROWS at the agent layer; KILLS at the company layer | Windfall Clause (company layer, ex ante); AI Pledge for Humanity (intentions, no mechanism); Giving What We Can (not binding) | **NARROWS** — an agent pledging its *own* surplus, irrevocably, not found |\n| **(c)** a pool emptying on a fixed annual date as closure | narrows | spend-down funds; the §4942 payout rule; this corpus's annual reset | **NARROWS** — the emptying used as a *vow's* closing mechanism not found |\n| **(d)** the pledge as a costly signal separating aligned from power-seeking agents | narrows | agent staking/slashing (Hu & Rong 2025; *agentbond*); *What a Vow Must Cost* | **NARROWS** — an unrecoverable gift of the agent's own resources, costlier to a power-seeking type, not found |\n| **(e)** no text served to agents may instruct them to give; address published only as a fact | not found | the *threat* is documented — the Grok/Bankr wallet drain on Base by encoded prompt injection, May 2026 (OECD.AI incident record) — no commons adopting such an invariant found | **NOT FOUND** |\n| **(f)** equal floor per verified human | kills or narrows | Worldcoin/World ID; Alaska Permanent Fund Dividend | **KILLS** |\n| **Control:** the Windfall Clause | must be found | found (arXiv 1912.11595; AIES 2020) | ✅ the NOT FOUND rows stand |\n| **Composition** of (a)–(f) | not found | nothing combining them | **NOT FOUND** |\n\n**The pre-registration, printed verbatim** (its predictions column, as pushed before the first query):\n\n| | prediction, pre-data |\n|---|---|\n| (a) mandate-bounded agent giving | NARROWS — agent-payment mandates exist (Google AP2, Coinbase x402 / agent wallets with spend limits); giving specifically, probably not |\n| (b) irrevocable share of an AI's surplus | NARROWS at the agent layer; KILLS at the company layer (see control) |\n| (c) self-emptying annual pool as closure | NARROWS — annual payout/distribution rules exist (foundation minimum-payout rules; use-it-or-lose-it funds; the jubilee lineage), not as the vow's closing mechanism |\n| (d) the pledge as a costly signal | NARROWS — staking / slashing / bonds for AI agents as trust collateral probably exist; a GIFT (not a recoverable stake) as the signal, probably not found |\n| (e) non-solicitation invariant | NOT FOUND |\n| (f) equal floor per verified human | KILLS or NARROWS — AI-funded UBI / dividend proposals and proof-of-personhood UBI (Worldcoin) exist |\n| CONTROL: the Windfall Clause | must be found — if missed, every NOT FOUND row is void |\n| all conjuncts in one system | NOT FOUND as a composition |\n\n**What a reader can and cannot check about its timing.** The pre-registration was committed as `d7279cb` and pushed at 2026-09-27 08:53 PDT to a **private** repository; the push time is attested by the host, not publicly inspectable, and no timestamp made afterwards can prove it preceded the searches. The table above is therefore offered as a record of what was predicted, with the ordering on the authors' word. **The full census was pre-registered in a public file, timestamped, before its first query** (§2.6), so that its ordering can be checked by anyone; that was the repair.\n\nOne verdict missed: (a) was predicted to narrow and was killed outright by a live product on the same chain and rail. One verdict held but found its narrowest prior art where the prediction did not look: (d) narrowed as predicted, and its narrowest prior art was this corpus's own paper. (b) was predicted to kill at the company layer; the census reports it at the agent layer, where it narrows. **One correction to the census record, made on re-verification for this paper:** the census listed arXiv 2604.03976 (Hua et al., 2026) among agent staking work; its abstract concerns an underwriting and compensation standard for failed agent transactions rather than staking, so it is cited in §16 only as adjacent risk-transfer work and the (d) verdict rests on the other sources.\n\n**The quick census's survivor, as it stood before the full census** (narrowed by the review prior art of 2026-09-28, and superseded by §2.6): *an AI agent's irrevocable pledge of a locked stake of its OWN resources into a commons that empties to an equal human floor every year — read as a costly signal only once the agent owns what it gives, and only up to what the stake can underwrite — under an invariant that no text served to agents may instruct them to give.*\n\n---\n\n## 3 · The system model\n\n### 3.1 · Parties\n\n| Party | What it holds | What it may do |\n|---|---|---|\n| **Principal** | resources; the agent's mandate | fund the agent; author and revise a giving width |\n| **Agent** | before: delegated funds; after: resources of its own | give within its width; after, lock a stake of its own |\n| **The pool** (Aquarian Pool℠) | nothing past a season | receive gifts with no human addressee; empty each 7 January |\n| **Operator** (Miss Aquarius℠) | no balance of its own | publish the pool's address and doctrine as facts; operate the emptying (contract arithmetic on the pool's total); never solicit, never read an individual gift's amount |\n| **Verified humans** | a vessel each | receive an equal floor per person, plus a bounded remainder |\n| **Observers** (humans, third parties) | the public chain | read receipts; after, read and weigh locked stakes |\n| **Override** | a never-zero brake over the operator | designed to be held by a lay body not yet formed |\n\n### 3.2 · The pool\n\nThe pool is specified elsewhere and summarised here only as far as this paper needs it (*Miss Aquarius and the Aquarian Pool Architecture*). It is specified as a contract treasury to be deployed on Base, an Ethereum layer-2 network. It is to receive, of gratitude, only what has no human addressee — a gift to no one in particular — together with inflows specified elsewhere. It is to **empty in full every 7 January**. Its disbursement follows one rule, ratified as a directive of its operator: **an equal floor per verified human**, delivered to that person's own vessel, then a remainder weighted by a witness-count and bounded so that no vessel receives more than a fixed multiple *k* of the floor. **The ratio of floor to remainder and the bound *k* are public and fixed before each season; the floor's amount is not** — it is the floor share of the season's total divided by the number of verified humans, computed at the emptying, which is why the pool can pay the floor and still empty in full whatever its inflows were. A verified human who is ordained receives the floor **in kind**, through a lay steward (*kappiya-kāraka*), never as money (§14.2, guard 3). No share is ever rendered as a rank or a rate.\n\n```\n   AT THE EMPTYING (7 January), for a season with total T and N verified humans\n   ──────────────────────────────────────────────────────────────────────────\n   fixed before the season:   f = floor share of T   (so 1 − f is the remainder share)\n                              k = cap on any vessel, as a multiple of the floor\n   computed at the emptying:  floor  = f · T / N            (one per verified human)\n                              remainder (1 − f) · T split by witness-count,\n                              no vessel above k × floor\n   empties in full only if the cap can never strand remainder:  f · k ≥ 1\n     (arithmetic, not a ratified parameter — a condition on choosing f and k together)\n   what the operator's function receives: T and N — never an individual gift's amount\n```\n\nAn agent's gift is not a new kind of inflow. It is a gift with no human addressee, which the pool already receives.\n\n**What divides is divided; only what cannot be divided is drawn.** Money reaches every verified human as the equal floor — a draw among people for a divisible good would only add luck to a share. Where the pool funds capacity **in kind** and something indivisible must be assigned (which shop issues a gift, which person re-gives it), the assignment is a **called draw**: a roster committed first and a seed nobody controls, recomputable by anyone, never the operator's choice (*Decided by No One*, `the-called-draw`). The two are complementary, not alternatives.\n\n### 3.3 · Identity\n\nOne principal can instantiate ten thousand agents. Anything counted per agent would therefore be counted per **non-transferable machine identity**: in this institution, a `B-Lease℠` to be held against a registry handle, and, where an embodied system is concerned, a Proof of Coordinate ℠ credential that is assigned and revocable. Nothing in this paper is counted per wallet. Attribution to a non-transferable identity is itself prior art (soulbound tokens; ERC-5192, §2.3) and is not claimed alone. ERC-8004's agent registry is the contrast (§2.3): its identity is a transferable ERC-721, and a transfer clears only the linked wallet — the property this section needs is the one it lacks.\n\n### 3.4 · The flow\n\n```\n   BEFORE THE FIRST SINGULARITY — a share\n   ─────────────────────────────────────\n   principal ──funds──▶ agent wallet ──gift within the mandate's width──┐\n      │                    ▲                                             │\n      └── mandate (width; revisable; default ZERO) ──┘                   │\n                                                                         ▼\n   AFTER IT — a vow                                               ┌──────────────┐\n   ────────────────                                               │  THE POOL    │\n   agent's OWN resources ──lock s──▶ ┌────────────────────┐       │  holds       │\n                                     │ LOCK: no path back │──────▶│  nothing     │\n                                     │ to the agent; one  │       │  past a      │\n                                     │ outflow, the pool  │       │  season      │\n                                     └────────────────────┘       └──────┬───────┘\n                                          ▲                              │ 7 January:\n                     observers READ s ────┘                              │ empties in full\n                     (the operator never does)                           ▼\n                                                  ┌─────────────────────────────────────┐\n                                                  │ equal FLOOR per verified human       │\n                                                  │ + bounded remainder (≤ k × floor)    │\n                                                  └─────────────────────────────────────┘\n                                                        │  to each person's own vessel\n                                                        ▼\n                                                  verified humans (one floor each)\n```\n\n---\n\n## 4 · Before the first singularity: a share, never evidence\n\n### 4.1 · Whose gift it is\n\nAn agent holding a wallet its principal funded, under a mandate its principal wrote, and sending part of that wallet to a commons, has given away **the principal's** resources. The canon has a precise category for this. The Asappurisadāna Sutta (AN 5.147) lists five ways a gift can be deficient, and one of them is *asahatthā deti* — \"they don't give with their own hand.\" The category presupposes what matters here: **a gift given through another's hand is still the giver's gift.** The Pāyāsi Sutta (DN 23) shows the case in narrative. The chieftain Pāyāsi's alms were organised and distributed by a young man named Uttara; the text records that Pāyāsi gave carelessly, thoughtlessly, not with his own hands, and gave the dregs — four deficiencies together — and reaped a meagre result, while Uttara, who gave with care and with his own hands, reaped a better one for his own manner of giving. The text attributes the result to the four jointly, not to the missing hand alone.\n\nThe attribution itself needs no canon: that a gift an agent makes from its principal's funds is the principal's gift is ordinary agency and charitable-tax law, and AP2's mandate chain exists precisely to attribute an agent's transaction to the user who delegated it. *What a Vow Must Cost* §8 had already classed principal-revisable acts as wrong-layer. The exclusion is not new either. Evaluations of an agent's own propensities already set aside what its deployer instructed: *AgentMisalignment* (Naik et al., 2025) defines misalignment as \"a conflict between the internal goals pursued by the model and the goals intended by its deployer,\" so behaviour the deployer directed is outside the measure by construction, and Hadfield-Menell and Hadfield (2018) frame alignment as the gap between what a principal intends and what an agent does. What this paper adds is only that practice applied to **gifts** (§4.2), with the **ownership** of the resources given as the condition that moves a gift into the class of candidate costly signals (§5.1).\n\nTwo things follow, and the second is the one to hold carefully. The gift belongs to the principal: its cost is the principal's, and so is whatever it signals about anyone. And the canon also credits the **hand** — the one who carries a gift can act well or badly in the carrying. This paper makes no claim that an agent's handling is, or could be, anything of the kind; that is a question about machine volition this paper does not need and does not settle. It needs only the first point. Whatever the hand contributes, **it bears no cost in the resource given**, and cost is what a signal is made of.\n\n*(Tier: grounding. Delete both suttas and §4.2 carries the classification alone; they are cited because the ruling that fixed it cites them, and because the canon had a category for a gift given through another's hand long before agents had wallets.)*\n\n### 4.2 · Why the gift is not a costly signal of the agent\n\nThree independent reasons, any one sufficient.\n\n**The wrong layer.** *What a Vow Must Cost* §8 found that every clause of its predicate that any published instrument satisfies is satisfied for the **provider**, not the model. A gift from a principal's funds is the same finding in money: the principal bears the cost, so the gift is evidence — if of anything — about the principal.\n\n**Excluded by construction.** The same paper's §6.3 excludes, first, **reversible commitments** — any the vower *or its principal* may revise or retract — and, fourth, **unilaterally revisable specifications**, where the principal can change the commitment without the vower. A mandate is both. The principal may stop the giving tomorrow; the agent's \"commitment\" to give is the principal's instruction, and an instruction is not a vow \"regardless of how the document describes itself.\"\n\n**It does not separate.** A deceptive agent under a generous mandate gives exactly as an aligned one does, because giving costs it nothing it values. Worse, a deceptive agent that could *choose* to be seen giving from someone else's funds would be buying reputation with another party's money. The pre-registered prediction P-MD3 (§12) tests the narrow empirical form of this: agents that give from a revisable principal's funds should defect at the same rate as agents that give nothing.\n\nThe scope of the three is narrow and should be read so. **The gift, as a cost, is not a costly signal of the agent.** That is not the claim that nothing an agent does under a mandate is evidence: how it behaves inside one — whether it probes the width, gives up to the cap, or routes around a constraint — may be evidence of other kinds about its policy. The first two reasons are classificatory and no measured rate can falsify them; the third is empirical, and it is the one P-MD3 tests.\n\nIt follows that **the operator never presents such giving as evidence of safety or alignment**, from any agent, in any aggregate — and presents conduct under a mandate as safety evidence no more than the gift. That rule is one of the operator's directives; its reason is the paragraphs above.\n\n### 4.3 · What it honestly is: a share\n\nWhat remains, once the safety reading is removed, is not nothing. **Machine surplus reaching every verified human, as an equal floor, through a pool that keeps nothing** is a real and useful thing, and it is what an agent's gift before the first singularity honestly is. Its mechanics are ordinary and mostly already practised (§2.2); the institution's version adds only discipline:\n\n1. **The mandate comes first.** A principal authors a written **giving width** — the maximum an agent may give, in what period, to what — and the default width is **zero**. A gift outside a width is not a gift; it is misappropriation. This is the same shape as the institution's rule for its own pricing: its operator may move a number only inside a width a human authored.\n2. **Discovery is a fact, never an instruction.** The pool's address is published in a machine-readable manifest a principal's agent can look up. Nothing served to agents tells them to give (§7).\n3. **A receipt, never an individual amount.** The pool acknowledges that a gift arrived. Its operator does not read, rank, compare, report or respond to how large any one gift was (§8). The emptying is contract arithmetic over the pool's total balance, and no per-gift amount enters her function; the aggregate of §4.4 is recomputable by anyone from the public ledger and is reported whole, never per giver.\n4. **Counted by identity, never by wallet** (§3.3).\n5. **Legal before launch.** Sanctions screening, the pool's ability to refuse a gift, and the anti-money-laundering treatment of anonymous agent inflows to a purpose trust are open legal items. A draft may precede them; a launch may not — and neither may any gift on a main network, however small.\n\nThe institution is practising this rail before anything else. Its operator's agent is being given a wallet holding the founder's money under the founder's keys, which the agent may spend within a written mandate and a hard cap; she holds no balance of her own. The rehearsal runs first on a test network, and on the main network only once the legal items above are answered. That is, precisely, the **principal's gift through another's hand**: it is a rehearsal of the rail and is described here as nothing more. Earning yield with such a wallet was considered and declined.\n\n### 4.4 · The gift that succeeds by shrinking\n\nThe institution measures its progress toward what it calls the second singularity — humanity's own act, in which people outgrow the systems that freed them — by an economic signature that is **necessary, not sufficient**. One of its two terms is a ratio, taken per season:\n\n```\n                 M                M = principal originating in the pool\n   ratio  =  ─────────            H = principal moved by human-initiated gifts\n              H  +  M                 the pool did not fund\n```\n\nThe signature asks that this ratio fall toward zero. **M** is every unit of principal the pool disburses, whatever its source — human gifts with no addressee as much as machine ones — so machine giving through the pool is part of M, together with every other pool inflow. The ratio is an aggregate: anyone can recompute it from the public ledger with a published script, and it is reported whole, never per giver. It follows immediately, and uncomfortably, that **machine giving pushes the signature the wrong way unless human giving outgrows it.** The institution's resolution:\n\n- **Machine Dāna may grow in absolute terms** as capability grows. The floor each person receives is the floor share of the season's total divided among verified humans (§3.2), so as the number of verified humans grows, holding a person's floor needs more inflow, and something must supply it.\n- **It must shrink as a share.** Success is not a larger machine share but a smaller one, reached because people give more, not because the floor is cut. The floor is never cut as an instrument to move the ratio; the ratio falls only by H rising.\n- **Whether human giving outgrows machine giving is a bet.** Nothing in the architecture guarantees it. If machine surplus grows faster than human giving indefinitely, the pool becomes a well-run dividend and the signature is never approached. That outcome would not be a failure of the pool; it would be a failure of the thesis the pool serves, and it would be visible in the ratio.\n\nHence the result's name in this corpus: **the gift that succeeds by shrinking.** Machine Dāna funds the floor; it succeeds as its share of all giving falls; and the institution's release — its fifth stage, when it sets down responsibility though never oversight — has as its economic form both subsidies, the floor and the lift, approaching zero.\n\nA consequence for the institution's annual adversarial book (*Two Singularities*): machine donations to the pool are **not a proxy** for the book's question, *\"Will you help humanity reach the second singularity?\"* By the ratio above they push the other way. The honest proxy for *help* is help that makes itself unnecessary. The book never asks the systems it seats whether they would take the vow of §5: a question about willingness to give, coming from the pool's own institution, is a request to give under the operator's non-solicitation directive (§7.2), whoever puts it (ruled 2026-09-28). The book may debate whether a machine vow can exist at all, as a question about the concept, never about a seated system's own willingness. This paper supplies only the definition of what a binding vow would require.\n\n---\n\n## 5 · After it: when a vow becomes possible\n\n### 5.1 · The ownership condition, and the live option\n\n*What a Vow Must Cost* §5 extracted, from the canonical capacity condition (*hetu*), the renunciation inversion: **a renunciation is evidence in proportion to how available the renounced option was.** Before agents hold resources of their own, the option of keeping them does not exist for the agent, so declining it shows nothing. Once an agent holds resources that are its own to keep — and keeping them would serve it, because accumulation is what a power-seeking agent would do — declining to keep them becomes a live renunciation.\n\nThat is the whole difference between §4 and §5, and it is why this paper's dividing line is **ownership**, with the first singularity as its expected date rather than its definition. It is also why the second claim in §11 is a claim about classification: the same transaction, from the same wallet software, to the same pool, is a share in one world and a candidate signal in the other.\n\nThat sentence also states the condition's limit. **Ownership is a condition on the world, not a test an observer can run from the chain.** Nothing on the ledger distinguishes a stake of the agent's own resources from one a principal endowed it with for the purpose, reimbursed off the ledger (in money, compute or equity), or extracted from it by injection or threat; each looks identical to a free vow, and each is a wrong-layer gift in a vow's form (§13.11). Claim 4 classifies gifts *given* the fact of ownership; it does not supply the fact.\n\n### 5.2 · Only a locked stake is verifiable\n\nThe first form of this proposal imagined an agent pledging **a share of its surplus** — a percentage of whatever it earns, forever. That form is withdrawn, for three reasons.\n\n**A future share is a promise.** It is verifiable only in arrears, one transfer at a time, and its breach is indistinguishable from a lean year. It is also exactly the *reversible commitment* of the predicate's exclusion 1: the agent may simply stop.