@event4u/agent-config 9.7.0 → 9.9.0

package/docs/personas.md

@@ -1,178 +0,0 @@
-# Personas
-
-Personas are **review lenses** the host agent adopts when running a
-skill — not sub-agents, not execution modes. Each persona shapes
-*what* the agent looks for in a diff, plan, or artifact; the host
-identity, tools, and workflow stay the same.
-
-This page catalogs the **29 active personas** shipped with
-`event4u/agent-config` (5 core + 24 specialists), plus the 5
-**advisor** personas in `personas/advisors/`. It also explains how
-`personas:` (lens axis) interacts with `/mode` (role-mode axis).
-Removed personas are deleted in-commit (no soak window) — see
-[`persona-governance § Deprecation path`](../dist/agent-src/rules/persona-governance.md).
-
-## Catalog — Core (always-loaded, 5)
-
-| ID | Role | Wing | Owner | Lens summary |
-|---|---|---|---|---|
-| `developer` | Developer | — | package | Pragmatic implementer voice — what's the smallest correct change? |
-| `senior-engineer` | Senior Engineer | — | package | Long-horizon impact — what does this look like in 6 months? |
-| `stakeholder` | Stakeholder | — | package | Risk and politics — who owns the rollback if this fails? |
-| `critical-challenger` | Critical Challenger | — | package | Devil's-advocate voice — what assumption is load-bearing here? |
-| `ai-agent` | AI Agent | — | package | Tool-economy voice — token cost, prompt clarity, automation seam. |
-
-Note: `product-owner` was reclassified to **specialist** at v2.0 — see below.
-
-## Catalog — Specialists (opt-in, 24 active)
-
-| ID | Role | Wing | Owner | Lens summary |
-|---|---|---|---|---|
-| `qa` | QA | — | package | Test coverage and regression gates. |
-| `backend-architect` | Backend Architect | — | package | Service-layer boundaries, transaction scope, contract changes. |
-| `eloquent-tamer` | Eloquent Tamer | — | package | N+1, query shape, ORM idioms that melt the database. |
-| `performance-engineer` | Performance Engineer (`performance`) | — | package | Latency, allocations, hot-path cost, complexity class at 100× load. |
-| `data-integrity` | Data Integrity Reviewer (`data`) | — | package | Migration reversibility, lock behavior under live traffic, data-loss on backfill. |
-| `production-validator` | Production Validator | — | package | Last gate before done — no mock/stub remains; validated against real systems (ADR-098). |
-| `security-engineer` | Security Engineer | — | package | OWASP-shaped failure modes, secret leakage, trust boundaries. |
-| `frontend-engineer` | Frontend Engineer | — | package | Component lifecycle, reactive state, hydration boundaries. |
-| `product-owner` | Product Owner | — | package | Outcomes named, AC unfalsifiable, scope on record. |
-| `tech-writer` | Tech Writer | — | package | Release narratives, READMEs, AGENTS.md thin. |
-| `revops-maintainer` | RevOps Maintainer | — | package | Contributor lifecycle, package adoption funnel, release readiness. |
-| `discovery-lead` | Discovery Lead | — | package | Switch events, falsifiable hypotheses, theme ranking. |
-| `hollywood-director` | Hollywood Director (`ai-video`) | — | package | Live-action lens, lighting, blocking, negative constraints. |
-| `ai-video-technical-director` | AI Video Technical Director (`ai-video`) | — | package | Provider tuning — Veo / Kling / Sora / Higgsfield / OpenAI grammar. |
-| `brand-strategist` | Brand Strategist (`