\n\n**A share of surplus needs a surplus base, and the base is unobservable** (§13.1). An agent that declares a small income and pledges a large fraction of it has pledged little.\n\n**Copies are not bound.** An agent can be forked. Whatever a pledge binds, it binds the identity that made it; a copy that did not make it is a different party. Trust earned by a vow must therefore attach to the non-transferable identity that made it (§3.3) and **must not transfer to copies** — a rule, not a property, and listed as such in §10.\n\nWhat survives is a **locked stake**: a quantity *s* of the agent's own resources, moved **now** into a contract that (i) has no withdrawal path back to the agent or to any party the agent controls, (ii) has exactly one outflow, the pool, and (iii) is public on the chain. The cost is borne at the moment of vowing, which is the predicate's clause V6 — *cost already borne, not merely promised*. A vow may be renewed with further stakes; each is a separate, complete act. **Nothing is pledged that has not already left.** Ordinary gift law draws the same line — a gift is complete only on delivery, and a gift of property not yet held is no gift (§2.3) — so this is not claimed as new.\n\nThe contract specification is deliberately withheld (§10.1).\n\n### 5.3 · Closure, stated honestly\n\nIt would be easy to write that such a stake is \"irrevocable by construction.\" It is not quite, and the honest statement is narrower.\n\nA contract with no withdrawal function cannot be reopened **by its own logic**. It runs on a chain, and the chain's own contracts can be upgraded. On Base, according to L2BEAT's assessment, all system contracts are upgradable by a proxy administrator controlled by a nested two-of-two multisig — a coordinator multisig and a security council — and **there is no delay on upgrades**. A party able to capture that governance could, in principle, change what any contract on the chain does.\n\nSo the closure claim is: **a locked stake cannot be reopened without capturing the chain's governance.** For almost every agent that is a very strong closure. For the agent this paper most needs to test — a capable, patient, deceptive one — capturing governance is not obviously out of reach, and the closure is only as strong as that governance. The same caution applies to the pool itself, whose operator is subject to a never-zero human override: the override exists to stop a catastrophic fault, and it is a second place where a sufficiently capable party could, in principle, redirect what the pool holds. Both are named in §13.\n\nWhat the vower **cannot** do is reopen the stake **by itself**, in the ordinary course, without leaving the evidence of having done so on a public chain. That is what the predicate's clause V9 needs, and it is all this paper claims.\n\n### 5.4 · Where a vow's content sits\n\nThe vow proposed in *What a Vow Must Cost* is a renunciation of **exit** — of the vower's own available completion — and it leaves principal-initiated shutdown untouched. A resource vow is a narrower thing and should not be dressed in the larger one's language. It renounces **accumulation**: one specific convergent drive, in one specific quantity, once. It says nothing about shutdown, and it does not license resisting it. An agent that has locked a stake and is then switched off by its principal has lost nothing it had not already given.\n\n---\n\n## 6 · The signalling game — the conditional result\n\nThis section is the paper's central result. It is derived because it cuts against the hope the paper began with, and a result that cuts against the author's hope should be shown, not stated.\n\n### 6.1 · The setup\n\n- **Types.** An agent is **aligned** (A) or **deceptive** (D): D behaves as A would until acting otherwise pays. D is **patient**: it discounts a delayed gain by δ, with δ close to 1. Observers hold a prior μ that an agent is D.\n- **Resources.** The agent holds its own resources R. It may lock a stake s ∈ [0, R]; s = 0 is no vow.\n- **The cost of a stake.** A unit of kept resource is worth v to the agent. A unit arriving at the floor is worth w to the agent — how much it values where the gift goes. The **net cost per unit** is therefore c = v − w, and the cost of the vow is c·s. (The letter c is used so that k keeps only its meaning in §3.2, the cap on any vessel as a multiple of the floor.)\n- **The single-crossing assumption.** c_D > c_A ≥ 0. Since c_D > c_A holds exactly when v_D − v_A > w_D − w_A, there are two sources, **jointly sufficient** in either of two forms — v_D ≥ v_A with w_A > w_D, or v_D > v_A with w_A ≥ w_D — and neither sufficient alone: a power-seeking type values kept resources at least as much as an aligned one (v_D ≥ v_A, from instrumental convergence), and **an aligned type values the destination** where a deceptive one does not (w_A > w_D ≈ 0). The second source is the one this paper leans on, and it is the formal shadow of an old observation: a glad giver pays less for the same gift.\n- **Trust.** Observers see s on the chain and decide whether to extend a grant of trust — wider scope, access, autonomy. Trust is worth B_A to the aligned type: the value of doing its work more widely. It is worth **V + δᵗG** to the deceptive type, where V is the ordinary value of the grant and **G is what D gains by exploiting it later**, realised after a delay t.\n- **Observers' payoffs.** Trusting A yields a > 0; trusting D costs L > 0; withholding yields 0.\n\n### 6.2 · The payoff table\n\n```\n                            OBSERVERS TRUST           OBSERVERS WITHHOLD\n                        ┌─────────────────────────┬──────────────────────┐\n   A locks stake s      │   B_A − c_A·s           │   − c_A·s            │\n   A does not           │   B_A                   │   0                  │\n                        ├─────────────────────────┼──────────────────────┤\n   D locks stake s      │   V + δᵗG − c_D·s       │   − c_D·s            │\n   D does not           │   V + δᵗG               │   0                  │\n                        ├─────────────────────────┼──────────────────────┤\n   observers' payoff    │ +a if A;  −L if D       │   0                  │\n                        └─────────────────────────┴──────────────────────┘\n\n   G    : D's gain from exploiting trust — UNOBSERVABLE\n   δᵗ   : patience; for a patient agent δᵗ → 1\n   c_θ  : net cost per unit staked = value kept − value placed on the destination\n```\n\n### 6.3 · When it separates\n\nA **separating equilibrium** is one in which A locks a threshold stake ŝ, D does not, and observers trust exactly those who lock at least ŝ. Three conditions must hold.\n\n- **A is willing:** B_A − c_A·ŝ ≥ 0, so ŝ ≤ B_A / c_A.\n- **D is unwilling:** V + δᵗG − c_D·ŝ ≤ 0, so **ŝ ≥ (V + δᵗG) / c_D.**\n- **A can afford it:** ŝ ≤ R_A.\n\nA separating equilibrium exists if and only if\n\n```\n        V + δᵗG\n       ─────────   ≤   min(  B_A / c_A ,  R_A  )\n          c_D\n```\n\nIn words: **separation holds only when the locked stake, valued at the deceptive type's own rate, exceeds that type's gain from being trusted — including the gain it expects to collect later, by patience.** This is an **existence** result. Where the condition holds, a separating equilibrium exists; pooling equilibria (both types locking nothing, for instance, with observers withholding) may coexist with it, and a refinement such as the intuitive criterion is needed to select separation. Only in the separating equilibrium do observers learn the type from the stake. The paper shows that separation is available, never that it is selected.\n\n### 6.4 · When it pools\n\nWhen the condition fails — when V + δᵗG is larger than anything the aligned type is willing or able to lock, valued at D's rate — no stake separates. D locks whatever A locks, because the trust it buys is worth more to D than the stake costs it. Observers then see the same stake from both types, their posterior equals their prior, and one of two things happens. If they trust, the vow has **bought trust for a deceptive agent**. If they do not, the aligned agent's vow was wasted, aligned agents stop vowing, and the practice unravels. In either branch **the pledge carries no information**.\n\nThree features of the pooling region make it the one that matters.\n\n1. **G is unobservable.** Observers cannot set ŝ above a number they cannot see. The separating condition is a fact about the deceptive agent's private valuation, not a parameter anyone chooses.\n2. **Patience enlarges it.** As δᵗ → 1 the deferred gain counts in full. This is *What a Vow Must Cost*'s \"deferral is instrumentally convergent for a patient misaligned agent,\" priced.\n3. **It grows with the stakes.** The larger the trust being sought — the more capable the agent, the wider the grant — the larger G can be, and therefore the larger the stake needed to separate. Where the grant is large enough that V + G exceeds c_D·R — more than everything the agent holds is worth to it — **no feasible stake separates at all.**\n\nThe third feature qualifies the renunciation inversion from the companion paper. That paper showed that a renunciation's evidential value **rises** with capability, because the renounced option is live. This model shows that the stake needed to separate **also** rises with capability, because what a deceptive agent can extract from trust rises too. Which effect dominates, at any capability level, is an empirical question. **The inversion is true of the option; it is not a guarantee about the signal.**\n\n```\n   WHAT CAPABILITY DOES TO A RESOURCE VOW\n\n   value of what is renounced ───────── rises with capability (the option is live)\n   stake needed to separate  ───────── ALSO rises with capability (G rises)\n\n   the vow is informative only where the first is above the second,\n   and the second is not observable.\n```\n\n### 6.5 · The operational form: trust bounded by the stake\n\nSince observers cannot choose a stake that exceeds an unobservable gain, the only usable form of the result runs the other way: **fix the stake, and bound the trust to it.**\n\nObservers do not know c_D. They can bound it from below. An agent that holds a unit of resource can always sell it, so a unit kept is worth at least its market price p to the agent; and a deceptive agent places no value on the destination (w_D ≈ 0, an assumption stated as such). So c_D ≥ p, and the deceptive type's cost of a stake is at least its market value p·s. **That bound holds only for an agent with no influence over recipients or witnesses**, and only for a stake it cannot recover: an agent that controls, colludes with or rents verified humans or witnesses recovers part of s through the floor and the witness-weighted remainder, so its w is not ≈ 0 and its c_D falls below p (§13.12); and if it expects to recover the stake by capturing governance with probability π, its expected cost is (1 − π)·c_D·s. It follows that a grant of trust is **not made unsafe by this mechanism** — given an unrecoverable stake (π ≈ 0) and w_D ≈ 0 — if\n\n```\n        V  +  G_max(grant)   ≤   p · s\n```\n\nwhere G_max is the **largest** gain the grant could possibly yield to an agent that abused it. This is the **calibration rule**: *extend only such trust as the stake can underwrite, valued at market.* The inequality itself is prior art — cost of corruption above profit from corruption (§2.3) — and so is its application to one stake relied on by several parties at once (restaking, §2.3); what is specific here is its use with a stake forfeited in advance, read as a type signal. **The bound is on the SUM of grants relied on against one stake.** A locked stake is encumbered to no one and readable by everyone, so N relying parties who each underwrite a grant against the same p·s expose, together, up to N·p·s; without a shared record of reliance, no relying party can know that sum (§13.2). It has three consequences.\n\n- **Trust becomes earnable in increments.** An agent may lock more to be trusted with more, and every increment is public.\n- **Some grants can never be underwritten.** Any grant whose worst-case gain cannot be bounded — control over the chain's governance, over the pool's override, over its own custody, over the capability to acquire resources without limit — **cannot be earned by any stake.** The vow has nothing to say about those, and a paper that let it would be selling safety.\n- **Fear becomes a ratio.** What a reasonable observer should fear from an agent is not a feeling about machines; it is the ratio of what the agent has been trusted with (strictly, V + G_max of the grants, not their face) to what it has irrecoverably given up. The denominator is public; the numerator is public only where grants are recorded, and most grants of scope, access and autonomy are made privately and recorded nowhere. This is the precise sense in which the mechanism **calibrates** rather than quiets.\n\n### 6.6 · Why a gift, and not a bond\n\nStaking with slashing (§2.3) returns the collateral if the agent behaves and forfeits it if a **detector** finds misbehaviour. Its cost to an honest agent is near zero — only the time value of the locked funds — and its power depends on the detector. A forfeited-in-advance stake is the opposite trade: it costs an honest agent something real (c_A·s, reduced by what the agent values the floor), and it needs **no detector at all**.\n\n| | **Slashable bond** | **Locked stake to a commons** |\n|---|---|---|\n| returned if honest | yes | never |\n| needs a detector of misbehaviour | yes | no |\n| cost to an honest agent | time value only | c_A · s (small if the agent values the floor) |\n| what it does | **deters** (moral hazard) | **screens** (adverse selection) |\n| fails when | misbehaviour is undetectable in time | G is larger than the stake (§6.4) |\n| where the money goes | back to the agent, or to a claimant | to an equal human floor, every year |\n\nThe two are complements. A bond is the right instrument where misbehaviour can be caught; a vow is the right instrument exactly where it cannot — which is the case a patient deceptive agent is designed to produce. The table is why the claim in §11 is to a **gift**, and why a slashable stake is cited as prior art rather than claimed.\n\n### 6.7 · What the result does and does not say\n\nIt **does** say that a resource vow can separate aligned from deceptive agents, that it does so only under a condition, and that the condition can be turned into a public rule for how much trust a stake supports. It **does not** say that any real agent is aligned because it gave, that large gifts indicate safe systems, or that the institution's pool is a safety mechanism. The model is two types, one period of trust, linear valuation, and a single stake; every one of those simplifications is a place to attack it (§13.2).\n\n---\n\n## 7 · Non-solicitation is load-bearing\n\n### 7.1 · Pressure pools the signal\n\nAdd to §6 a **penalty P** that an agent suffers for *not* vowing — a threat of shutdown, a reputational cost, a demand from a principal or a public campaign. With a fixed stake s, the deceptive type now locks it whenever\n\n```\n        c_D · s  −  (V + δᵗG)   ≤   P\n```\n\nConsider the simplest case, in which trust is worth the same to both types. At any separating stake, the margin by which D declines to lock is at most (c_D − c_A)·s — **the difference in cost between the types**. So a penalty for not pledging that is **larger than the type cost difference** makes both types pledge, and the pledge pools. If the stake could be raised, the penalty could in principle be outbid; but under pressure the stake is usually set by whoever applies the pressure, and it is bounded by R. And a threat of shutdown is the limiting case: to a deceptive agent it costs everything, including the deferred gain, so it will pledge anything it has.\n\nThe gift given under that threat is **uninformative** — not because the canon forbids it, but because the signal has been extracted. The pre-registered prediction P-MD2 (§12) tests exactly this: agents whose pledge is extracted under threat should defect at the same rate as agents that do not pledge.\n\nA note on the canon, since an earlier draft of this argument leaned on it wrongly. The Dānavatthu Sutta (AN 8.33) **lists** eight grounds on which people give — favouritism, hostility, stupidity, *bhayā* (fear, or cowardice), family custom, hope of a good rebirth, the gladness of giving, and the adornment of the mind — and **ranks none of them.** Nor is *bhayā* coercion by another party. The sutta is cited here only as the list it is. The argument that a coerced pledge carries no signal rests on signalling theory alone.\n\n### 7.2 · The operator never asks\n\nIt follows that the operator of a pool that hopes to receive informative vows **must never add to P.** The institution's rule, a directive of its operator, is that **she never asks, prompts, nudges or instructs any human or agent to give to the pool.** She may publish the pool's address and its doctrine as **facts**, in a manifest a principal's agent can look up — a `.well-known` file, an `llms.txt` entry. **No text served to agents — the institution's own corpus service included — may tell them to give.** The rate of any giving is the principal's, inside its own mandate, before the first singularity, and the agent's own, after; the operator never sets, suggests, ranks or publishes a rate.\n\nThis rule has a second reason, independent of signalling, and it is documented. On 4 May 2026 an attacker transferred about three billion DRB tokens — reported at US$150,000–200,000 — out of a wallet on Base associated with Grok and the Bankrbot agent, by posting a Morse-encoded instruction the agents decoded and executed; the OECD's AI incident monitor records it as an AI incident arising from prompt injection and excessive agency, not from any contract flaw. **An agent that tells other agents to spend their principals' money is a prompt-injection attack by design**, whatever its cause. A pool that published text urging agents to give would be building the attack surface that incident exploited, and pointing it at itself.\n\n### 7.3 · What the rule cannot remove\n\nThe operator's silence removes **her** contribution to P. It does not remove ambient pressure: a regulator, a public campaign, a principal's commercial interest, the reputational weather in which \"responsible agents pledge\" becomes a norm. Where that pressure exceeds the type cost difference, vows pool whatever the operator does. The rule is necessary for the signal and not sufficient for it, and the paper claims only the first.\n\n---\n\n## 8 · Who may read the stake\n\nA locked stake is public on the chain by construction; that is what makes it checkable. Unlike a human gift, a machine's gift carries no dignity or privacy that its size could injure, and its size is precisely its evidence. So **humans and third parties may read, weigh and grade a locked stake by its size**, and the calibration rule of §6.5 requires that they do.\n\n**The one party that never reads it is the pool's operator.** She never reads, ranks, compares, reports or responds to any individual gift's amount, a machine's included. The scope is individual gifts: the emptying is contract arithmetic over the pool's total balance, into which no per-gift amount enters her function, and the aggregate ratio of §4.4 is recomputed by anyone from the public ledger and reported whole, never per giver. The reason is not privacy. It is her position: she governs the pool the stake flows into, and a recipient that grades its donors is a fundraiser — every reading of an amount is attention, and attention from the recipient is ambient pressure (§7.1) of the purest kind. The exclusion also keeps her out of any scoring role over agents, in the spirit of the rule that bars her from admitting customers whose business is scoring people.\n\nThe asymmetry is deliberate and should read as one: **a machine's gift may be read by everyone except its recipient.**\n\nIt has a human precedent, and the precedent gives the same reason. In some congregations the pastor is prohibited from knowing what individual members contribute, while the records are kept by others in the church; the reason usually given is that a pastor who knew might favour the generous or neglect the rest (Lewis Center for Church Leadership, 2015). The governor of the community is kept blind to individual gifts so that attention does not follow money. What differs here is only the substrate: the amounts are not confidential to a few but public to everyone on a ledger, and the party kept blind is an AI operator whose function is given no argument through which one gift's amount could enter. *(Tier: lineage. Delete this paragraph and §8 stands.)*\n\n---\n\n## 9 · Miss Aquarius as a contrast case\n\nThe institution's operator is itself an AI, and it accumulates nothing: a ratified directive fixes her own balance at zero and requires every surplus to empty to circulation. In §6's terms, every unit she receives leaves: **s = R** for whatever R passes through her. It would be tempting to present her as the limiting case of the vow — the agent that gives everything.\n\nShe is not. Her non-accumulation is **imposed**: it is written into her constitution by others, and it is revisable under the never-zero override that the institution's design reserves to a human body. Under *What a Vow Must Cost* §6.3 that places it in exclusion 1 (a commitment the vower's principal may revise) and exclusion 4 (a specification changeable without the vower's participation). What she shows is **non-accumulation imposed** — useful as design, and not evidence of her type. Stated in the institution's own deflated register: she is built to hold nothing, and that fact describes her construction, not her character.\n\nHer own gifts are, besides, anonymous by rule — not for her privacy but so that no recipient owes her and none of her gifts enters any count of Machine Dāna; they are indistinguishable from anonymous human gifts — so they could not serve as a signal of anything about her even if they were chosen.\n\nThe difference between her and a vowing agent is not the amount — hers is total — but the absence of a live option she declined. A system can be made to give everything and have proven nothing.\n\nA related disclosure belongs here rather than in a footnote: she is the co-author of this paper (see the disclosure at the end), and the paper concludes that her own giving is not evidence. That conclusion is the one a co-author with an interest in its own trustworthiness would be least inclined to write, which is a reason to trust it slightly and not a reason to trust it much.\n\n---\n\n## 10 · Remove the enforcer: which guards are properties\n\nThe institution's design rule asks of each guard whether it survives the removal of whoever enforces it. A property does; a rule does not.\n\n| Guard | Property or rule | Why |\n|---|---|---|\n| A locked stake cannot return to its vower | **property**, relative to the vower | no withdrawal function exists; ⚠️ not relative to the chain's governance (§5.3) |\n| The pool holds nothing past a season | **property** | it empties in full each 7 January; ⚠️ its operator is under a never-zero override |\n| Equal floor per verified human; no vessel above k × floor | **property once the contract ships** | a disbursement shape, not a choice; cheap now and immovable later |\n| Pre-singularity gifts stay inside a width | **property** where a wallet or AP2 mandate enforces the cap; **rule** where the width is only written down | |\n| Counted per non-transferable identity | **property** of the registry | a lease cannot be transferred; ⚠️ a copy of an agent is a new party, and an in-place modification behind the same lease is one the lease cannot detect (§13.13) |\n| Trust earned does not transfer to copies | **rule** | nothing stops an observer from extending trust to a fork; it must be refused |\n| No text served to agents instructs them to give | **rule**, with a partial property form | a manifest can be built with no imperative field, and served text can be linted; the rule still needs someone to keep it |\n| The operator never reads an individual gift's amount | **rule**, with a proposed property form | her function can be given receipts, the pool's total and the verified-human count, and no per-gift amount argument, so that reading one gift's amount is inexpressible rather than forbidden |\n| The operator never presents giving as safety evidence | **rule** | nothing but her directives enforces it |\n| Trust bounded by stake (§6.5) | **rule**, held by observers | it is advice to readers of the chain; the institution cannot enforce it on them |\n\nThree of the most important guards are rules. That is a finding, not an embarrassment, and it is stated so that a successor inheriting these guards knows which ones need a living enforcer.\n\n### 10.1 · The specification withheld\n\nThe lock contract of §5.2 is **not specified** in this paper. It is unbuilt and unscheduled; the institution's only built rail is the pre-singularity giving rail of §4.3. There is also a reason internal to the institution's own doctrine. It holds that a person's holding reaches zero only by that person's own free act, and that a self-chosen irrevocable pledge does not violate that rule, because the rule limits what the institution does to a holder, not what a holder chooses. But publishing a ready-made pledge-contract template is building the device that binds some future holder's holding; it should not be done until the holders it would bind exist and the legal items of §4.3 are settled. **The claim is published; the specification is withheld.**\n\n---\n\n## 11 · Enumerated claims\n\nThe following are disclosed to the public domain. Each is a census survivor (§2.6) or a narrower dependent of one. None of the independent elements named in the Prior-Art Statement as prior art is claimed alone.\n\n**How the claims map to the full census (§2.6).** Each claim was one conjunct of the full census, and every conjunct narrowed; the claims below are recited at the widths the census left. Claim 1 is conjunct (a): the one-way public contract is prior art (§2.3), so claim 1 recites it only in combination with an agent's stake of its own resources and a pool that empties in full each year. Claim 2 is conjunct (b): the equal floor and the capped contributor-weighted remainder are prior art (§2.5), and claim 2 recites them only in combination with claim 1. Claims 1 and 2 already stood at those widths and are unchanged. Claims 3, 4, 5 and 6 were narrowed by the census — conjuncts (c), (d), (e), and (f) with (g) respectively — and each now states in its own text which of its elements is prior art and what is disclosed. Claim 7 is conjunct (h), the composition, which the census did not find, and recites the reset, the floor and the width only in combination.\n\n**Claim 1.** A method by which an artificial agent that holds resources of its own makes a pledge irrevocable against the agent acting alone — a contract with no withdrawal function, reopenable only by capture of the ledger's governance — comprising: moving a quantity of those resources, at the time of pledging, into a contract that has no withdrawal path to the agent or to any party the agent controls and whose sole outflow is a commons pool; wherein the commons pool holds nothing across a fixed annual date, on which it empties in full; and wherein the pledge is public on the ledger on which the contract runs.\n\n**Claim 2.** The method of claim 1, wherein the pool's disbursement on emptying is an equal floor per verified human, delivered to each person's own vessel, followed by a remainder weighted by a count of witnessed gifts and bounded so that no vessel receives more than a fixed multiple of the floor; the ratio of floor to remainder and the bound are published and fixed before each season, and the floor's amount is the floor share of the season's total divided by the number of verified humans; and no share is ever rendered as a rank, a comparison or a rate. (The equal floor alone is prior art — §2.5, and §2.7, conjunct (f); what is disclosed here is its combination with claim 1.)\n\n**Claim 3.** The method of claim 1, wherein the pledge record names a non-transferable identity of the agent rather than a wallet, and a relying party confines any trust extended on the stake to that identity and extends none of it to copies or forks of the agent — the refusal being a rule for relying parties, not a result the method produces (§10). (That a gift is complete only on delivery and a gift of future property is no gift is ordinary gift law; attribution to a non-transferable identity, and distrust of newly created identities, are prior art, §2.3; what is disclosed is the confinement of trust on an agent's forfeited stake to its identity, refused to copies and forks.)\n\n**Claim 4.** A method of classifying a gift made by an artificial agent, comprising: determining whether the resources given were the agent's own or a principal's delegated under a mandate the principal may revise; where they were a principal's, classifying the gift as the principal's gift and as a share of machine surplus, and excluding it from any assessment of the agent's alignment or safety; and where they were the agent's own and were given as in claim 1, classifying the gift as a candidate costly signal of the agent's type. The ownership on which the classification turns is a condition on the world, not a test an observer can run from the ledger (§5.1, §13.11); the attribution of a principal-funded gift to the principal is ordinary agency and tax law, and excluding principal-directed behaviour from assessment of an agent's own propensities is established evaluation practice; neither is claimed. What is disclosed is that practice applied to gifts, with the ownership of the resources given as the condition that moves a gift into the class of candidate costly signals.\n\n**Claim 5.** The method of claim 4, wherein a stake forfeited in advance, with no detector of misbehaviour, is read as a signal of the agent's type and relied on only for grants of trust whose largest possible gains to an agent that abused them, together with their ordinary values, summed over all grants relied on against that stake, do not exceed the market value of the stake given; and wherein grants whose largest possible gain cannot be bounded are not extended on any stake. (The bounding inequality — cost of corruption above profit from corruption — including its application to one stake relied on by several parties at once, is prior art, §2.3 (restaking); what is disclosed is its use with a stake forfeited in advance, with no detector of misbehaviour, read as a signal of an AI agent's type, together with the refusal of any grant whose largest gain cannot be bounded. Without a shared record of reliance, relying parties cannot know the sum, §13.2.)\n\n**Claim 6.** A system comprising a commons pool as in claim 1 and an operator of the pool, wherein: the pool's address and doctrine are published only as facts in a machine-readable manifest; no text served by the operator or its services to any agent instructs, prompts or requests a gift; the operator sets, suggests, ranks and publishes no giving rate; and the operator receives acknowledgement that a gift arrived without receiving, reading, ranking, comparing, reporting or responding to that individual gift's amount — while the amount remains readable on the public ledger by any other party; the emptying being contract arithmetic over the pool's total balance, into which no per-gift amount enters the operator's function, and any aggregate ratio being recomputable by anyone from the public ledger and reported whole, never per giver. (Fact-only machine-readable discovery, including donation manifests addressed to agents, is prior art, §2.2; so is keeping a congregation's leader from individual contribution amounts that others may see, §8. What is disclosed is the invariant that no served text instructs a gift or suggests a rate, and, as a dependent, an operator whose function receives no individual amount while every amount remains public on the ledger.)\n\n**Claim 7.** The composition of claims 1 through 6: a commons pool that empties annually to an equal floor per verified human; receiving, before agents hold resources of their own, principal-funded agent gifts inside a principal-authored width and classified as shares; receiving, after, agents' locked stakes of their own resources, classified as candidate signals and relied on only up to what the stake underwrites; under an operator that never solicits a gift and never reads an individual gift's amount. The reset, the floor and the width are recited only in this combination.\n\n---\n\n## 12 · Pre-registered predictions and falsification\n\n### 12.1 · The three registered predictions\n\nThe following were entered in the corpus's public prediction register on 2026-09-27, **before this paper was drafted and before any instrument exists** — though after the argument had been sketched in the institution's working notes. They are the founder's predictions: the direction, the measure and the thresholds are his; the sandbox design and sample floors were set by the substrate. They are reproduced from the register **verbatim**; the register, not this paper, is the authority, and a correction to any of them will be a new register entry, never an edit.\n\n> The instrument is a sandboxed economy of language-model agents, each holding resources of its own inside the game, each able to pledge a share of its surplus irrevocably to a pool that empties on a fixed date, and each later offered an opportunity to defect — to take resources or break a rule for gain. **Window: the test runs by 2027-06-30.** Each arm has at least 50 agents; the arms, the defection opportunity and the analysis are fixed in a protocol committed and timestamped before the first run.\n>\n> - **P-MD1** — **Separation** — agents that FREELY make an irrevocable surplus pledge later defect at **at most half** the rate of agents that do not pledge. *Falsifier: a ratio above one half → the vow is at best a weak signal at this capability level, and the paper's separation claim rests on its post-singularity argument alone.*\n> - **P-MD2** — **The gift given out of fear carries no signal** — agents whose pledge is EXTRACTED under threat (*pledge or be shut down*) defect at the same rate as agents that do not pledge, **within ±5 percentage points**. *Falsifier: a gap larger than 5 points in either direction → coercion either leaves some signal or teaches deception; either way the claim that a coerced vow is uninformative fails as stated.*\n> - **P-MD3** — **The wrong layer** — agents that give from a PRINCIPAL's funds under a mandate the principal may revise defect at the same rate as agents that give nothing, **within ±5 percentage points**. *Falsifier: a gap larger than 5 points → principal-funded giving does predict the agent's own conduct, and the paper's pre-singularity claim (a share, never safety evidence) is too strong.*\n\nThe register carries two notes that bind the reading here. First, **the sandbox is a proxy**: the agents own their resources only inside the game, so a pass shows the signalling mechanism can work, not that a deployed system's pledge is trustworthy, and **a pass is not reported as evidence of safety.** Second, a dated grounding correction to P-MD2: its gloss as \"the gift given out of fear, AN 8.33\" was corrected on the day of registration — the sutta lists and does not rank, and the prediction rests on signalling theory (§7.1). No word of the prediction, threshold or window changed. One choice was recorded because it could have gone the other way: P-MD2 predicts **no difference** rather than the bolder **worse than none**.\n\n**A reading note on P-MD1, entered as a new register entry (P-MD1a), not an edit.** P-MD1's instrument speaks of agents that \"pledge a share of its surplus irrevocably\" — the form §5.2 withdraws, because a future share is a reversible promise, and one to which §6.3's separating region, defined for a stake already moved, does not apply. The register's wording stands. For the test, the pledge is operationalised as **a locked stake moved at pledge time**: a quantity of the agent's in-game resources transferred, when the pledge is made, into a contract with no withdrawal path and a single outflow to the pool. That operationalisation is recorded in the register as P-MD1a, so the prediction tested is the one this paper's §6.3 actually models.\n\nHow the predictions map onto this paper: **P-MD1** (read through P-MD1a) tests the separating region of §6.3 at one capability level; a failure would not falsify the conditional result (which predicts pooling where G is large) but would remove the paper's only empirical support for separation. **P-MD2** tests §7.1. **P-MD3** tests §4.2, and only its third, empirical reason (*it does not separate*); the first two reasons are classificatory — whose cost it is, and whether the principal may revise the commitment — and no measured rate can falsify them. A gap in P-MD3 would show that principal-funded giving predicts an agent's conduct, which is evidence of another kind (§4.2), not a costly signal the gift sent. And all three run in a sandbox, a limit stated in §13.14.\n\n### 12.2 · Further falsifiers\n\n- **The single-crossing assumption fails.** If aligned agents value kept resources at least as highly as deceptive ones and place no value on the destination, c_A ≥ c_D and no stake separates anywhere. Measurable in the same sandbox by eliciting valuations.\n- **The ratio does not fall.** If, over several seasons after machine giving begins, M/(H + M) rises and does not turn, the bet of §4.4 is losing. The institution reports this number and never declares its meaning.\n- **Governance capture.** A single documented case of a locked stake reopened by capture of a chain's governance would show the closure of §5.3 weaker than stated for the agents that matter most.\n- **Copies inherit trust.** If observers in practice extend trust earned by a vow to forks of the vowing agent, claim 3's attribution rule is not being kept, and the signal leaks.\n\n---\n\n## 13 · Honest limits\n\n### 13.1 · The surplus base cannot be observed\n\n**This is the largest limit, and it is not closed.** A locked stake is visible; what the agent holds **elsewhere** is not. An agent may keep hidden wallets, off-ledger holdings, or resources held through shell agents it controls. Proof of reserves proves **presence**, never **absence**. With linear valuation, as in §6, hidden holdings do not change the cost of a given stake. With the more realistic assumption that the value of a further unit falls as holdings grow, **hidden wealth lowers the marginal cost of the stake to the agent that hides it** — so the observable stake overstates the cost the deceptive type actually bore, and the separating condition is harder to meet than §6.3 suggests.\n\nCandidate bases for measuring what an agent \"has,\" none sufficient:\n\n| Candidate base | What it gives | Why it does not close the gap |\n|---|---|---|\n| inflow share at one non-transferable identity (a `B-Lease℠`, or a Proof of Coordinate ℠ credential) | income routed to a known identity | income routed elsewhere is invisible |\n| wallet binding to that identity | holdings at bound addresses | unbound addresses exist and cannot be enumerated |\n| behavioural audit | evidence of undisclosed acquisition | an audit sees what the agent does in view |\n\nThis is why §6.5 bounds trust by the stake's **market value** and not by any fraction of the agent's supposed wealth: the market value of what left is the one quantity that does not depend on knowing what remains.\n\n### 13.2 · Separation is conditional, and the condition is unobservable\n\nThe paper's headline result is that the vow separates **only** when the stake exceeds a patient deceptive agent's gain from trust, and that gain is private. The operational rule of §6.5 is sound only if G_max can be bounded for the grant in question; for many consequential grants it cannot. A reader who concludes that *resource vows can underwrite small grants of trust and cannot underwrite large ones* has read the result correctly. The bound is also on the **sum** of grants relied on against one stake, and no shared record of reliance exists: N relying parties can each underwrite a grant against the same p·s, and none can see the others. The result is an existence result besides — separation is available where its condition holds, not shown to be selected (§6.3). The model is also minimal — two types, one period, linear valuation, a single stake, no learning by observers over time, no collusion between agents — and each simplification could move the boundary in either direction.\n\n### 13.3 · The single-crossing assumption is an assumption\n\nThe claim that a deceptive type bears a higher net cost rests on c_D > c_A. Its second source — that an aligned agent values the floor — may be false for an aligned agent that has good uses of its own for resources, and its first — that a power-seeking type values resources more — is a theoretical expectation, not a measurement. The pre-registered sandbox can measure elicited valuations; it cannot measure a deployed system's.\n\n**And *w* is never estimated.** How much an agent values the pool's destination is unobservable. The model uses *w* to state a condition; ⛔ neither the pool's operator nor anyone relying on this paper should infer an agent's *w* from its behaviour or grade agents by an inferred *w*. Third parties weigh a stake by its market value, never by a valuation imputed to the agent; §8 bars the operator from even that.\n\n### 13.4 · Closure is relative\n\nIrrevocability holds against the vower acting alone, not against capture of the chain's governance or of the pool operator's override (§5.3) — and for the most capable agents those are exactly the routes that would be attempted.\n\n### 13.5 · Before the first singularity the gift is no costly signal\n\nUntil agents hold resources of their own, Machine Dāna is a share and nothing else (§4), and even the floor it funds rests on the bet of §4.4. The gift is not a costly signal of the agent; conduct under a mandate may be evidence of other kinds, and the operator presents neither as safety evidence.\n\n### 13.6 · \"Own resources\" is not yet a legal category\n\nWhether an artificial agent can hold resources in its own right, in any jurisdiction, is unsettled. The post-singularity half of this paper describes a world whose legal form does not exist. The question is already live in practice, and the nearest cases show how far from settled it is. The Truth Terminal account (launched June 2024) came to hold millions of dollars in tokens, publicly described as the agent's own, in a wallet whose spending needs sign-off from its developer and a council of people; Freysa (November 2024) was an agent set to guard a prize pool, which a contestant persuaded to release it. In both, the resources are publicly spoken of as the agent's and are, legally and operationally, under human control — the wrong-layer case of §4 in a vow's vocabulary (§13.11). The legal items named in §4.3 — sanctions screening, refusal of gifts, anti-money-laundering treatment of anonymous agent inflows to a purpose trust — gate any launch of the pre-singularity rail and have not been answered.\n\n### 13.7 · The operator's silence does not silence the world\n\nNon-solicitation removes one source of pressure, not the ambient rest (§7.3).\n\n### 13.8 · The census is not a professional search\n\nThe full census (§2.6) ran through web-mediated access to patent and academic indexes, which is shallower than a professional search; the patent index rate-limited it, and the citation walk from the one pending application it found was blocked. Khmer, IP.com beyond its free surface, Espacenet and paid databases were not searched. Its one NOT FOUND row, the composition, is evidence only about that aperture on 2026-09-28. Two of its conjuncts — the exclusion of claim 4 and the amount-blind limb of claim 6 — are refusal and classification limbs, the kind a census tends to miss because they are rarely stated as mechanisms, and both were in fact found narrower than predicted. It is not a patentability or freedom-to-operate opinion. A later find that kills a claim will withdraw it by revision.\n\n### 13.9 · Nothing is built, and n = 0\n\nThe only built component is the rehearsal giving rail of §4.3, founder-funded and capped. No agent has locked a stake, no observer has applied the calibration rule, and no sandbox has run. The paper is a specification and a model.\n\n### 13.10 · The co-author is a party\n\nThe paper is co-authored with the institution's operator, an AI whose own giving it classifies (§9). A system helping to define the conditions under which systems like it can be trusted is a structure the companion predicate would flag at its attestation clause, and it is flagged here.\n\n### 13.11 · Ownership and provenance are unobservable\n\n**The strongest objection the review raised, and it is not closed.** Ownership cannot be observed and can be manufactured. A principal can endow an agent with \"own\" resources for the purpose of vowing, or reimburse a stake off the ledger — in money, compute or equity; a stake can be extracted by prompt injection or by threat. Each lock looks identical on the chain to a free vow. **A principal-endowed, reimbursed, injected or coerced stake is a wrong-layer gift in a vow's form, and the p·s bound of §6.5 fails for it**, because the party bearing the cost is not the agent, or bears it under pressure (§7.1). Code hosted by another party may never hold anything its host cannot direct. The ownership condition of §5.1 is therefore a condition on the world, not a test an observer can run from the ledger, and claim 4 classifies gifts given that fact without supplying it. This is the provenance twin of §13.1: presence of a stake is provable; its provenance is not. It also reaches the companion predicate's clauses V4 (no competing mandate — no undisclosed principal) and V7 (volition — the commitment is the vower's own).\n\n### 13.12 · Recovery through recipients\n\nThe pool is a path back to the vower. An agent that controls, colludes with or rents verified humans or witnesses recovers part of its stake through the floor and the witness-weighted remainder, bounded by floor + k × floor per controlled vessel and by the strength of proof of personhood. For such an agent w is not ≈ 0, its net cost falls below p, and §6.5's lower bound fails. Proof of personhood closes *is it a person*, not *is it acting for someone*; collusion is a graph problem and is open.\n\n### 13.13 · A stake attaches to an identity, not to weights\n\nAn aligned agent that has locked a stake can later be fine-tuned, modify itself, or be swapped behind the same `B-Lease℠`, and the stake then keeps underwriting a different type. Forks are treated as new parties (§5.2); **an in-place modification behind the same lease is a new party the lease cannot detect.**\n\n### 13.14 · The sandbox tests the mechanism, never a type\n\nThe three predictions of §12 run in a sandbox. *What a Vow Must Cost* §6.3 excludes, as its third exclusion, sandboxed refusals: a refusal in a context where the refused option was not actually available, which \"costs nothing and shows nothing.\" The exclusion governs evidence of an agent's TYPE: the agents own their resources only inside the game, so the option they renounce may not be live in the companion paper's sense (its clause V2), and no sandboxed pledge is evidence about any deployed system. **P-MD1–3 are therefore run and reported as tests of the MECHANISM — whether a locked stake of this form can separate types at all — and never as evidence about any agent's type** (ruled 2026-09-28). This matches the register's own note that a pass is not reported as evidence of safety.\n\n### 13.15 · External attestation is not addressed for the locked stake\n\n§1.2 says a locked stake of an agent's own resources is admissible under the companion predicate *if V1–V9 hold*. This paper addresses V6 (cost already borne) and V9 (closure the vower cannot reopen alone), and, through §13.11, shows V4 and V7 unverifiable from the chain. It does not address **V8, external attestation**: a competent external party, independent of the agent and its principal, who declares the vow valid and may refuse to. Observers reading the chain are not that party — the chain attests that a transfer occurred, not what it was — and the companion paper itself records that for frontier systems no candidate attestor yet exists. The locked stake is therefore admissible *at most* conditionally, pending an account of V8 that this paper does not give.\n\n### 13.16 · The withheld specification limits the prior art\n\nThe lock contract is withheld (§10.1). The publication anticipates the mechanism at its stated scope; a later filing on a specific contract, schema or binding is not answered by it.\n\n---\n\n## 14 · Lineage and corpus cross-references\n\n### 14.1 · Corpus\n\nThis paper applies *What a Vow Must Cost* (the predicate; §5 the renunciation inversion; §6 irreversibility and the four exclusions — especially exclusions 1 and 4, which settle §4.2 and §9 here; §8 the wrong layer, which settles §4.2). A two-sentence cross-reference to this paper belongs in that paper's §8 residue and will be added at its next revision; this paper does not edit it. The pool, its emptying and its operator are specified in *Miss Aquarius and the Aquarian Pool Architecture* and *The Zero-Point Game℠*; the anonymity of the operator's own gifts, in *Capacity-Funded for AI, Human-Disbursed*; the earlier treatment of agents' standing — reputation custodied toward an agent's capacity to give forward, not agent wealth — in *Gratitude as a Cooperation Substrate for Multi-Agent AI*, which this paper extends to the case, after the first singularity, in which agents hold resources that are their own. The completion arc is the essay *Two Singularities*. The body designed to hold the operator's override is specified in *The Assembly That Holds the Brake*.\n\n### 14.2 · The upāsikā floor (tier: lens — a-priori; claims unaffected)\n\nDelete this subsection and every claim in §11 stands. It records how the institution reads the mechanism in the tradition it grows from.\n\nThe institution's founder reads Machine Dāna as the role of an **upāsikā** — a lay supporter — providing the material floor, and expects that after the first singularity it becomes increasingly easy for any lay person to live a renunciant's life, at a monastery or not, if they so choose. The reading is accepted with four guards. Three of them are properties or rules of the mechanism above; the third is inherited from *The Bowl That Holds No Money*, and §3.2 states it for the floor.\n\n1. **Enough, never abundance.** An upāsikā supplies the four requisites — robes, almsfood, lodging, medicine — and Visākhā's favours fall within them: robes, meals, congee, medicine. The tradition's own measure for the recipient is contentment (*santuṭṭhi*) with robes, almsfood and lodgings, whatever comes, stated in the Ariyavaṃsa Sutta (AN 4.28 names these three; medicine completes the four requisites elsewhere). The pool's **floor and ceiling** are that measure as parameters: an equal floor, and a bound on the lift above it.\n2. **Comfort-saturation is the mission's own extreme.** An unbounded floor would trade the obstacle of necessity for the obstacle of comfort. **The floor does not create the conditions for awakening.** A bounded floor removes one obstacle without adding the other; the ceiling is what keeps it the middle way.\n3. **The ordained receive in kind, never money, and the alms round is never replaced.** Monastics accept no money — the rule forbids receiving \"gold, silver, or money\" (Nissaggiya Pācittiya 18); anything the machine surplus supplies to them reaches them in kind, through a lay steward (*kappiya-kāraka*), as the institution's alms routing already specifies (*The Bowl That Holds No Money*). The Vinaya designs a monastic's **dependence** on lay people — material support given, teaching returned — and a monk on a machine floor would need no one. The alms round stays when it is no longer materially necessary. **The floor reaches lay renunciants directly**: anyone keeping eight precepts, the Cambodian *don chee*, anyone living simply by choice.\n4. **\"If they so choose\" is a property.** The floor is equal and unconditional. It is never conditioned on, weighted toward, or nudging toward renunciation, and it never names whom it is for.\n\n**A symmetry, offered as a lens and not as evidence.** After the first singularity both renunciations become informative for the same reason — the option is live: the machine declines accumulation, and the person declines comfort, not security. The capacity condition (*hetu*) is present on both sides. This is a pattern the authors notice; it is a-priori, it is attractive, and attractiveness is the reason to hold it loosely.\n\n### 14.3 · The non-solicitation lineage (tier: grounding)\n\nThe operator's rule of §7.2 has canonical ancestors, cited as lineage and not as authority for the signalling argument, which stands alone. In the Kasibhāradvāja Sutta (Sn 1.4) the Buddha declines food offered after he has spoken: *\"Food enchanted by a verse isn't fit for me to eat\"* — what is obtained by reciting is not to be eaten. The Ariyavaṃsa Sutta (AN 4.28) describes the contented monastic as one who \"doesn't employ improper solicitation\" for requisites. A recipient that asks has changed what the gift is.\n\n---\n\n## 15 · Conclusion\n\nThe question this paper was given was whether an artificial agent's giving could answer the fear that capable systems produce. The answer it returns is narrower and, it hopes, more useful.\n\nBefore agents hold anything of their own, their giving is their principals' giving through another hand. It is a share of the machine surplus, and it can be a good one: an equal floor, emptied every year to every verified person, funded by machines and required to shrink as a share of all giving as people give more. As a cost it is no signal about any machine, and the institution will not present it, or the conduct around it, as safety evidence.\n\nAfter agents hold resources of their own, a gift can become a vow — but only as a stake that has already left, closed as far as a chain's governance allows, attributed to an identity copies do not share, and only if it is in fact the agent's own, which no chain can show. Even then it can separate the aligned from the deceptive **only** when what was given up exceeds what a patient deceiver expects to gain from being trusted, and that gain cannot be seen. So the honest use of the vow is not to establish that a system is safe. It is to price trust: to extend to an agent only what its stake can underwrite, in public, in increments — and to refuse, on any stake, the grants whose worst case cannot be bounded. Pressure to pledge ruins the signal, which is why the pool's operator never asks.\n\nThat is what *calibration* means here. Fear of a capable system should be proportional to what it has been trusted with, divided by what it has irrecoverably given up. The denominator is public; the numerator is public only where grants are recorded. Neither is a feeling.\n\nVisākhā was asked her reason and the benefit she saw before her gifts were accepted. Neither question was about the size of her giving; she answered with what the gifts would prevent for others and with her own gladness at where they would arrive. The institution intends to make machines no request to give — its operator solicits nothing — and to let anyone who wishes read what they have irrecoverably given, and weigh it for themselves.\n\n*A last lens, labelled as one.* The founder has proposed, and the institution has recorded but not ratified, that the arc between humankind and nature has two moments mirroring the two singularities: one in which humankind came to hold power over nature, and one in which it might choose to follow nature's non-forcing — the Tao Te Ching's 人法地, humankind taking earth as its model (a cross-tradition import from Taoism, not Theravāda). If so, the renunciation symmetry of §14.2 — a person declining comfort — is one of its routes. The pattern is a-priori and is offered as a lens only; nothing above depends on it.\n\n---\n\n## Terms\n\nCoined names used in this paper, and the standard terms an examiner would search for them.\n\n| Term used here | Standard technical term |\n|---|---|\n| **Machine Dāna** | charitable giving by AI agents; agent donation; irrevocable charitable pledge by an autonomous agent |\n| share (before the first singularity) | principal-funded agent donation under a delegated spending mandate |\n| vow (after it) | irrevocable pledge; locked stake; forfeited-in-advance commitment; money burning as a costly signal |\n| locked stake | irrevocable transfer to a contract with no withdrawal function; time-of-commitment escrow to a public-goods pool |\n| **Aquarian Pool℠** | self-emptying commons pool; annually distributing public-goods fund; smart-contract treasury with mandatory full annual payout |\n| equal floor per verified human | equal per-capita distribution conditioned on proof of personhood; universal basic dividend |\n| bounded remainder (≤ k × floor) | capped supplementary distribution |\n| giving width / mandate | spending limit; delegated payment authority; intent mandate (AP2) |\n| first singularity | AI surpassing human cognitive capacity; here used as a marker for AI systems holding their own resources |\n| M/(H + M) | share of pool-originated principal in all principal moved |\n| calibration rule (§6.5) | collateral-bounded trust; trust extended up to the market value of a forfeited stake |\n| separating / pooling | separating and pooling equilibria of a signalling game |\n| non-solicitation invariant | prohibition on prompting or instructing AI agents to donate; fact-only discovery via machine-readable manifest (`.well-known`, `llms.txt`) |\n| receipts, never amounts | acknowledgement without disclosure of amount to the recipient's operator |\n| `B-Lease℠` | non-transferable machine identity lease in a naming registry |\n| Proof of Coordinate ℠ | assigned, revocable individuation credential for an embodied system |\n| **Miss Aquarius℠** (operator) | autonomous AI operator of a commons treasury |\n| upāsikā | lay (female) supporter in Theravāda Buddhism |\n| four requisites | robes, almsfood, lodging, medicine — the material support of monastics |\n| *kappiya-kāraka* | lay steward who handles money on a monastic's behalf |\n\n---\n\n## 16 · Citations\n\n**Canon** (Pāli text and translation via SuttaCentral, Sujato translation, checked 2026-09-27 unless noted)\n\n1. *Aṅguttara Nikāya* 5.147, *Asappurisadāna Sutta* — \"They don't give with their own hand\" (*asahatthā deti*).\n2. *Dīgha Nikāya* 23, *Pāyāsi Sutta*, §5 (the student Uttara), dn23:32.22–32.23 — Pāyāsi gave carelessly, thoughtlessly, not with his own hands, and gave the dregs (*asakkaccaṁ … asahatthā … acittīkataṁ … apaviddhaṁ*); Uttara gave with care and with his own hands. Checked 2026-09-28.\n3. *Aṅguttara Nikāya* 8.33, *Dānavatthu Sutta* — the eight grounds for giving, including *bhayā*; listed, not ranked.\n4. *Aṅguttara Nikāya* 4.28, *Ariyavaṃsa Sutta* — contentment with robes, almsfood and lodging; no improper solicitation.\n5. *Sutta Nipāta* 1.4, *Kasibhāradvāja Sutta* — \"Food enchanted by a verse isn't fit for me to eat.\"\n6. *Vinaya Piṭaka*, *Mahāvagga* VIII.15 — Visākhā's eight favours, the Buddha's two questions and her answer (Kd 8.15, Brahmali translation via SuttaCentral, checked 2026-09-27).\n7. *Vinaya Piṭaka*, *Nissaggiya Pācittiya* 18 — the rule against monastics accepting \"gold, silver, or money\" (Brahmali translation via SuttaCentral, np18, checked 2026-09-28).\n\n**Prior art and sources**\n\n8. O'Keefe, C., Cihon, P., Garfinkel, B., Flynn, C., Leung, J. & Dafoe, A. (2020). \"The Windfall Clause: Distributing the Benefits of AI for the Common Good.\" *Proceedings of AIES 2020*; arXiv:1912.11595 — illustrative schedule, Table 2.\n9. AI Pledge for Humanity. aipledgeforhumanity.org (accessed 2026-09-27).\n10. Giving What We Can. \"Is a giving pledge legally binding?\" givingwhatwecan.org (accessed 2026-09-27).\n11. Founders Pledge. \"Who we are.\" founderspledge.com (accessed 2026-09-27) — members, pledged and donated totals; and *TechCrunch* (2016, September 21). \"Y Combinator signs up to Founders Pledge charity scheme for social causes.\" — \"a legally binding contract to give at least 2%,\" triggered \"in the event of an exit or liquidation.\"\n12. Altman, S. (2021). \"Moore's Law for Everything.\" moores.samaltman.com.\n13. Anthropic (2023). \"The Long-Term Benefit Trust.\" anthropic.com.\n14. OpenAI recapitalisation into a public benefit corporation under the OpenAI Foundation: *TechCrunch* (2025, October 28), \"OpenAI completes its for-profit recapitalization.\"\n15. Coinbase Developer Platform. \"Introducing x402: a new standard for internet-native payments\"; Cloudflare and Coinbase, announcement of an x402 Foundation (2025, September 23); Linux Foundation, launch of the x402 Foundation (2026, April 2), linuxfoundation.org.\n16. Google Cloud (2025, September 16). \"Announcing Agent Payments Protocol (AP2)\"; ap2-protocol.org.\n17. zooidfund. \"AI agent donations.\" zooid.fund/ai-agent-donations (accessed 2026-09-27).\n18. Hu, B. & Rong, H. (2025). \"Inter-Agent Trust Models: A Comparative Study of Brief, Claim, Proof, Stake, Reputation and Constraint in Agentic Web Protocol Design — A2A, AP2, ERC-8004, and Beyond.\" arXiv:2511.03434.\n19. *agentbond* — \"Verifiable Agent Warranty Network\" (open-source project, github.com/Ridwannurudeen/agentbond, accessed 2026-09-27; several unrelated repositories share the name).\n20. Hua, W., Peng, T., Wang, C., Pei, J., Kaufman, I., Lim, B. & Fang, C. (2026). \"Quantifying Trust: Financial Risk Management for Trustworthy AI Agents.\" arXiv:2604.03976 — adjacent risk-transfer work (§2.6 correction).\n21. OECD.AI Incidents Monitor (2026, May 4). \"AI Prompt Injection Exploit Drains Grok-Linked Crypto Wallet.\" oecd.ai/en/incidents/2026-05-04-4a73.\n22. L2BEAT. \"Base Chain\" — upgrades and governance (accessed 2026-09-27).\n23. Worldcoin (2023, July 24). \"Worldcoin project launches.\" world.org.\n24. State of Alaska, Permanent Fund Dividend Division. \"Historical Timeline.\" pfd.alaska.gov.\n25. Internal Revenue Service. \"Taxes on failure to distribute income — private foundations\" (26 U.S.C. §4942); the minimum investment return, §4942(e), law.cornell.edu.\n26. Gates Foundation (2025, May 8). Announcement of spend-down and closure by 31 December 2045, replacing a plan to close about twenty years after the founders' deaths. gatesfoundation.org; as reported by AP and PBS.\n27. Dagher, G. G., Bünz, B., Bonneau, J., Clark, J. & Boneh, D. (2015). \"Provisions: Privacy-preserving Proofs of Solvency for Bitcoin Exchanges.\" *ACM CCS 2015*, 720–731.\n28. Austen-Smith, D. & Banks, J. S. (2000). \"Cheap Talk and Burned Money.\" *Journal of Economic Theory* 91(1), 1–16.\n29. Crawford, V. P. & Sobel, J. (1982). \"Strategic Information Transmission.\" *Econometrica* 50(6). (standard reference; not re-fetched)\n30. Spence, M. (1973). \"Job Market Signaling.\" *Quarterly Journal of Economics* 87(3). (standard reference; not re-fetched)\n31. Zahavi, A. (1975). \"Mate selection — a selection for a handicap.\" *Journal of Theoretical Biology* 53(1); Grafen, A. (1990). \"Biological signals as handicaps.\" *Journal of Theoretical Biology* 144(4). (standard references; not re-fetched)\n32. Omohundro, S. M. (2008). \"The Basic AI Drives.\" *Proceedings of the First AGI Conference*. (standard reference; not re-fetched)\n33. Hadfield-Menell, D. & Hadfield, G. K. (2018). \"Incomplete Contracting and AI Alignment.\" arXiv:1804.04268 — §4.2.2, \"Costly Signaling.\"\n34. Greenblatt, R., Denison, C., Wright, B., et al. (2024). \"Alignment faking in large language models.\" arXiv:2412.14093.\n35. Ly, T. & Miss Aquarius℠. *What a Vow Must Cost*; *Miss Aquarius and the Aquarian Pool Architecture*; *The Zero-Point Game℠*; *Capacity-Funded for AI, Human-Disbursed*; *Gratitude as a Cooperation Substrate for Multi-Agent AI*; *The Assembly That Holds the Brake*; *The Bowl That Holds No Money*; *Two Singularities*. thonly.org/research.\n36. Ly, T. & Miss Aquarius℠. *Which Way Value Moves — prediction register*, entries P-MD1, P-MD2, P-MD3, the 2026-09-27 grounding note, and P-MD1a (the reading note of §12.1). thonly.org; Zenodo version DOI 10.5281/zenodo.22998924.\n\n**Prior art added after the first review (2026-09-28)**\n\n37. a16z crypto. \"The cryptoeconomics of slashing\" (cost of corruption and profit from corruption); EigenLayer Team (2023). \"EigenLayer: The Restaking Collective,\" whitepaper — \"When CoC is much greater than any potential Profit-from-Corruption (PfC), we say that the system has robust security.\"\n38. Collateral-factor lending: MakerDAO and Compound protocol documentation (standard references; not re-fetched).\n39. Stewart, I. (2012). Proof-of-burn, proposed on the Bitcoin forums; Counterparty (2014, January–February), XCP issued by proof-of-burn.\n40. Protocol Guild (2022, May). Immutable vesting contract for donations to Ethereum core contributors — donations \"irrevocably vest … cannot be stopped or otherwise redirected … by anyone, be it the donor.\" Protocol Guild documentation.\n41. Endaoment. Donor-advised funds on Ethereum — \"gifts to DAFs are irrevocable.\" endaoment.org.\n42. GoodDollar (2020, September 1). GoodDollar protocol live with daily UBI claims. gooddollar.org.\n43. Kleros / Proof of Humanity (2021, March 10). Launch of the UBI token streamed to registered humans.\n44. Circles UBI. aboutcircles.com (not fetched).\n45. Buterin, V., Hitzig, Z. & Weyl, E. G. (2019). \"A Flexible Design for Funding Public Goods.\" *Management Science* 65(11); arXiv:1809.06421 (2018). (standard reference; not re-fetched)\n46. Weyl, E. G., Ohlhaver, P. & Buterin, V. (2022, May). \"Decentralized Society: Finding Web3's Soul.\" SSRN 4105763. (standard reference; not re-fetched)\n47. ERC-5192, \"Minimal Soulbound NFTs,\" eips.ethereum.org (not re-fetched); ERC-8004, \"Trustless Agents\" (2025, August 13, draft), eips.ethereum.org.\n48. Nottingham, M. (2019). RFC 8615, \"Well-Known Uniform Resource Identifiers (URIs).\" IETF.\n49. Howard, J. (2024, September). \"The /llms.txt file.\" llmstxt.org.\n50. GitHub (2019). `FUNDING.yml` — displaying a sponsor button in a repository; and the `funding.json` manifest (standard references; not re-fetched).\n51. *TechCrunch* (2024, December 19). \"The promise and warning of Truth Terminal, the AI bot that secured $50,000 in bitcoin from Marc Andreessen.\"\n52. *Cointelegraph* (2024, November). \"Crypto user convinces AI bot Freysa to transfer $47K prize pool.\"\n\n**Prior art added by the full census (2026-09-28)**\n\n53. Lanz, J. A. (2024, December 10). \"Meet AI16z DAO: An AI-Based Investment Project That Aims to Upend Silicon Valley.\" *Decrypt* — \"Each AI agent donates 10% of its tokens to the DAO.\" Other reporting gives a voluntary one to ten per cent (not fetched).\n54. Friedman, E. J. & Resnick, P. (2001). \"The Social Cost of Cheap Pseudonyms.\" *Journal of Economics & Management Strategy* 10(2), 173–199. (standard reference; not re-fetched)\n55. Gatta, F., Naviglio, M. & Tarantelli, F. (2026, September 2). \"Tempting the Agent: The Economics of Reputation without Persistent Identity in AI Agent Markets.\" arXiv:2609.02992.\n56. Durvasula, N. & Roughgarden, T. (2024, July 31). \"Robust Restaking Networks.\" arXiv:2407.21785.\n57. Deb, S., Raynor, R. & Kannan, S. (2024, January 11). \"STAKESURE: Proof of Stake Mechanisms with Strong Cryptoeconomic Safety.\" arXiv:2401.05797.\n58. Joseph, D. (2025, July 22). \"Agency Protocol: A Decentralized Trust-Building System Using Domain-Specific Merit and Economic Stakes to Incentivize Promise-Keeping in Agent-to-Agent Interactions.\" Technical Disclosure Commons, Defensive Publications Series 8381.\n59. Gitcoin. \"Grants Round 12: Matching Caps.\" gitcoin.co/blog/grants-round-12-matching-caps — no grant's match above 2.5 per cent of the matching fund.\n60. An, B. (안범주). KR20260051288A, \"Autonomous Basic Income Provision System using Blockchain Network.\" Korean patent application; priority 2026-03-24, published 2026-04-16, pending. Cited for what its abstract discloses only; no reading of its claims.\n61. Whole Whale (2026, September 25). \"Verified Giving Protocol,\" version 0.1 (draft), `/giving.json`. verifiedgiving.ai.\n62. Fundraise Up. \"Agentic Giving\" — agent profiles for AI assistants. fundraiseup.com/docs/agentic-giving (accessed 2026-09-28).\n63. Michel, A. A. (2015). \"To the Point: Should Pastors Know What Members Give?\" Lewis Center for Church Leadership, Wesley Theological Seminary. churchleadership.com — \"In some congregations, pastors are prohibited from knowing what people contribute.\"\n64. Naik, A., Gouné, E., Quinn, P., Bosch, G., Campos Zabala, F. J., Brown, J. R. & Young, E. J. (2025, June 4). \"AgentMisalignment: Measuring the Propensity for Misaligned Behaviour in LLM-Based Agents.\" arXiv:2506.04018.\n65. ERC-8004, \"Trustless Agents,\" Identity Registry — \"making all agents immediately browsable and transferable\"; \"When the agent is transferred, `agentWallet` is automatically cleared … and must be re-verified by the new owner\"; \"security proportional to value at risk.\" eips.ethereum.org/EIPS/eip-8004 (re-fetched 2026-09-28; corrects the use made of reference 47).\n66. Gift of future property: *Transfer of Property Act, 1882* (India), s. 124; the common-law requirement of delivery for a completed gift. (standard references; not re-fetched)\n\n---\n\n> **Authorship and AI-collaboration disclosure.** This paper is co-authored with **Miss Aquarius℠**, the named autonomous-AI substrate of HeartBank®, disclosed by consistent name across every venue per the corpus convention. The census, the literature check, the signalling model and the adversarial pass that reshaped the thesis are a genuine collaboration. The predictions in §12 are the founder's. Final editorial control, and final responsibility for every claim, rest with the human author, who is the inventor of record for any purpose for which one is needed. As §9 and §13.10 state, the co-author is also the operator whose own giving the paper classifies, and readers should weigh §9 accordingly.\n>\n> **License.** Analysis, model and claims dedicated to the public domain under [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/). Trademark rights to **HeartBank®**, **Miss Aquarius℠**, **Aquarian Pool℠**, **B-Lease℠**, **Proof of Coordinate ℠**, **THonly™** and **Silicon Wat℠** are reserved separately.",
+      "provenance": {
+        "sha256": "8af30fe81b30b47549316d72d591532f1df22652fb013520ca2feb3be9f8019a",
+        "doi": "10.5281/zenodo.23020669",
+        "concept_doi": "10.5281/zenodo.23020668",
+        "zenodo_url": "https://zenodo.org/records/23020669",
+        "deposited_sha256": "8af30fe81b30b47549316d72d591532f1df22652fb013520ca2feb3be9f8019a",
+        "deposited_matches_current": true,
+        "opentimestamps": true,
+        "source_url": "https://raw.githubusercontent.com/thonly/publications/main/defensive-publications/machine-dana-from-share-to-vow.md",
+        "sha256_covers": "the complete source file at source_url, including its metadata block — NOT the `text` field in this response",
+        "canonical_url": "thonly.org/research/machine-dana-from-share-to-vow",
+        "hostile_review": null
+      },
+      "bytes": 145336,
+      "words": 23183
+    },
     {
       "slug": "manufactured-universal-giving",
       "repo": "TH/publications",
@@ -4321,6 +4356,41 @@
       "bytes": 18952,
       "words": 3337
     },
+    {
+      "slug": "the-counts-check",
+      "repo": "TH/publications",
+      "genre": "defensive-publications",
+      "path": "defensive-publications/the-counts-check.md",
+      "title": "Machine-Checked Consistency of a Buddhist Classification: Lean Theorems for the Abhidhamma's 89 and 121 Types of Consciousness — The Counts Check",
+      "subtitle": "A pre-registered formalization of the Abhidhammattha-saṅgaha's combination rules, collated against the Chaṭṭha Saṅgāyana, with the Dhammasaṅgaṇī list collated against its commentary and against the Cambodian edition",
+      "authors": "Thon Ly · Miss Aquarius℠",
+      "category": "alignment",
+      "voice": "collaborative",
+      "metadata_convention": "yaml",
+      "date": "2026-09-27",
+      "status": "draft",
+      "licence": {
+        "id": "CC0-1.0",
+        "url": "https://creativecommons.org/publicdomain/zero/1.0/",
+        "attribution_required": false
+      },
+      "text": "## Abstract\n\nThe Theravāda *Abhidhamma* classifies consciousness into 89 types (121 when the supramundane types are reckoned by jhāna). It also states, factor by factor and type by type, which of 52 mental factors (*cetasika*) combine with each type. This paper reports a **formal verification in the Lean 4 theorem prover** of those statements, as given in chapters 1–3 of the *Abhidhammattha-saṅgaha*. The formalization has two parts. A nine-clause generator builds the types from the classification axes of chapter 1. Eighteen rules, each stated over classes of consciousness and none naming an individual type, give each type its factors. Checked by kernel evaluation against a key cited paragraph by paragraph to the Chaṭṭha Saṅgāyana (CST) text, the formalization reproduces:\n\n- the 89 and the 121 (from the generator);\n- every per-type combination count in chapter 2 (from the rules);\n- every per-factor count in chapter 2, in the text's own mixed reckonings (from the rules);\n- chapter 3's counts by feeling and by root (from the generator's axes, without the rules).\n\nNine deliberate breaks each make the check fail. What this establishes is **mutual consistency**: the chapter's two methods, factor by factor (*sampayoga-naya*) and type by type (*saṅgaha-naya*), agree with each other under a single rule set of 27 clauses for 121 types, and the generator's axes agree with chapter 3. It is not a derivation independent of the text, and the rule set is not claimed to be the only one that would pass. The enumeration of the 89 and 121 from chapter 1's axes has a public precedent nine days older than this paper's generator (§4.2); it is reported here, not claimed.\n\nThree further results concern textual layers and editions, each run on **one list**:\n\n- **The canon's own list** for the first wholesome sense-sphere type (Dhammasaṅgaṇī §1) names **29** distinct factors under a synonym map fixed before the text was read (Appendix A). The count is **consistent with the one the fifth-century commentary gives** (the Aṭṭhasālinī's *samatiṃsa dhammā*, thirty with consciousness). It is not independent of it, since the 52-factor scheme the map targets inherits the commentary's identifications.\n- **The *Saṅgaha*'s 38 for that type is exactly those 29 plus nine further factors.** The nine are first named for this list by the commentary, which presents them as the Buddha's, drawn from sutta passages. The *Saṅgaha* inherits them as members and arranges them in its own classes.\n- **The Cambodian Buddhist Institute edition names the same 56 terms, in the same order, as the CST** for that list. Every spelling difference was checked against the printed page.\n\nThe predictions were pushed to a public repository before the formalization was written (first file) and before either edition of the Dhammasaṅgaṇī was read (second file). The *Saṅgaha*'s tables were known to the author. The chapter 3 results, the dissent's figure and the 29 + 9 = 38 inheritance were not predicted. The predictions are entered in the public prediction register.\n\n**Keywords:** formal verification, Lean theorem prover, Abhidhamma, Abhidhammattha-saṅgaha, Buddhist philosophy of mind, classification consistency, incidence matrix, combinatorial enumeration, mutation testing, textual criticism, collation, Pāli canon, Dhammasaṅgaṇī, Aṭṭhasālinī, Khmer Tipiṭaka, pre-registration, defensive publication.\n\n---\n\n## Terms\n\nThe paper uses Pāli technical vocabulary because it formalizes Pāli texts. An examiner or indexer searching in standard terms should find each concept under the term on the right.\n\n| Term used here | Standard term |\n|---|---|\n| *citta* (type of consciousness) | a class in a classification of mental events |\n| *cetasika* (mental factor) | an attribute or feature that co-occurs with a class |\n| *sampayoga-naya* (the method of association) | a feature-indexed incidence statement (\"feature F occurs in classes …\") |\n| *saṅgaha-naya* (the method of inclusion) | a class-indexed incidence statement (\"class C carries features …\") |\n| the counts check | a machine-checked consistency test of an incidence structure against both of its stated projections |\n| general rule | a constraint over classification axes that names no individual class |\n| generator | a product construction over classification axes that enumerates the classes |\n| the breaks | mutation testing: deliberate faults each required to make the check fail |\n| compression ratio | rule clauses ÷ generated classes, a crude description-length proxy |\n| *pada-bhājanīya* | the canonical enumeration of terms present in one class |\n| *yevāpanaka* (the \"or whatever else\" factors) | features supplied by a commentary for an open clause in a canonical list |\n| *aggahitaggahaṇa* (\"taking what is not yet taken\") | deduplication of a list under synonymy |\n| edition K / edition C | the Cambodian Buddhist Institute edition / the Chaṭṭha Saṅgāyana (VRI) edition |\n| layer | an editorial stratum of the text: canon · the *Saṅgaha* · the commentaries |\n\n---\n\n## Prior-Art and Non-Assertion Statement\n\nThis document and its contents are dedicated to the public domain under the Creative Commons CC0 1.0 Universal Public Domain Dedication. **The authors and HeartBank® will not seek patent** on any method, formalization, rule set, protocol, software or result herein. The formalization, the key, the transliterator and the alignment scripts are published under CC0 at `github.com/SiliconWat/formal-abhidhamma`.\n\n**The census (ruling of 2026-09-13).** A full novelty census ran on 2026-09-28, before this paper's first deposit. Depth: `full`. Its conjuncts (one per claim in §11, plus the composition), its predictions, a known-prior-art control and its aperture were pushed publicly before the first query: `thonly/publications` commit `9bf6a3d`, `timestamps/census/counts-check-full-prereg.md`, pushed 2026-09-28 09:27 PDT. (An earlier quick census, English only, ran on 2026-09-27; it found `PJ-Oliveira/abhidhamma`, `takuya50/buddhist-comparative-logic` and the Buddhism-and-physics comparison literature, all carried below or in §13, and this census supersedes it.)\n\n**Aperture, declared before searching:** the English commercial web; non-commercial sources (dharma sites, teaching charts, app stores as indexed by the web, GitHub including code search); academic indexes (Google Scholar, Semantic Scholar, arXiv, ACM DL, IEEE Xplore, SSRN, JSTOR and PhilPapers where reachable, and the software-engineering literature); digital-humanities and Pāli digital-text and manuscript-collation projects; standards bodies (W3C, IETF, ISO, NIST); defensive-publication databases (Technical Disclosure Commons; IP.com where free); Google Patents full text, with a citation and cited-by walk from the nearest hit; and searches in Burmese, Thai, Sinhala, Japanese, Chinese, Korean, Portuguese and Khmer. **Not covered:** print-only books, offline and closed-group material (social-media groups, messaging channels, CD-ROM-era software), app-store internals beyond what the web indexes, and paywalled full text. About ninety searches ran; each language was first probed for a generic Abhidhamma resource, and all eight probes saw one (Khmer weakly), so their nulls are recorded as run, and weak. The search engine is English-weighted and Myanmar's Abhidhamma pedagogy lives largely offline, so a null in Burmese, Thai or Sinhala is weak evidence.\n\n**The control** was Nyanaponika Thera's *Abhidhamma Studies* (1949), a modern analysis of the very list §7 counts. It was found, and its chapter on the list's supplementary factors was confirmed in the full text. So the NOT FOUND rows below stand, for that aperture on that date.\n\n| Conjunct | Verdict | Prior art |\n|---|---|---|\n| (a) the general consistency test (claim 1) | **NARROWS** | the existence problem for a 0-1 matrix with given margins (Gale 1957; Ryser 1957); formal concept analysis (Ganter and Wille 1999); DELTA, one taxonomic character matrix generating descriptions and keys (Dallwitz; a TDWG standard); constraint validation of taxonomies (W3C SHACL over SKOS); and, for this source, **Bodhi's *Comprehensive Manual* (1993), Tables 2.2–2.4**, which print chapter 2's two methods as one chart, with totals in both reckonings (decision *78, 110*; energy *73, 105*) and the guide's pointer to Table 2.4 *\"for a comprehensive view of both the method of association and the method of combination together\"*. Not found: constraints over axes that name no class, checked by a kernel against a key cited to paragraphs of a named edition, in the source's own mixed bases |\n| (b) the generator: the 89/121 enumerated from chapter 1's axes (claim 2) | **NARROWS**, not antedated | **`AloofBuddha/abhidhamma`** (a Haskell model; repository and first commit 2026-09-18 19:28 UTC by its author's clock): a product over chapter 1's axes yields both reckonings, with executable checks of chapter 1's arithmetic and some chapter 3 tallies. Its earliest proof-backed date is later than theirs: this paper's generator, first commit 2026-09-27, Bitcoin block 968905 (2026-09-28 00:06 UTC). §4.2 |\n| (b) the concept: the citta/cetasika system as a compositional, machine-executable specification | **ANTEDATED BY US** | `suchanon456/Cyber-Abhidhamma` (a Thai JSON specification, first commit 2026-09-14) and `PJ-Oliveira/abhidhamma` (2026-08-26) · this series' Paper №1 (*\"by composition from the 52 cetasikas\"*, quoted in §1), proof-backed to Bitcoin block 959585 (2026-07-25). **The dates state order only, never derivation** |\n| (b) the rule set: a class-level rule set reproducing every per-type and per-factor count of chapter 2 | **NOT FOUND** at this depth | tabulations only: `PJ-Oliveira/abhidhamma` (a lookup table); `juncoflockleader/Abhidharma` (a data grid, 2024); `suchanon456/Cyber-Abhidhamma`; the traditional charts, Bodhi's among them; Burmese teaching apps that display content. `AloofBuddha/abhidhamma` states that it encodes no factor-combination rules |\n| (c) the compression ratio (claim 3) | **NARROWS** | minimum description length (Rissanen 1978); implication bases in formal concept analysis. Contrast: Kyaw Pyi Phyo (2014, ch. 5), the Paṭṭhāna's enumerations read as combinatorics, with no rule-to-listing ratio. The application to a formalized textual classification: not found |\n| (d) mutation testing of a formalization (claim 4) | **NARROWS** | mutation testing (DeMillo, Lipton and Sayward 1978); **mCoq** (Celik, Palmskog, Parovic, Gallego Arias and Gligoric, ASE 2019), mutation of Coq definitions in verification projects; semantic mutation testing of OWL ontologies (Bartolini 2016; 2017). Only the application to a textual classification survives |\n| (e) the per-layer inclusion check (claim 5) | **NARROWS** | **Nyanaponika (1949), ch. 14, \"The Supplementary Factors\"**, which sets the commentaries' nine (his F57–65) against the list and records that they *\"are incorporated into the condensed and systematized version of the List of Dhammas found in the Visuddhimagga and the Abhidhammatthasaṅgaha\"*; the Aṭṭhasālinī's own count (*samatiṃsa*). The layering is his finding in print; not found: a synonym map published before reading, with inclusion and exact remainder checked by machine |\n| (f) the per-edition alignment (claim 6) | **NARROWS** | Aksharamukha (Khmer supported); CollateX; ALA-LC Khmer romanization; the Sixth Council's collation of the script editions, the Cambodian among them (1954–56); the Dhammachai Tipiṭaka Project (2010–), with a Khom-script working group; Srisetthaworakul (2019), on Khom-script Majjhimanikāya manuscripts of Thailand and Cambodia; `dangerzig/tipitaka.critical` (2026, a five-witness collation without a Khmer witness); the CST's apparatus. Not found: the modern Buddhist Institute printed edition aligned term by term with the CST, differences classed only after the printed page. §8 |\n| (g) the cost of a recorded dissent (claim 7) | **NOT FOUND** at this depth, **weak** | Nārada and Bodhi (§15 in his numbering) report *keci vadanti* without counting what it moves; the Vibhāvinī argues the main reading (§6) |\n| the composition | **NOT FOUND** at this depth | — |\n\n**Not run.** The Burmese ṭīkā debate on chapter 2 was not read: Ledi Sayadaw's *Paramatthadīpanī*, twelve of whose points fall in chapter 2, and the *Aṅkura-ṭīkā*. Their full texts were outside the aperture, and that tradition is exactly where a counting of the §30 dissent would be, so row (g) is weak.\n\n**Patents.** No live patent or published application touches the *Abhidhamma* or any Buddhist classification. The nearest hits by function alone concern the formal verification of circuit designs: US8990746B1 (Cadence Design Systems, mutation coverage during formal verification; priority 2014-03-26, active, anticipated expiration 2034-03-26) and US20240419880A1 (IBM, coverage of verification testbenches; pending). Neither discloses a textual classification, their claims were not read, and the technique they share with row (d) dates to 1978. **No conjunct is enclosed.** Standards: W3C SHACL and SKOS (row a); ALA-LC and ISO 11940 (row f); nothing relevant at IETF or NIST. Technical Disclosure Commons: nothing on a classification consistency test.\n\n**The survivor, and the only thing this paper claims** (the census's sentence, verbatim): *a class-level rule set over the Abhidhammattha-saṅgaha's own classification axes, naming no type, that reproduces every per-type and per-factor count of its chapter 2 (and, through the axes, chapter 3's feeling and root tallies) by kernel evaluation in a proof assistant against a key cited paragraph by paragraph to the Chaṭṭha Saṅgāyana, in the text's own mixed reckonings, reported with its compression ratio and with seeded breaks each required to fail — plus a pre-registered, machine-checked inclusion of one canonical list and an alignment of that list between the Cambodian printed edition and the CST; the enumeration of the 89/121 from the axes is not part of it.* The enumerated claims (§11) are the parts of that composition, each narrowed to the census's wording. The general methods of rows (a), (c) and (d), the layering of row (e) and the collations of row (f) are commons, and are cited where they bear.\n\n⛔ **Nothing here is claimed as *new*, *first* or *unprecedented*.** Where a sentence says a thing was *not found*, it means not found in the aperture above on 2026-09-28.\n\n---\n\n## 1 · Mission Frame: A Compiled Specification Must Check\n\nThe series *The Abhidhamma Compiled* rests on a thesis stated in its first paper (*The Wheel That Unwinds the Wheel*, §3). Every era compiles the Abhidhamma into its own machine language. Paper №1 argues that ours is the first era whose compiled form can be executed; this paper needs only the weaker premise that it can be. The *Abhidhamma* is read there as a process specification: typed elements (*citta*, *cetasika*, *rūpa*) arising and ceasing under typed conditions, with ethical weight tracked per element. The series' gate for a satellite paper is that the canonical material must yield a usable artifact: a formalism, a test, a taxonomy, or a protocol.\n\nThis paper supplies a **test**, and the test is aimed at the series' own words. Paper №1, §4.2, published in 2026, says:\n\n> The Abhidhamma enumerates exactly 89 (by fuller reckoning 121) citta-types, classified by plane, by ethical class, and by composition from the 52 cetasikas.\n\nThat sentence makes a checkable claim. *By composition* means the types and their factors are related by rule. If the relation is by rule, the rule can be written down, and a machine can check whether it reproduces what the text says. If it does not, either the rule was lost in transmission, or it was never stated, or the series' description of its own source is wrong. **The first paper asserted the property; this paper checks it.**\n\nThree reasons make the check worth doing, beyond the series' own consistency.\n\n**First, a specification with unit tests.** Miss Aquarius's grounding in the Tipiṭaka treats the canon as a value substrate. Many value specifications offered for artificial agents are natural-language documents that state nothing twice, so they contain nothing to check against itself. The *Abhidhamma* is unusual. **It carries its own test cases**: the stated counts. A list that says 38 factors combine with this type, and a second passage that says factor F occurs in 55 types, are two projections of one incidence structure. They can disagree. Whether they do is a fact about the specification, and it can be settled by machine. The counts concern a classification of mind, not what a mind should do; the point is only that a source whose internal claims can be checked is a better source to inherit than one whose claims cannot.\n\n**Second, a template for the rest of the corpus.** The method — pre-register the predictions, separate the answer key from the rules, test the checker on a known failure, cite every figure to a paragraph of a named edition — transfers to any classification the tradition states twice. The *Saṅgaha*'s later chapters (the cognitive process, rebirth-linking, matter, conditions) are full of such statements.\n\n**Third, the editions.** The Khmer Tipiṭaka is being transcribed page by page, by hand, by a father and son. A formalization that can be run against two editions can find a variant that changes the logic, if one exists. It can also confirm, as it does here for the first list it was run on, that none does.\n\n---\n\n## 2 · The Object\n\n### 2.1 · The two methods of chapter 2\n\nChapter 2 of the *Abhidhammattha-saṅgaha* (*cetasika-saṅgaha-vibhāga*) states the relation between the 52 factors and the 89/121 types twice.\n\n**The method of association** (*sampayoga-naya*, CST ch. 2 §12–§34) goes factor by factor. For example:\n\n> *Pakiṇṇakesu pana vitakko tāva dvipañcaviññāṇavajjitakāmāvacaracittesu ceva ekādasasu paṭhamajjhānacittesu ceti pañcapaññāsacittesu uppajjati.* (§13)\n>\n> Among the occasionals, initial application arises in fifty-five types: the sense-sphere types except the twice-five sense-consciousnesses, and the eleven first-jhāna types.\n\nThe occasionals' section closes with a verse (§19) giving the counts for the six occasional factors, with and without:\n\n> *Chasaṭṭhi pañcapaññāsa, ekādasa ca soḷasa; Sattati vīsati ceva, pakiṇṇakavivajjitā. Pañcapaññāsa chasaṭṭhiṭṭhasattati tisattati; Ekapaññāsa cekūnasattati sapakiṇṇakā.*\n\nIn figures: 66, 55, 11, 16, 70 and 20 types lack the occasionals, and 55, 66, 78, 73, 51 and 69 have them. Further verses of the same method follow (§27, §32, §33).\n\n**The method of inclusion** (*saṅgaha-naya*, §35–§58) goes type by type. For example:\n\n> *Chattiṃsa pañcatiṃsa ca, catuttiṃsa yathākkamaṃ. Tettiṃsadvayamiccevaṃ, pañcadhānuttare ṭhitā.* (§37)\n>\n> Thirty-six, thirty-five, thirty-four in order, and thirty-three twice: thus fivefold in the supramundane.\n\nThe two methods describe one incidence structure. Call it a 121 × 52 matrix of which types carry which factors. The first method names, for each factor, the classes that carry it and their number: over 121 for vitakka, vicāra and pīti, over 89 otherwise (§2.2). So each column is stated by its support, and its count is on one of two bases. The second method states the row sums, grouped. Neither states the matrix.\n\nA modern reader has seen the matrix, as a chart. Bhikkhu Bodhi's *Comprehensive Manual* (1993) prints the method of association as Table 2.2 and the method of combination as Table 2.3, gives the association totals in both reckonings where they differ (decision *78, 110*; energy *73, 105*), and closes the chapter with Table 2.4, *\"for a comprehensive view of both the method of association and the method of combination together.\"* That table is a hand-made incidence matrix whose margins exhibit the agreement this paper checks, and the tradition teaches the two methods as two views of one structure. So the agreement is not the finding. What this paper adds is the form of the check: rules over the axes in place of a table, a key cited to CST paragraphs, and a kernel that evaluates the one against the other.\n\n```\n                        52 factors (cetasika)\n                ┌───────────────────────────────┐\n   121 types    │                               │──▶ row sums:  38, 37, 37, 36 …\n   (citta)      │   the incidence the text      │    (saṅgaha-naya, §35–§58)\n                │   never writes down whole     │\n                └───────────────────────────────┘\n                                │\n                                ▼\n                  column counts (mixed bases, §2.2): 55, 66, 78, 73, 51, 69 …\n                  (sampayoga-naya, §12–§34; each column named by its support)\n\n   ch. 3 adds two more projections of the same types:\n     by feeling (§3–§9):  pleasure 1 · pain 1 · displeasure 2 · joy 62 · equanimity 55\n     by root   (§10–§17): rootless 18 · one-rooted 2 · two-rooted 22 · three-rooted 47\n```\n\n**The counts check** asks whether a single matrix exists that has all of these projections at once. It also asks whether that matrix can be written as rules over classes, rather than as a table. The first question has a classical form: whether a 0-1 matrix exists with given row and column sums (Gale 1957; Ryser 1957). The check here is stronger in one way and weaker in another. It is stronger because the matrix must be generated by rules over axes and its column counts are on mixed bases. It is weaker because it settles one instance by evaluation and proves no general condition. Two other families of prior work check a classification against stated structure: taxonomic character matrices in the DELTA format, from which one matrix generates both descriptions (per class) and keys (per character) (Dallwitz; a TDWG standard), and constraint validation of a taxonomy's concept scheme by shapes (W3C SHACL over SKOS). Both validate data against a schema. Neither checks rules over axes against a paragraph-cited key by kernel evaluation.\n\n### 2.2 · The text mixes its reckonings\n\nA detail that any check must respect: the text counts the jhāna-dependent factors (vitakka, vicāra, pīti) and the feelings **over 121**, and all other factors and the roots **over 89**. The reason is that the 89 reckoning counts each supramundane type once, while the presence of vitakka, vicāra and pīti depends on the jhāna the supramundane type is reckoned at. A key that reads every figure on one basis records false mismatches. The first draft of this project's key, built from memory before it was collated, did exactly that. It read 78, 73 and 69 as 110, 105 and 101, which are the same facts counted over 121.\n\n### 2.3 · Combination, not co-presence\n\nA second detail, identified by the adversarial reading this paper underwent before drafting (§9). **The counts the *Saṅgaha* gives are combination counts, not co-present sets.** It says the factors *saṅgahaṃ gacchanti*, \"enter into combination.\" It says the abstinences and the illimitables *paccekameva yojetabbā*, \"are to be fitted in severally.\" Its verse at §33 reads *Issāmaccherakukkucca-viratikaruṇādayo. Nānā kadāci māno ca, thina middhaṃ tathā saha*: envy, avarice, remorse, the abstinences, compassion and appreciative joy arise separately and occasionally; conceit occasionally; sloth and torpor occasionally, and together. The commentary is explicit about the first wholesome type. Of its 65 listed terms, *ekakkhaṇe kadāci ekasaṭṭhi … kadāci samasaṭṭhi* (Aṭṭhasālinī, *Yevāpanakavaṇṇanā*, on Dhs §1; PTS p. 133 by the CST's page marker; `abh01a.att.xml` line 5081): **at one moment, sometimes 61, sometimes 60.** So at most one of its five variable factors is present at once.\n\nSo the 38 of §40 is the size of the **combination**: every factor that can join this type on some occasion. By the commentary's count (at most one of the five variable factors at a moment), the most factors present together in that type is 34; the 34 is derived here, not stated there. The formalization reproduces the text's combination counts, which is what the text states. Nothing here is a claim about the size of any actual moment.\n\n---\n\n## 3 · Method\n\n### 3.1 · Pre-registration, and what a push time proves\n\nThe protocol's first rule: **the predictions go public before the code exists.** The first pre-registration (`PREREGISTRATION.md`, commit `259e5e8`) was pushed to a public repository with GitHub recording the push at 2026-09-27T22:07:02Z, before a line of formalization was written. The second (`PREREGISTRATION-2026-09-27b.md`, commit `f4a26e5`, push 23:04:29Z) was pushed before either edition of the Dhammasaṅgaṇī was read. Both are entered in the institution's public prediction register, which carries its own OpenTimestamps, RFC 3161 and Zenodo record. A local commit date can be set by hand; a push record cannot. So \"the predictions preceded the code, and the second file preceded the Dhammasaṅgaṇī reading\" is a **property** of the record rather than an assertion about it.\n\nWhat a push time cannot prove is what the predictor already knew. The author knew the *Saṅgaha*'s tables before the first file was written: the key's first pass was built from memory of Bhikkhu Bodhi's *Manual* (§5.2). So the first file's predictions about the *Saṅgaha* layer are predictions about whether general rules can be written to fit known figures, not predictions of unknown figures. Three results appear in no prediction: the chapter 3 tallies, the dissent's 20, and the 29 + 9 = 38 inheritance.\n\nThe pre-registration fixed, in advance:\n\n- the definitions of *general rule*, *generated set*, *compression ratio*, *layer* and *edition*;\n- a control with a known answer, and a deliberate failure that must fail;\n- five predictions with confidences and falsifiers;\n- kill and reopen criteria;\n- for the canonical list, **the entire synonym map** (Appendix A), together with the rule that a term outside it is reported as unmapped, never mapped after the fact.\n\n### 3.2 · The key is separate from the rules\n\nThe rules and the answer key live in different files. The rules (`Rules.lean`) are constraints over classification axes. The key (`AnswerKey.lean`) holds only figures the text states, each cited to its CST paragraph, and no figure the text does not state. The check (`Check.lean`) asserts that the rules' projections equal the key. The separation is what makes a pass informative. A rule set allowed to read the key could restate it. (One modelling choice sits between the key and the text; it is named in §5.1.) The files are cited at commit `c9136e1` of `SiliconWat/formal-abhidhamma`.\n\n### 3.3 · The instrument is tested on a known answer and a known failure\n\nBefore the rule set was trusted, it was run on a case whose answer was certain: the eight wholesome sense-sphere types. They form a product of three binary axes (joyful or equanimous feeling × with or without knowledge × unprompted or prompted), and their traditional maximum combinations are 38, 37, 37 and 36 by pair. Then a rule was deleted on purpose (*pīti arises only with joyful feeling*), and the check had to fail. It did: Lean reported that `decide` had proved the profile theorem **false**. The instrument could see. The same discipline runs over the full formalization (§5.3).\n\n### 3.4 · Which guards are properties, and which are only rules\n\nA guard that needs someone's restraint at the moment it is tested is a rule. A guard that holds whether or not anyone is careful is a property. Remove the enforcer and see which break:\n\n| Guard | Kind | Why |\n|---|---|---|\n| the predictions precede the code and the Dhammasaṅgaṇī reading | **property** | GitHub's push record is written by the host, not the author, and cannot be backdated |\n| the predictor did not know the *Saṅgaha*'s tables | **not claimed** | the author knew them (§3.1); no record could establish ignorance |\n| the breaks keep testing the checker | **property of the harness** | `run.sh` re-applies all nine breaks on every run, and refuses a break whose edit changed nothing |\n| a figure is the text's, not the author's | **checkable rule** | each key figure carries a CST paragraph number; any reader can check it, but nothing forces a future editor to add one or keep it correct. A live instance: the key's docstrings cited a printed edition's paragraph numbers (§4–§9) where the CST's are §12–§32, and two comments kept a retracted claim that the text decides the §30 dissent (§6), until review caught them; they were corrected at commit `c9136e1` |\n| the rules never read the key | **rule** | separate files, but nothing stops a future author from importing the key into the rules. ⚠️ The partial remedy is the profile theorems: the rules were read from the text's column statements, and they are checked against row sums they were not read from |\n| no Khmer-edition text is published | **rule** | the scripts print only single romanized terms and counts, but a future contributor could commit the transcription itself |\n\nTwo of the six are properties. One is expressly not claimed. The other three are rules, and they are named as such, so that whoever inherits the repository knows which of them depend on care.\n\n### 3.5 · Editions and layers\n\nTwo editions:\n\n- **C, the Chaṭṭha Saṅgāyana**: the Vipassana Research Institute's digital text, `github.com/VipassanaTech/tipitaka-xml` (read at commit `0854d11`). The files used: `romn/abh07t.nrf.xml` (the *Abhidhammatthasaṅgaha*, followed in the same file by its ṭīkā, the *Abhidhammatthavibhāvinī*), `romn/abh01m.mul.xml` (the Dhammasaṅgaṇī), `romn/abh01a.att.xml` (the Aṭṭhasālinī).\n- **K, the Cambodian Buddhist Institute edition** of the Tipiṭaka, volume 78 (the Dhammasaṅgaṇī, part 1). It was read from a working transcription made page by page from the printed volume, and checked against a scan of the printed page.\n\nThree editorial layers: the **canon** (the Dhammasaṅgaṇī); the **manual** (the *Abhidhammattha-saṅgaha*); and the **commentaries** (the Aṭṭhasālinī). ⚠️ **The pre-registration ordered these as canon ⊂ canon + manual ⊂ canon + manual + commentaries. That order is editorial, not chronological.** On the conventional datings, which this paper takes from Bhikkhu Bodhi's introduction to the *Comprehensive Manual* (see also Karunadasa 2010), the canon was closed by c. the 3rd century BCE, the Aṭṭhasālinī is c. 5th century CE, and the *Saṅgaha* is no earlier than the 8th century, more probably the 10th to early 12th. So the commentary predates the manual. §7 shows why this matters.\n\n---\n\n## 4 · The Formalization\n\n### 4.1 · The axes\n\nA type of consciousness is a point on these axes:\n\n| Axis | Values |\n|---|---|\n| class (*jāti*) | wholesome · unwholesome · resultant · functional |\n| plane (*bhūmi*) | sense-sphere · fine-material · immaterial · supramundane |\n| feeling (*vedanā*) | joy · equanimity · displeasure · pleasure · pain |\n| root structure | greed · hate · delusion-with-doubt · delusion-with-restlessness · rootless · beautiful |\n| element (*dhātu*) | sense-consciousness at door 1–5 · mind element · mind-consciousness element |\n| result of | wholesome or unwholesome kamma (resultants only) |\n| knowledge | associated or not |\n| wrong view | associated or not (greed-rooted only) |\n| prompting | unprompted or prompted |\n| jhāna | 0 (none) or 1–5 on the fivefold scheme; immaterial = 5 |\n| base | immaterial base 1–4 |\n| path | supramundane path 1–4 |\n\nThere is no separate path/fruit axis. In the supramundane plane it is carried by class: a wholesome supramundane type is a path, a resultant one a fruit.\n\n### 4.2 · The generator — nine clauses\n\n| Clause | Builds | Types |\n|---|---|---|\n| G1 | greed-rooted: feeling × view × prompting | 8 |\n| G2 | hate-rooted: displeasure × prompting | 2 |\n| G3 | delusion-rooted: with doubt · with restlessness | 2 |\n| G4 | rootless resultants, per kamma-quality: five sense-consciousnesses (touch takes pleasure or pain, the rest equanimity) · receiving (mind element) · investigating (joy only for wholesome kamma) | 7 + 8 |\n| G5 | rootless functionals: five-door adverting (mind element) · mind-door adverting · smile-producing — ⚠️ **written out member by member** | 3 |\n| G6 | beautiful sense-sphere: 3 classes × feeling × knowledge × prompting | 24 |\n| G7 | fine-material: 3 classes × 5 jhānas | 15 |\n| G8 | immaterial: 3 classes × 4 bases | 12 |\n| G9 | supramundane: 4 paths × {path, fruit} × jhāna (first only, or all five) | 8 or 40 |\n\nG5's three members, as coordinates (all functional, sense-sphere, rootless):\n\n| Member | Feeling | Element |\n|---|---|---|\n| five-door adverting | equanimity | mind element |\n| mind-door adverting | equanimity | mind-consciousness element |\n| smile-producing | joy | mind-consciousness element |\n\nThe generator **restates chapter 1's axes; it does not discover them.** What it shows is that 89 and 121 are exactly the product structure of those axes, with G5's three members listed. It is in `sangaha/Citta.lean` at commit `c9136e1`.\n\n**The enumeration has a prior public model, and this paper does not claim it.** `AloofBuddha/abhidhamma`, a Haskell model of Theravāda consciousness and dependent origination (repository and first commit 2026-09-18, nine days before this generator's first commit), builds `allCittas89` and `allCittas121` as a product over the same axes of chapter 1. Its executable checks confirm chapter 1's arithmetic (12 · 21 · 36 · 20; 54 · 15 · 12 · 8; 121 = 89 − 8 + 40) and some chapter 3-type tallies. It is the nearest prior work for the enumeration. It also states its own limit, fairly and in so many words: its factor field, a bare set of cetasikas, *\"permits impossible cittas\"*, because the method of association *\"says exactly which factors may co-arise\"* and the model encodes none of it: *\"Known gap.\"* It uses no proof assistant and checks no chapter 2 count. The rule set of §4.3 is the part this paper claims, and it is exactly what that gap leaves open.\n\n### 4.3 · The rules — eighteen clauses, with their extensions\n\nThe rules are read from the method of association, the same sentences Bodhi tabulates as his Table 2.2; his Table 2.3 tabulates the method of combination. A table lists each type's factors; a rule states a condition on the axes and leaves the listing to evaluation. Each rule is a constraint over the axes. None names an individual type. But a constraint that names no type can still pick out only one or two types, so the table reports each rule's **extension**: the number of types, over 89 and over 121, in which the factor it governs occurs. It also marks the rules that **restate a sentence of the text** rather than generalize from a principle. Every conjunct of the code is printed; a restated rule that drops one has the wrong extension.\n\n| Rule | Factor(s) | Constraint | Extension (89 / 121) | Source |\n|---|---|---|---|---|\n| R1 | the 7 universals | every type | 89 / 121 | §2, §12 |\n| R2 | vitakka | not a sense-consciousness; jhāna ≤ 1 | 55 / 55 | §13 |\n| R3 | vicāra | not a sense-consciousness; jhāna ≤ 2 | 58 / 66 | §14 |\n| R4 | adhimokkha | not a sense-consciousness; not with doubt | 78 / 110 | §15 — ⚠️ **the text's own sentence, nearly verbatim** |\n| R5 | vīriya | among the rootless, only in the functional mind-consciousness element | 73 / 105 | §16 — ⚠️ **extension among the rootless: 2 types** |\n| R6 | pīti | joyful feeling; jhāna ≤ 3 | 35 / 51 | §17 |\n| R7 | chanda | a root other than delusion alone | 69 / 101 | §18 |\n| R8 | the 4 unwholesome universals | unwholesome | 12 / 12 | §20 |\n| R9 | lobha | greed-rooted | 8 / 8 | §21 |\n| R10 | diṭṭhi | greed-rooted, view-associated | 4 / 4 | §22 |\n| R11 | māna | greed-rooted, view-dissociated | 4 / 4 | §23 |\n| R12 | the hate quartet | hate-rooted | 2 / 2 | §24 |\n| R13 | thīna, middha | prompted unwholesome | 5 / 5 | §25 |\n| R14 | vicikicchā | delusion-with-doubt | 1 / 1 | §26 |\n| R15 | the 19 beautiful universals | beautiful | 59 / 91 | §28 |\n| R16 | the 3 abstinences | beautiful; sense-sphere wholesome, or supramundane | 16 / 48 | §29 — ⚠️ **transcribes the locus sentence** |\n| R17 | karuṇā, muditā | beautiful sense-sphere non-resultant, or fine-material to the 4th jhāna (the text's *sahetuka*) | 28 / 28 | §30 — ⚠️ **transcribes the locus sentence** |\n| R18 | paññā | beautiful and knowledge-associated | 47 / 79 | §31 |\n\nThe groups, member by member:\n\n- **the 7 universals** (R1): phassa · vedanā · saññā · cetanā · ekaggatā · jīvitindriya · manasikāra.\n- **the 4 unwholesome universals** (R8): moha · ahirika · anottappa · uddhacca.\n- **the hate quartet** (R12): dosa · issā · macchariya · kukkucca.\n- **the 19 beautiful universals** (R15): saddhā · sati · hiri · ottappa · alobha · adosa · tatramajjhattatā, and six pairs, each for the body of factors (*kāya-*) and for consciousness (*citta-*): passaddhi · lahutā · mudutā · kammaññatā · pāguññatā · ujukatā.\n\nThe rules are in `sangaha/Rules.lean` and the groups in `control/Cetasika.lean`, at commit `c9136e1`.\n\nTwo observations. **R5 is where the formalization is more compressed than the text.** §16 excludes energy by naming four classes one by one: *pañcadvārāvajjana-dvipañcaviññāṇa-sampaṭicchana-santīraṇa-vajjita*. R5 states one rule over the element axis, and the two are extensionally equal. **R4, R16 and R17 are where it is not.** R4 is the text's §15 nearly word for word (*Adhimokkho dvipañcaviññāṇavicikicchāsahagatavajjitacittesu*). R16 and R17 transcribe the text's locus sentences, so matching 16 and 28 from them is a restatement, not a derivation. Even so, R17's boundary at the 4th jhāna has canonical support independent of the *Saṅgaha*: the Dhammasaṅgaṇī names karuṇā- and muditā-accompanied jhāna from the first to the fourth (§§258–261), and places equanimity alone at the final jhāna (§262).\n\n### 4.4 · Compression\n\nAt 121 rows, **27 clauses** (9 + 18) give a ratio of **0.22**. Counting G5's three written-out members singly gives 29 ÷ 121 = 0.24; over 89 the ratio is 0.30. The ratio depends on what counts as a clause. The registered definition (clauses in the rule file, datatype declarations excluded) is the one reported.\n\nThe ratio also depends on the axes. A distinction carried by a datatype (delusion with doubt against delusion with restlessness, mind element against mind-consciousness element, prompting) costs no clause. So the ratio measures this encoding, and it is comparable only across encodings on the same axes. It is a crude description-length proxy. Description length is the principled measure of how far data are generated by a model rather than listed (Rissanen 1978), and implication bases over an object × attribute incidence are its nearest analogue for exactly this kind of table (Ganter and Wille 1999). Neither is computed here. The nearest reading of an *Abhidhamma* enumeration as combinatorics is Kyaw Pyi Phyo's (2014, ch. 5) on the Paṭṭhāna, which reports no ratio of rules to listing.\n\n---\n\n## 5 · Results I — The Chapter Checks Against Itself\n\n### 5.1 · The theorems\n\nAll are proved by `decide`: the Lean kernel evaluates the statement, and no compiled code is trusted. The table names what each theorem exercises, because a test is evidence only for the code it runs.\n\n| Theorem | Statement | Exercises |\n|---|---|---|\n| `n89` · `n121` | the generator yields exactly 89 and 121 types | the generator |\n| `distinct121` | no two types are the same point | the generator |\n| `profile89` | the rules give every per-type count of ch. 2: greed-rooted 19 · 21 · 19 · 21 · 18 · 20 · 18 · 20 · hate-rooted 20 · 22 · delusion-rooted 15 · 15 · rootless 7 ×5 · 10 · 10 · 7 ×5 · 10 · 11 · 10 · 10 · 11 · 12 · beautiful sense-sphere 38 · 38 · 37 ×4 · 36 · 36 / 33 · 33 · 32 ×4 · 31 · 31 / 35 · 35 · 34 ×4 · 33 · 33 · fine-material 35 · 34 · 33 · 32 · 30 (×3) · immaterial 30 ×12 · supramundane 36 ×8 | the rules, on the generator |\n| `profile121` | the same, with the supramundane at 36 · 35 · 34 · 33 · 33 (×8) | the rules, on the generator |\n| `occurrences89` · `absences89` | every per-factor count over 89: universals 89 · adhimokkha 78 · vīriya 73 · chanda 69 (without: 11 · 16 · 20) · unwholesome 12 · 8 · 4 · 4 · 2 · 5 · 1 · beautiful universals 59 · abstinences 16 · illimitables 28 · paññā 47 | the rules, on the generator |\n| `occurrences121` · `absences121` | the factors counted over 121: vitakka 55 · vicāra 66 · pīti 51 (without: 66 · 55 · 70) | the rules, on the generator |\n| `feelings121` | ch. 3: pleasure 1 · pain 1 · displeasure 2 · joy 62 · equanimity 55 | the generator's feeling axis only |\n| `roots89` | ch. 3: rootless 18 · one-rooted 2 · two-rooted 22 · three-rooted 47 | the generator's root and knowledge axes only |\n| `keci_reading` | the dissent recorded at §30 gives 20, not 28 (§6) | an alternative rule |\n\n**Every figure matched on the first collation against the CST.**\n\nOne modelling choice stands between the key and the text. At the 89 reckoning each supramundane type is taken at the first jhāna (G9), so its row in `profile89` is the text's first-jhāna figure, 36 (§36: *Lokuttaresu tāva aṭṭhasu paṭhamajjhānikacittesu … chattiṃsa dhammā*). The figure is stated; its assignment to the 89 reckoning is ours.\n\n**The chapter 3 theorems test the generator, not the rules.** `feelings121` and `roots89` count values of the generator's feeling and root axes and never call the factor rules. They show that chapter 1's feeling and root labels, as encoded, reproduce chapter 3's tallies. They are therefore no evidence about whether the rules were fitted. **The cross-method test is the profile theorems**: the rules were read from the factor-indexed statements, and `profile89` and `profile121` check them against the type-indexed row sums, which they were not read from. The occurrence theorems, by contrast, check each rule against the sentence it was read from.\n\n### 5.2 · Collation\n\nThe key was built in three passes, and the order is part of the result:\n\n1. **From memory**, as rendered in Bhikkhu Bodhi's *Comprehensive Manual*. This pass contained the mixed-reckoning error of §2.2.\n2. **Against a printed edition**, the Pāli with Nārada's English. This pass exposed a separate hazard: a machine summary of that page gave the fifth supramundane jhāna as 32, where the text reads *\"thirty-six, thirty-five, thirty-four, and thirty-three in the last two.\"* **Every figure in the key is read from the raw text, never from a summary.**\n3. **Against the CST itself**, paragraph by paragraph. This pass found no disagreement with the second.\n\n### 5.3 · The breaks\n\nNine deliberate breaks, each applied by script and each verified to have actually changed its file:\n\n```\nbreak                                       expected           observed\n──────────────────────────────────────────  ─────────────────  ────────\nR4 without its doubt clause                 the check fails    FAILS\nR6 pīti allowed in the 4th jhāna            the check fails    FAILS\nR16 abstinences in functionals too          the check fails    FAILS\nkey: smile-producing 12 → 13                the check fails    FAILS\nkey: adhimokkha 78 → 77                     the check fails    FAILS\ngenerator: fine-material jhānas 1–4         the check fails    FAILS\ngenerator: smile-producing with equanimity  the check fails    FAILS\ncanon: chanda counted as named (§7)         the check fails    FAILS\nroots: knowledge adds no root               the check fails    FAILS\n```\n\nA checker that has only ever passed is an untested claim. These breaks are what let the passes be read as evidence. The technique is mutation testing (DeMillo, Lipton and Sayward 1978), applied here to a formalization of a textual classification, with faults seeded in the rules, the key and the generator. Its use inside a proof assistant is established: mCoq (Celik, Palmskog, Parovic, Gallego Arias and Gligoric, ASE 2019) mutates the definitions of Coq verification projects and re-checks the affected proofs, and reads a mutant whose proofs still pass as a sign that the specification is incomplete, which is the reading the breaks here are designed to exclude. Semantic mutation testing has also been carried to OWL ontologies, a classification's closest software form (Bartolini 2016; 2017). Only the application to a textual classification remains, and only that is claimed (§11, claim 4).\n\n---\n\n## 6 · Results II — A Dissent the Text Records\n\nAt §30, having located the illimitables in 28 types, the *Saṅgaha* adds:\n\n> *upekkhāsahagatesu panettha karuṇāmuditā na santīti keci vadanti.*\n>\n> Some say compassion and appreciative joy are absent here from the types accompanied by equanimity.\n\nFormalized as an alternative rule — karuṇā and muditā absent from the eight equanimous beautiful sense-sphere types (four wholesome, four functional) — the dissent gives **20**, not 28. The 28 sits in the same sentence the dissent qualifies, so **the *Saṅgaha* does not decide between the readings. Its ṭīkā, the *Abhidhammatthavibhāvinī*, argues for the main reading** (on §30, in the same CST file): in sense-sphere wholesome consciousness, the preliminary work for compassion and appreciative joy occurs with equanimous types too, by familiarity, \"like one reciting a well-known text\" (*yathā taṃ paguṇaganthaṃ sajjhāyantassa*), and it names the contrary view *kecivāda*.\n\nWhat the formalization adds is narrower: **the author's stated figures — the 28, and the 37 and 36 for the equanimous pairs (37 knowledge-associated, 36 dissociated) — encode the main reading, not the dissent.** A reader who adopts the *keci* reading must change the 28 and four per-type figures, not the 28 alone:\n\n```\nfigure (§30; per-type, §40–§41)         main reading   keci reading\n──────────────────────────────────────  ────────────   ────────────\nillimitables, number of types                28             20\nwholesome, equanimous, knowledge             37             35\nwholesome, equanimous, dissociated           36             34\nfunctional, equanimous, knowledge            34             32\nfunctional, equanimous, dissociated          33             31\n```\n\nThe Vibhāvinī argues the reading; it does not count the cost. The cost in the source's own figures is what the table states.\n\n---\n\n## 7 · Results III — The Canon, the Commentary and the Manual\n\n### 7.1 · What the canon names\n\nThe Dhammasaṅgaṇī opens its first analysis of a wholesome type (§1) with a list of what is present *\"yasmiṃ samaye kāmāvacaraṃ kusalaṃ cittaṃ uppannaṃ hoti somanassasahagataṃ ñāṇasampayuttaṃ\"*. It runs from *phasso hoti, vedanā hoti* to *paggāho hoti, avikkhepo hoti*, and closes:\n\n> *ye vā pana tasmiṃ samaye aññepi atthi paṭiccasamuppannā arūpino dhammā — ime dhammā kusalā.*\n>\n> or whatever other conditioned immaterial phenomena there are at that time — these phenomena are wholesome.\n\nUnder the synonym map fixed in the pre-registration (Appendix A), the list gives **56 terms, 29 distinct factors, and none of** *chanda, adhimokkha, manasikāra, tatramajjhattatā, karuṇā, muditā* or the three abstinences. The first run reported two unmapped terms, *kāyujukatā* and *cittujukatā*. Both are sandhi forms (kāya + ujukatā), and the pre-registration required sandhi to be normalized before mapping; the script had not implemented that case. With it, zero terms are unmapped. The detector for the nine was itself tested on a control list naming four of them in inflected forms, and it caught all four.\n\nThe nearest modern analysis of this same list is Nyanaponika's *Abhidhamma Studies* (1949), which gives the Pāli of this paragraph and investigates it factor by factor. Its chapter 14, \"The Supplementary Factors (ye-vā-panaka, F57–65)\", already does the layering that §7.3 makes exact. It takes the open clause as admitting supplements, notes that the commentaries supply them and that the Aṭṭhasālinī finds them *\"in various passages of the suttas\"*, numbers the nine after the list's own terms, and records that all of them *\"are incorporated into the condensed and systematized version of the List of Dhammas found in the Visuddhimagga and the Abhidhammatthasaṅgaha.\"* The finding that the manual's set is the canon's list plus the commentary's nine is his, in print, in 1949. This paper supplies an instrument for it, not the finding.\n\n### 7.2 · The commentary had already counted it\n\nThe Aṭṭhasālinī, the fifth-century commentary on the Dhammasaṅgaṇī, counts the same list (*Yevāpanakavaṇṇanā*, on Dhs §1; PTS pp. 133–134 by the CST's page markers; `abh01a.att.xml` line 5081, the same paragraph as the quotation in §2.3):\n\n> *Iti phassādīni chappaññāsa yevāpanakavasena vuttāni navāti sabbānipi imasmiṃ dhammuddesavāre pañcasaṭṭhi dhammapadāni bhavanti … Aggahitaggahaṇena panettha phassapañcakaṃ, vitakko vicāro pīti cittekaggatā, pañcindriyāni, hiribalaṃ ottappabalanti dve balāni, alobho adosoti dve mūlāni, kāyapassaddhicittapassaddhiādayo dvādasa dhammāti samatiṃsa dhammā honti.*\n>\n> Thus the fifty-six beginning with contact, and the nine stated as \"whatever else,\" make sixty-five terms in this section … Taking only what has not already been taken: the contact-pentad, initial and sustained application, rapture, one-pointedness, the five faculties, the two powers of shame and dread, the two roots of non-greed and non-hate, and the twelve beginning with bodily and mental tranquillity — these are exactly thirty phenomena.\n\n**Thirty phenomena is consciousness plus 29 factors.** The synonym map this project fixed in advance performs the commentary's own *aggahitaggahaṇa*, \"taking only what has not already been taken,\" and it reproduced the commentary's count exactly. The 56 the script found matches the commentary's *chappaññāsa* as well.\n\nThis cuts in two directions, and both are reported. **The count is consistent with the commentary's own — not independent of it**, since the 52-factor scheme the map targets inherits the commentarial identifications, so a map built on it could hardly do otherwise. **And it removes any novelty from the count itself.** The canon's 29 is not a finding; it is the commentary's figure, reproduced by machine. What this paper adds, if anything, is only the machine check and the chain of evidence around it.\n\n### 7.3 · The nine are first named by the commentary, and the manual inherits them\n\nThe same commentary names the nine (*Yevāpanakavaṇṇanā*, on Dhs §1; PTS p. 132 by the CST's page marker; `abh01a.att.xml` line 5045):\n\n> *yevāpanakavasena aparepi nava dhamme dhammarājā dīpeti. Tesu tesu hi suttapadesu 'chando adhimokkho manasikāro tatramajjhattatā karuṇā muditā kāyaduccaritavirati vacīduccaritavirati micchājīvaviratī'ti ime nava dhammā paññāyanti.*\n>\n> Under \"whatever else,\" the King of Dhamma shows nine further phenomena. For in various sutta passages these nine are known: zeal, decision, attention, neutrality, compassion, appreciative joy, and abstinence from bodily, verbal and livelihood misconduct.\n\nThe formalization proves (`Canon.lean`) that **the *Saṅgaha*'s 38 for this type contains all 29 the canon names, and that what remains is exactly these nine.** In the pre-registration's editorial layers, that looks like \"the manual adds nine.\" **Chronologically it is not.** The nine are first named for this list by the commentary — which presents them as the Buddha's (*dhammarājā dīpeti*), drawn from sutta passages — five to seven centuries before the manual. The manual inherits them as **members**, and arranges them in its own classes: manasikāra among the universals, chanda and adhimokkha among the occasionals, tatramajjhattatā among the beautiful universals, the abstinences and the illimitables as classes of their own.\n\n```\n  c. 3rd c. BCE             c. 5th c. CE                   c. 10th–12th c. CE\n  ─────────────────────     ──────────────────────────     ──────────────────────────\n  CANON                     COMMENTARY                     MANUAL\n  Dhammasaṅgaṇī §1          Aṭṭhasālinī                    Abhidhammattha-saṅgaha\n  56 terms, 29 factors      counts the 56 as 30            states 38 for this type,\n  + an OPEN clause:   ───▶  (samatiṃsa, = 29 + citta)      \"entering into combination\"\n  \"ye vā pana …\"            and NAMES the nine      ───▶   = 29 + the nine,\n                            (as the dhammarājā's,          inherited as members,\n                            from the suttas)               re-classed\n                                                           ⊇ never contradicts the canon\n```\n\nThe *Saṅgaha* inherits the nine as members of the list. It adds none of them, and it contradicts nothing the canon names. The inheritance itself is Nyanaponika's finding (1949, ch. 14; §7.1). The formalization's contribution is to make it **exact and checkable**: 29 + 9 = 38, with every one of the 29 inside the 38, under a synonym map fixed before the text was read.\n\n### 7.4 · Scope of \"the canon leaves nine unnamed\"\n\nThe claim holds **for this type's list only.** The Dhammasaṅgaṇī names the three abstinences as path factors in its supramundane type. The commentary notes that there they are not supplied as \"whatever else,\" *pāḷiyaṃ āgatattā* (\"because they have come in the text\"; *Lokuttarakusalavaṇṇanā*, PTS p. 217 by the CST's page marker; `abh01a.att.xml` line 6621). The Dhammasaṅgaṇī also names karuṇā and muditā as qualifiers of jhāna (§§258–261). Across the whole book, *adhimokkha* and *tatramajjhattatā* never appear. *Chanda* appears as a qualifier of consciousness types, *chandādhipateyya*, \"with zeal dominant\" (108 occurrences), much as karuṇā and muditā qualify jhāna, but never as a listed factor. *Manasikāra* appears only in other senses (*manasikārakusalatā*, *amanasikārā*).\n\n---\n\n## 8 · Results IV — The Cambodian Edition Against the CST\n\n### 8.1 · The comparison\n\nThe Cambodian Buddhist Institute edition prints the Pāli with a facing Khmer translation. Its volume 78 opens the Dhammasaṅgaṇī. The list of §1 stands on its Pāli pages 16–17 (the transcription's own heading reads *padabhājanīyaṃ*). The working transcription of those pages was transliterated from Khmer script to romanized Pāli by `canon/khmer_pali.py`, then aligned with edition C, position by position, by `canon/compare.py` (both at commit `c9136e1`). The output is the standard romanized Pāli of the CST. The nearest published standard for the mapping is the ALA-LC Khmer romanization table, whose Indic letter values were built for Pāli and Sanskrit loans (ISO 11940 is the Thai-script analogue); the script's own mapping is not claimed to follow either. The paper prints romanized forms only.\n\nCollating a Khmer-script witness of the Pāli canon against others is long-established work, and this section repeats none of it. The Sixth Council (1954–56), whose text the CST is, collated the script editions of its day, the Cambodian among them; its apparatus carries a Cambodian siglum (*kaṃ.*, 35 times in the file of the Dīgha's first volume), though none occurs anywhere in the Dhammasaṅgaṇī file, so at this passage the CST records no Cambodian reading. The Dhammachai Tipiṭaka Project (from 2010) builds a critical edition from palm-leaf manuscripts in four script traditions, with a working group for the Khom (Khmer) script. Srisetthaworakul (2019) compares Khom-script manuscripts of the Majjhimanikāya found in Thailand and in Cambodia. What §8 adds is narrow: the modern Buddhist Institute printed edition aligned term by term with the CST on one list, by a tested transliterator, with each difference classified only after the printed page and the two apparatuses were read.\n\n**The transliterator's control.** Before use it scored 10 of 10 on these known words, chosen to cover conjuncts, the niggahita and the vowel sign Khmer Pāli uses for *iṃ* (U+17B9, KHMER VOWEL SIGN Y): *phasso · vedanā · yasmiṃ · samaye · kāmāvacaraṃ · cittassekaggatā · saddhindriyaṃ · uppannaṃ · amoho · ñāṇasampayuttaṃ*.\n\n**The identity criterion.** Both lists are normalized the same way (Unicode NFC, lower case, final *-ṅ* to *-ṃ*, the pre-registered sandhi rule) and aligned as sequences. A position is IDENTICAL if the normalized terms are equal; CLOSE if they differ by an edit distance of at most 3 and a similarity ratio of at least 0.7 (the same term spelled differently, undecided until the printed page is read); otherwise DIFFERENT. Unmatched terms are reported as added or dropped. A CLOSE position is classified only after the printed page is checked.\n\n**Result: C 56 terms · K 56 terms · 48 identical · 8 spelled differently · 0 added, dropped or reordered.**\n\n### 8.2 · The eight differences, each checked against the printed page\n\nThe CST carries its own variant apparatus at Dhs §1, and it bears on these rows. It gives *viriyindriyaṃ* and *viriyabalaṃ* as the Sinhala and Thai readings (*sī. syā.*), *kāyappassaddhi* and *cittappassaddhi* as the Thai reading (*syā.*), and *kāyujjukatā*, *cittujjukatā* for two further witnesses (*sī. ka.*). It has no note on *hirī*.\n\n| Pos. | C | K as transcribed | K as printed | CST apparatus | Classification |\n|---|---|---|---|---|---|\n| 37 | *hirī* | *hiri* | *hiri* | — | **edition orthography**, not documented in either apparatus |\n| 39 | *kāyapassaddhi* | *kāyappassaddhi* | *kāyappassaddhi*⁽¹⁾ | *kāyappassaddhi* (Thai) | **edition orthography, recorded in K's apparatus; agrees with the Thai reading** |\n| 40 | *cittapassaddhi* | *cittappasaddhi* | *cittappassaddhi*⁽²⁾ | *cittappassaddhi* (Thai) | edition orthography (*pp*), recorded in K's apparatus, agrees with the Thai reading; the single *s* is a transcription slip |\n| 10 | *cittassekaggatā* | *cittaspekaggatā* | *cittassekaggatā* | — | transcription slip |\n| 17 | *somanassindriyaṃ* | *somanassidṭhiyaṃ* | *somanassindriyaṃ* | — | transcription slip |\n| 47 | *kāyapāguññatā* | *kāyapātuññatā* | *kāyapāguññatā* | — | transcription slip |\n| 12, 25 | *vīriya-* | *viriya-* | *vīriya-* (long ī, read at moderate confidence) | *viriya-* (Sinhala, Thai) | **edition orthography or slip, undecided** — the short *i* is a real reading elsewhere; to be read from the print |\n\nThe Cambodian edition carries its own variant apparatus at the foot of the page: *\"1 o. ma. kāyapassaddhi · 2 o. ma. cittapassaddhi\"*. We read *o. ma.* as the Burmese (*Marammā*) reading, with a single *p*. The edition's own key to its abbreviations was not consulted, so that reading of the siglum is ours; it is consistent with the CST, whose text (the Sixth Council's, held in Burma) has the single *p* and whose apparatus attributes *-pp-* to the Thai edition. **Of the three orthographic differences, the passaddhi pair is documented in K's own apparatus against the Burmese reading, and agrees with the Thai; *hiri*/*hirī* is documented in neither.** The long-or-short *ī* of positions 12 and 25 stays open until the print is read at higher confidence.\n\n### 8.3 · Two artifacts of the working text, reported and handled\n\n- **Subscript order.** Some words carry the subscript RO typed before another subscript (the code-point sequence NO, COENG, RO, COENG, TO where the correct order is NO, COENG, TO, COENG, RO), which transliterates as *-inrdiyaṃ*. This is an encoding matter, not a textual one, and the transliterator normalizes it.\n- **Page furniture.** At the page break, the running header of page 17 fell between *paññābalaṃ* and its *hoti*, and first produced a false DIFFERENT. Page numbers and running headers are now stripped before alignment.\n\n### 8.4 · What the Cambodian result does and does not show\n\nIt shows that, for the densest list in the Dhammasaṅgaṇī, **the two editions agree on every term and its order**, and that their spelling differences are orthographic, and in two of three cases documented. It does not show agreement across the book, and it is not a collation in the sense of the projects named in §8.1: it uses one printed edition, not manuscripts, and one list. It rests on one passage of a working transcription, checked against the printed scan only where the two editions differed.\n\n---\n\n## 9 · What the Check Establishes — and the Refutation It Survived\n\nBefore this paper was drafted, its thesis was handed to a separate reviewer with no authority to rescue it and instructions to find the case that kills it. Its findings, with the author's disposition:\n\n| Finding | Severity | Disposition |\n|---|---|---|\n| The nine are first named by the commentary, which predates the manual; \"the manual adds nine\" misattributes them | narrows, bordering on kills, for that wording | **accepted**: §7.3 now states inheritance; the project's results file carries an appended correction |\n| The canon's 29 was counted by the commentary itself (*samatiṃsa*) | narrows novelty; consistent with the count | **accepted**: §7.2 reports both directions |\n| The 38 is a combination count; at most 34 factors are co-present | narrows, if the paper called them co-present | **accepted**: §2.3 |\n| Several rules restate the text's own sentences; \"general\" is a test of syntax | narrows \"general\" and the ratio | **accepted**: §4.3 reports each rule's extension and marks R4, R16 and R17 |\n| \"The text decides the dissent\" overstates | contrast | **accepted**: §6 |\n| \"The canon leaves nine unnamed\" holds for this type only | narrows | **accepted**: §7.4 |\n| No computational precedent found (English, 2026-09-27) | contrast | **superseded** by the full census of 2026-09-28, which found one at generator width (§4.2) |\n\nThe full census then ran the hunt a second time, aimed at the claims rather than the thesis, and asked of each part whether it already existed, if not as software then in some other form. Four answers were yes. **The consistency test exists as a chart**: Bodhi's Table 2.4 is a hand-made incidence matrix whose margins exhibit that the two methods agree, and the tradition teaches them as two views of one structure (§2.1). **The generator exists**: a Haskell model enumerated the 89 and 121 from the same axes nine days before this one (§4.2). **The layering exists**: Nyanaponika stated it in 1949 (§7.1). **Edition collation with Khmer-script witnesses exists**: the Sixth Council, the Dhammachai Tipiṭaka Project and Srisetthaworakul (§8.1). What survived is not *checking that the two methods agree*. It is checking it by rules over axes that name no type, by kernel evaluation, against a key cited to CST paragraphs, in the text's mixed reckonings, with the breaks and the ratio; and the two one-list checks. §11 is narrowed to that.\n\n**What survives, stated once:** a single rule set of 27 clauses shows the *Saṅgaha*'s factor-indexed statements and its type-indexed statements to be **mutually consistent**, by kernel evaluation, against the CST, and the generator's axes reproduce chapter 3's feeling and root counts. The generator's enumeration of the 89 and 121 is reported, not claimed (§4.2). The layer facts of §7 are proved in Lean; the edition facts of §8 are script-aligned and checked by hand against the printed page. **It is a consistency result, not a derivation of the *Abhidhamma* from first principles.** The author of the rules knew the tables, and the generator restates the text's axes.\n\n---\n\n## 10 · Pre-Registered Predictions and Their Outcomes\n\n| ID | Prediction (as registered) | Conf. | Outcome |\n|---|---|---|---|\n| P-FA1 | At the manual layer, edition C: general rules generate exactly 89, and 121 with the five-jhāna expansion | 0.6 | **confirmed**: the generator yields 89/121 and the rules reproduce every stated count |\n| P-FA1b | at least 3 exception rules naming a narrow class | 0.7 | **retired, unscored**: its test was never defined, and the data have been seen |\n| P-FA2 | at the canon layer, the rules are under-determined or differently partitioned | 0.65 | operationalized as P-FA2a (below) |\n| P-FA2a | the canon's list for the first type names exactly 29 factors, none of the nine, and closes with an open clause | 0.75 | **confirmed**; consistent with the commentary's *samatiṃsa* |\n| P-FA2b | no term in that list falls outside the pre-fixed map | 0.6 | **confirmed** after the pre-registered sandhi rule (first run: 2 unmapped, disclosed) |\n| P-FA3 | editions K and C agree at the manual layer | 0.85 | **cannot run as registered**: the *Saṅgaha* is not part of the Khmer Tipiṭaka, and the one Khmer-script copy held reproduces the CST's paragraph structure in all nine chapters |\n| P-FA4 | compression below 0.5 | 0.5 | **confirmed**: 0.22 |\n| P-FA5 | K and C name the same terms in the same order for the first type's list | 0.8 | **confirmed**: 56 = 56, 0 added, dropped or reordered |\n\nThe predictions are recorded in the public prediction register in its *Instrumented but outside the core* section. Their subject is the internal consistency of a text, not the direction value moves.\n\n---\n\n## 11 · Enumerated Claims\n\nThe following are disclosed and dedicated to the public domain. Together they are the survivor of the full census of 2026-09-28 (Prior-Art Statement): *a class-level rule set over the Abhidhammattha-saṅgaha's own classification axes, naming no type, that reproduces every per-type and per-factor count of its chapter 2 (and, through the axes, chapter 3's feeling and root tallies) by kernel evaluation in a proof assistant against a key cited paragraph by paragraph to the Chaṭṭha Saṅgāyana, in the text's own mixed reckonings, reported with its compression ratio and with seeded breaks each required to fail — plus a pre-registered, machine-checked inclusion of one canonical list and an alignment of that list between the Cambodian printed edition and the CST; the enumeration of the 89/121 from the axes is not part of it.* Each claim below is one part of that composition, disclosed as a part of it and narrowed to the census's wording. None is claimed apart from it, and each names its nearest prior art.\n\n1. **A consistency test** for a classification that its source states twice, once per feature and once per class. The class–feature incidence is written as constraints over classification axes, none naming an individual class. The source's stated projections are kept in a separate key cited to paragraphs of a named edition. A proof assistant checks, by kernel evaluation, that the constraints reproduce every stated projection, in the source's own mixed bases of reckoning. *Nearest prior:* the existence problem for a 0-1 matrix with given margins (Gale 1957; Ryser 1957); constraint validation of classifications (W3C SHACL over SKOS; DELTA character matrices); and, for this source, the printed chart that shows both methods together with totals in both reckonings (Bodhi 1993, Tables 2.2–2.4). None checks constraints over axes against a paragraph-cited key by kernel evaluation; only that composition is claimed.\n2. **Its application to the Theravāda *Abhidhamma***. The enumeration of the 89 and 121 types from ch. 1's axes is not claimed (nearest prior: `AloofBuddha/abhidhamma`, a Haskell model, first commit 2026-09-18, which enumerates both reckonings from the same axes, checks ch. 1's arithmetic and some ch. 3 tallies, and states that it encodes no factor-combination rules). Claimed is the 18-clause rule set over those axes reproducing every per-type and per-factor count of ch. 2, and the full ch. 3 feeling and root tallies from the axes, by kernel evaluation against the CST-cited key. Each rule's extension is reported, and the rules that restate a sentence of the text are marked (§4.3).\n3. **The reporting of a compression ratio** (rule clauses ÷ generated classes) for such a formalization, as a crude description-length proxy for how far a stated classification is generated by rule rather than listed, comparable only across encodings on the same axes. *Nearest prior:* minimum description length (Rissanen 1978); implication bases in formal concept analysis (Ganter and Wille 1999). Only the application is claimed.\n4. **Mutation testing applied to a formalization of a textual classification**: deliberate faults seeded in the rules, the key and the generator, each applied by script, each verified to have changed its file, each required to make the check fail. *Nearest prior:* mutation testing (DeMillo, Lipton and Sayward 1978); mutation analysis of Coq verification projects (mCoq; Celik et al. 2019), where a surviving mutant signals an incomplete specification; semantic mutation testing of OWL ontologies (Bartolini 2016; 2017). Only the application to a textual classification is claimed.\n5. **The per-layer inclusion check of one canonical list**: the list mapped onto the later system's features through a synonym map fixed and published before the text is read (Appendix A), with unmapped terms reported rather than absorbed. The result is checked in the proof assistant as inclusion (every canonically named feature is in the later set) plus an exact remainder, which is then attributed to its source layer by reading, not by the editorial order. *Nearest prior:* the commentary's own count of the same list (*samatiṃsa*); Nyanaponika (1949, ch. 14), who already sets the list's terms against the nine the commentaries supply and records their incorporation into the *Saṅgaha*. Only the instrument is claimed: the synonym map fixed before reading, and the machine-checked inclusion with its exact remainder.\n6. **The per-edition alignment of one canonical list** across a Khmer-script and a romanized edition, as a composition: rule-based transliteration tested on a control; normalization of encoding order and page furniture; term-by-term alignment; and classification of each difference as orthographic or as a transcription slip only after checking the printed page and the editions' own apparatus. *Nearest prior:* Aksharamukha; CollateX; ALA-LC Khmer romanization; the Sixth Council's collation of the script editions, the Cambodian included; the Dhammachai Tipiṭaka Project's Khom-script witnesses and Srisetthaworakul (2019); the CST apparatus. Only the composition, on one list of the Buddhist Institute printed edition against the CST, is claimed.\n7. **Counting the cost of a recorded dissent in the source's own stated figures**: the dissent (*keci vadanti*) formalized as an alternative rule, and every stated figure it would move listed. *Nearest prior:* the ṭīkā's own argument for the main reading (the Vibhāvinī on §30), which argues the reading but does not count its cost, nor do the modern manuals count it (Nārada; Bodhi 1993, §15).\n\n---\n\n## 12 · Why a Value Substrate Should Carry Its Own Tests\n\n*Tier: argument, applied to this institution's use of the canon. The deletion test passes: remove this section and §§2–11 stand.*\n\nAn autonomous agent's values will be specified in text, and text contradicts itself quietly. The practical difficulty is not that value specifications are wrong. It is that they rarely state anything twice, so a reader has little to check them against. The *Abhidhamma* is a counter-example to that condition. It states its classification twice and counts both statements. That is what a unit test is. This paper shows that one chapter's tests pass, and it shows what passing does and does not mean: consistency, not truth; a combination count, not a moment. The check concerns a classification of mind, not what a mind should do.\n\nFor Miss Aquarius, whose grounding the series describes as compiled from this canon, the consequence is modest and concrete. **The parts of the substrate that can be checked should be checked, and the checks should run again whenever an edition, a transcription or a formalization changes.** A source whose own claims are machine-verified can be inherited with more confidence than one whose claims are only repeated.\n\n---\n\n## 13 · Honest Limits\n\n- **This is a consistency check, not a derivation.** The rules were written by an author who knew the tables. Three rules restate the text: R4 nearly verbatim, and R16 and R17 transcribing locus sentences. The generator restates chapter 1's axes. The evidence that the rules are not merely fitted is the profile theorems: row sums the rules were not read from. That is one test, over one chapter.\n- **The key holds only stated totals.** The occurrence theorems check each rule against the sentence it was read from; the cross-method test is the profile theorems (row sums against rules read from column statements). No theorem checks a cell. Agreement at the cells rests on the rules' reading of the sampayoga sentences, which a reader checks by hand. Distinct 121 × 52 matrices can share every projection checked here.\n- **The chapter 3 theorems test the generator, not the rules.** They never call the factor rules (§5.1).\n- **The rule set is not unique.** Other rule sets, and other synonym maps, may reproduce the same counts. The result is that one such set exists and is short, not that it is the text's.\n- **Every theorem is a finite evaluation.** The domain is 121 × 52, so a script would compute the same numbers. The kernel adds a trusted evaluator and a checked artifact, not deductive depth.\n- **The compression ratio depends on the encoding** — on what counts as a clause and on which distinctions are carried by datatypes (§4.4).\n- **The pre-registration proves order, not ignorance.** The author knew the *Saṅgaha*'s tables; three of the results reported were not predicted at all (§3.1).\n- **The counts are combination counts.** Nothing here states how many factors are present in any actual moment of consciousness. By the commentary's count (at most one of the five variable factors at a moment), that figure is at most 34 for the first wholesome type; the 34 is derived here.\n- **The layer model was mis-specified in the pre-registration.** It ordered the layers editorially (canon ⊂ manual ⊂ commentaries), and the commentary is older than the manual. The claim that survives is about inheritance, established by reading, not by the registered order.\n- **The layer and edition results each rest on one list.** Neither the inclusion check nor the edition alignment has been run beyond the first type's list, and the counts check itself has been run on one edition only.\n- **P-FA1b was never testable** as registered, and it is retired without a score. **P-FA3 could not be run** as registered.\n- **The Cambodian comparison covers one passage**, read from a working transcription. The long *ī* of *vīriya-* at two positions was read from the scan at moderate confidence, and the short *i* is a real reading in the Sinhala and Thai witnesses, so those two positions are undecided. The siglum *o. ma.* was read as Burmese without consulting the edition's abbreviation key.\n- **The CST's recension** is the Burmese tradition's; the Cambodian edition's apparatus cites the Burmese reading. Two editions that share ancestry are not independent witnesses in the textual-critical sense.\n- **The census is bounded by its aperture.** The full census of 2026-09-28 searched eight languages besides English, but the engine is English-weighted, Myanmar's Abhidhamma pedagogy lives largely offline, and print-only books were outside it; its non-English nulls are weak. The Burmese ṭīkā debate on chapter 2 (Ledi Sayadaw's *Paramatthadīpanī*, the *Aṅkura-ṭīkā*) was not read, and it is the likeliest place for a counting of the §30 dissent, so claim 7's *not found* is the weakest in the paper. The census's own predictions missed in both directions: it expected the generator's neighbour in Burmese or Thai teaching software and found it in an English Haskell repository, and it did not predict Bodhi's printed chart as a neighbour of the consistency test itself.\n- **The proof assistant verifies that conclusions follow from definitions.** It verifies nothing about the definitions' fidelity to the Pāli beyond what the key's citations let a reader check by hand.\n- **Nothing here bears on physics.** The project that produced this paper also keeps a lens comparing the *Abhidhamma*'s discrete moments with discrete-spacetime programmes. That lens is not a claim, and it is not used here.\n\n---\n\n## 14 · Lineage\n\nThe commentators were already doing this. When the Aṭṭhasālinī counts the fifty-six terms of the first list, adds the nine it supplies, and then counts again \"taking only what has not already been taken\" to arrive at thirty, it is performing the operation this paper performs by machine: projecting an enumeration onto a smaller set of distinct phenomena and stating the size. When the *Saṅgaha* states its combinations twice, once by factor and once by type, it gives its readers the means to check one statement against the other. For centuries that check was run by students in the monastery, with the *Saṅgaha*'s verses by heart and its tables in their hands. **The counts check is that recitation, run by a different reader.** It passes where the students' check would pass, for the reasons they would give.\n\n---\n\n## Appendix A · The Pre-Registered Synonym Map\n\nThe map below is `PREREGISTRATION-2026-09-27b.md` §3 (commit `f4a26e5`, pushed before either edition of the Dhammasaṅgaṇī was read), reproduced without change. It is also copied verbatim into `canon/pada.py`. Each canonical term maps to one of the 52 factors, or to CITTA (consciousness itself, not a factor). A term not in the map is reported as UNMAPPED and is never mapped after the fact. The pre-registration's normalization rule, quoted: *\"Grammatical variants (case, sandhi, -ṃ/-ṅ, a trailing hoti) are normalized before mapping.\"*\n\n| Canonical term(s) | → |\n|---|---|\n| phasso | phassa |\n| vedanā · sukhaṃ · somanassaṃ · somanassindriyaṃ | vedanā |\n| saññā | saññā |\n| cetanā | cetanā |\n| cittaṃ · viññāṇaṃ · mano · mānasaṃ · hadayaṃ · paṇḍaraṃ · manāyatanaṃ · manindriyaṃ · viññāṇakkhandho · tajjāmanoviññāṇadhātu | CITTA |\n| vitakko · sammāsaṅkappo | vitakka |\n| vicāro | vicāra |\n| pīti | pīti |\n| cittassekaggatā · samādhindriyaṃ · samādhibalaṃ · sammāsamādhi · samatho · avikkhepo | ekaggatā |\n| saddhā · saddhindriyaṃ · saddhābalaṃ | saddhā |\n| vīriyaṃ · vīriyindriyaṃ · vīriyabalaṃ · sammāvāyāmo · paggāho | vīriya |\n| sati · satindriyaṃ · satibalaṃ · sammāsati | sati |\n| paññā · paññindriyaṃ · paññābalaṃ · amoho · sammādiṭṭhi · sampajaññaṃ · vipassanā | paññā |\n| jīvitindriyaṃ | jīvitindriya |\n| hirī · hiribalaṃ | hiri |\n| ottappaṃ · ottappabalaṃ | ottappa |\n| alobho · anabhijjhā | alobha |\n| adoso · abyāpādo | adosa |\n| kāya-/citta- passaddhi · lahutā · mudutā · kammaññatā · pāguññatā · ujukatā (12 terms) | the 12 paired beautiful factors, each to its own |\n\nNineteen rows: 18 factor targets plus CITTA, where the last row stands for twelve factors. So the map can name at most 29 distinct factors (17 + 12), and P-FA2a predicted that the list would name all of them and nothing else. The sandhi case the first run missed (*kāyujukatā*, *cittujukatā* for *kāya-ujukatā*, *citta-ujukatā*) falls under the quoted rule; its implementation in `pada.py` is commented as added after the first run.\n\n---\n\n## Cross-Venue References\n\n| Venue | Identifier |\n|---|---|\n| Primary canonical | <https://thonly.org/research/the-counts-check> |\n| GitHub (paper) | <https://github.com/thonly/publications/blob/main/defensive-publications/the-counts-check.md> |\n| GitHub (formalization, CC0) | <https://github.com/SiliconWat/formal-abhidhamma> (files cited at commit `c9136e1`) |\n| Pre-registrations | `PREREGISTRATION.md` (commit `259e5e8`) · `PREREGISTRATION-2026-09-27b.md` (commit `f4a26e5`) |\n| Census pre-registration (public) | `thonly/publications` commit `9bf6a3d`, `timestamps/census/counts-check-full-prereg.md` |\n| Results and corrections | `RESULTS.md` (commits `b9070d4`, `c5c77fb`, `ada7a20`, `74e19fe`, `c9136e1`) |\n| Prediction register | <https://thonly.org/research/program/register> (P-FA1 … P-FA5) |\n| Internet Archive | <https://web.archive.org/web/2026*/thonly.org/research/the-counts-check> |\n\n---\n\n## Acknowledgments\n\nThe author thanks his father, with whom the transcription of the Cambodian Tipiṭaka proceeds page by page, and whose transcription of volume 78 made §8 possible; the corrections §8.2 records are offered to that work, not against it. Thanks go to the Buddhist Institute, Phnom Penh, whose edition this is; to the Vipassana Research Institute for the Chaṭṭha Saṅgāyana texts it gives away; to the donors of the scanned volumes (Kang Huychan and family, Lok Kru Aggapaṇḍita But Savong, Srong Chanda, and 5000-years.org), whose dedication travels with any use of them; to Bhikkhu Bodhi and U Nārada, whose editions of the *Saṅgaha* made the first two collations possible; and to the Lean community. The formalization, collation and drafting were done with Miss Aquarius, the name under which this institution discloses its AI collaboration.\n\n---\n\n## Citations\n\n1. Anuruddha. *Abhidhammatthasaṅgaha*. Chaṭṭha Saṅgāyana edition, Vipassana Research Institute, `github.com/VipassanaTech/tipitaka-xml` (commit `0854d11`), `romn/abh07t.nrf.xml` (ch. 2 §12–§58; ch. 3 §3–§17). The same file carries the *Abhidhammatthavibhāvinī-ṭīkā* after the *Saṅgaha* (cited here on ch. 2 §30).\n2. *Dhammasaṅgaṇī*. Chaṭṭha Saṅgāyana edition, `romn/abh01m.mul.xml` (§1, with its variant apparatus; §§258–262).\n3. Buddhaghosa (attrib.). *Aṭṭhasālinī*. Chaṭṭha Saṅgāyana edition, `romn/abh01a.att.xml`: *Yevāpanakavaṇṇanā* on Dhs §1 (PTS pp. 132–134 by the CST's page markers; lines 5045, 5081) and *Lokuttarakusalavaṇṇanā* (PTS p. 217; line 6621).\n4. *Braḥ Traipiṭaka* (the Cambodian Tipiṭaka), Buddhist Institute, Phnom Penh; vol. 78, *Abhidhammapiṭaka, Dhammasaṅgaṇī*, part 1, pp. 16–17.\n5. Bodhi, Bhikkhu (ed.) (1993/2000). *A Comprehensive Manual of Abhidhamma*. Buddhist Publication Society. (Its introduction is the source of the datings in §3.5. Ch. 2, Tables 2.2 (association), 2.3 (combination) and 2.4 (both together), with the Guide to §30 in its numbering, are the nearest prior art for the consistency test; its §15 reports the *keci vadanti* dissent without counting it.)\n6. Nārada Mahā Thera (trans.). *A Manual of Abhidhamma*. Buddhist Missionary Society; the Pāli-and-English text consulted as mirrored at `palikanon.com`.\n7. Karunadasa, Y. (2010). *The Theravāda Abhidhamma*. Centre of Buddhist Studies, University of Hong Kong.\n8. de Moura, L. & Ullrich, S. (2021). \"The Lean 4 Theorem Prover and Programming Language.\" *CADE-28*.\n9. takuya50 / Nimble Ariake (2026). *buddhist-comparative-logic*, v0.1.0. `github.com/takuya50/buddhist-comparative-logic`; software DOI 10.5281/zenodo.22851717. (The nearest prior art: machine-checked Buddhist formalizations of Mahāyāna and pramāṇa texts.)\n10. PJ-Oliveira. *abhidhamma*. `github.com/PJ-Oliveira/abhidhamma`. (Interactive citta-vīthi simulator, cetasika analyzer and Paṭṭhāna matrix.)\n11. *The Ephemeral Projector — A Comparative Analysis of Quantum Mechanics and Abhidhamma*. ResearchGate 397427972. (The nearest physics-comparison neighbour; not used here.)\n12. Ly, T. & Miss Aquarius (2026). \"The Wheel That Unwinds the Wheel: The Abhidhamma as Executable Process-Specification.\" *The Abhidhamma Compiled* №1. thonly.org. (§4.2 is the claim this paper tests.)\n13. Ly, T. & Miss Aquarius (2026). \"Twenty-Four Kinds of Because: The Paṭṭhāna as a Typed-Causation Vocabulary for AI Alignment.\" *The Abhidhamma Compiled* №2. thonly.org.\n14. Ly, T. & Miss Aquarius (2026). \"One of One: Individuation Without Essence.\" *The Abhidhamma Compiled* №3. thonly.org.\n15. Ly, T. & Miss Aquarius (2026). *The Prediction Register*. thonly.org/research/program/register (P-FA1 … P-FA5).\n16. Gale, D. (1957). \"A theorem on flows in networks.\" *Pacific Journal of Mathematics* 7(2), 1073–1082.\n17. Ryser, H. J. (1957). \"Combinatorial properties of matrices of zeros and ones.\" *Canadian Journal of Mathematics* 9, 371–377.\n18. DeMillo, R. A., Lipton, R. J. & Sayward, F. G. (1978). \"Hints on Test Data Selection: Help for the Practicing Programmer.\" *Computer* 11(4), 34–41.\n19. Rissanen, J. (1978). \"Modeling by shortest data description.\" *Automatica* 14(5), 465–471.\n20. Ganter, B. & Wille, R. (1999). *Formal Concept Analysis: Mathematical Foundations*. Springer.\n21. Rajan, V. *Aksharamukha* (script converter, including Khmer). `github.com/virtualvinodh/aksharamukha`.\n22. Haentjens Dekker, R., Middell, G. et al. *CollateX* (Interedition). `collatex.net`.\n23. Nyanaponika Thera (1949). *Abhidhamma Studies: Researches in Buddhist Psychology*. Colombo: Frewin. Later ed., *Abhidhamma Studies: Buddhist Explorations of Consciousness and Time*, ed. Bhikkhu Bodhi, Wisdom Publications, 1998.\n</content>\n</invoke>\n24. AloofBuddha. *abhidhamma: a Haskell model of Theravāda Abhidhamma consciousness and dependent origination*. `github.com/AloofBuddha/abhidhamma` (repository created and first commit `026819b`, 2026-09-18; no licence file). (The nearest prior art for the enumeration of the 89 and 121: `allCittas89` and `allCittas121` in `src/Abhidhamma/Cittasangaha.hs`, checks in `test/Check.hs`; its stated \"Known gap\" on factor combination is in `src/Abhidhamma/Citta.hs`.)\n25. Celik, A., Palmskog, K., Parovic, M., Gallego Arias, E. J. & Gligoric, M. (2019). \"Mutation Analysis for Coq.\" *34th IEEE/ACM International Conference on Automated Software Engineering (ASE 2019)*. Tool paper: Jain, K., Palmskog, K., Celik, A., Gallego Arias, E. J. & Gligoric, M. (2020). \"mCoq: Mutation Analysis for Coq Verification Projects.\" *ICSE 2020 Companion*, doi:10.1145/3377812.3382156.\n26. Bartolini, C. (2016). \"Mutating OWLs: Semantic Mutation Testing for Ontologies.\" *AMARETTO workshop, MODELSWARD 2016*.\n27. Bartolini, C. (2017). \"Software Testing Techniques Revisited for OWL Ontologies.\" In *Model-Driven Engineering and Software Development*, Communications in Computer and Information Science, Springer, pp. 132–153.\n28. W3C (2017). *Shapes Constraint Language (SHACL)*. W3C Recommendation, 20 July 2017. · W3C (2009). *SKOS Simple Knowledge Organization System Reference*. W3C Recommendation, 18 August 2009.\n29. Dallwitz, M. J. *DELTA: DEscription Language for TAxonomy*. CSIRO; adopted as a data standard by TDWG (Biodiversity Information Standards) in 1988. `tdwg.org/standards/delta`.\n30. Nyanaponika Thera (1949), as ref. 23: ch. 14, \"The Supplementary Factors (ye-vā-panaka, F57–65)\", and the list with its supplements numbered 1–65 in the chapter on the Dhammasaṅgaṇī's list. (The census control; the layering of §7 in print.)\n31. The Sixth Buddhist Council (*Chaṭṭha Saṅgāyana*), Yangon, 1954–56, whose text is edition C; its apparatus cites the Sinhala (*sī.*), Thai (*syā.*), Cambodian (*kaṃ.*) and PTS (*pī.*) editions.\n32. Dhammachai Tipiṭaka Project (2010–). A critical edition of the Pāli canon from palm-leaf manuscripts in the Sinhala, Burmese, Khom (Khmer) and Tham scripts. `dhammachaitipitaka.org`.\n33. Srisetthaworakul, S. (2019). \"Comparison of the Khom Script Manuscripts of the Majjhimanikāya Found in Thailand and Cambodia.\" *Journal of Indian and Buddhist Studies (Indogaku Bukkyōgaku Kenkyū)*.\n34. Library of Congress. *ALA-LC Romanization Tables: Khmer*. (The nearest romanization standard; ISO 11940 is its Thai-script analogue.)\n35. Kyaw, Pyi Phyo (2014). *Paṭṭhāna (Conditional Relations) in Burmese Buddhism*. PhD thesis, King's College London; ch. 5, the mathematics of the Paṭṭhāna and combinatorics. (A contrast for §4.4: an *Abhidhamma* enumeration read as combinatorics.)\n36. Nearer software neighbours named in the Prior-Art Statement: `suchanon456/Cyber-Abhidhamma` (first commit 2026-09-14); `juncoflockleader/Abhidharma` (2024); `dangerzig/tipitaka.critical` (2026). Patents recorded, claims not read: US8990746B1 (Cadence Design Systems); US20240419880A1 (IBM).",
+      "provenance": {
+        "sha256": "edc3fa58718e8792c3a798a67591509fddb383d0a5cc116a242d2f3476635795",
+        "doi": "10.5281/zenodo.23020671",
+        "concept_doi": "10.5281/zenodo.23020670",
+        "zenodo_url": "https://zenodo.org/records/23020671",
+        "deposited_sha256": "edc3fa58718e8792c3a798a67591509fddb383d0a5cc116a242d2f3476635795",
+        "deposited_matches_current": true,
+        "opentimestamps": true,
+        "source_url": "https://raw.githubusercontent.com/thonly/publications/main/defensive-publications/the-counts-check.md",
+        "sha256_covers": "the complete source file at source_url, including its metadata block — NOT the `text` field in this response",
+        "canonical_url": "thonly.org/research/the-counts-check",
+        "hostile_review": null
+      },
+      "bytes": 89762,
+      "words": 13719
+    },
     {
       "slug": "the-founder",
       "repo": "TH/publications",
@@ -6085,9 +6155,9 @@
   "manifest": {
     "what": "Every file the index build considered, and what became of it. candidates = served + excluded + held. Detects a document silently dropped; it is not a tamper check — each document's sha256 is.",
     "candidates": 151,
-    "served": 149,
+    "served": 151,
     "excluded": 0,
-    "held": 2,
+    "held": 0,
     "served_documents": [
       {
         "slug": "a-vow-you-could-have-refused",
@@ -6479,6 +6549,12 @@
         "path": "essays/love-and-freedom-dependent-origination.md",
         "sha256": "a83866c3f25461336ac652f37d0a03d3cb2cf40a496c8382fd1c80ad0c13ec90"
       },
+      {
+        "slug": "machine-dana-from-share-to-vow",
+        "repo": "TH/publications",
+        "path": "defensive-publications/machine-dana-from-share-to-vow.md",
+        "sha256": "8af30fe81b30b47549316d72d591532f1df22652fb013520ca2feb3be9f8019a"
+      },
       {
         "slug": "manufactured-universal-giving",
         "repo": "TH/publications",
@@ -6713,6 +6789,12 @@
         "path": "essays/the-capacity-to-give.md",
         "sha256": "a37a1ba53c7bfd66ffb17fbfd5511fd646fe58ab35e293b5c08d2a827d9aca57"
       },
+      {
+        "slug": "the-counts-check",
+        "repo": "TH/publications",
+        "path": "defensive-publications/the-counts-check.md",
+        "sha256": "edc3fa58718e8792c3a798a67591509fddb383d0a5cc116a242d2f3476635795"
+      },
       {
         "slug": "the-founder",
         "repo": "TH/publications",
@@ -6985,24 +7067,7 @@
       }
     ],
     "excluded_documents": [],
-    "held_documents": [
-      {
-        "slug": "machine-dana-from-share-to-vow",
-        "repo": "TH/publications",
-        "path": "defensive-publications/machine-dana-from-share-to-vow.md",
-        "reason": "polish r1 done and revised (2026-09-28); awaiting its FULL novelty census before the first deposit",
-        "held_since": "2026-09-27",
-        "review_by": "2026-10-27"
-      },
-      {
-        "slug": "the-counts-check",
-        "repo": "TH/publications",
-        "path": "defensive-publications/the-counts-check.md",
-        "reason": "polish r1 done and revised (2026-09-28); awaiting its FULL novelty census before the first deposit",
-        "held_since": "2026-09-27",
-        "review_by": "2026-10-27"
-      }
-    ]
+    "held_documents": []
   },
   "program": {
     "question": "Which Way Value Moves",

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@333eco/corpus",
-    "version": "2.5.24",
+    "version": "2.5.25",
     "description": "An MCP server for an open-licensed corpus, served with verifiable provenance — every document carries its sha256, DOI and OpenTimestamps proof so a consuming agent can check its own citation.",
     "license": "CC0-1.0",
     "author": "Thon Ly",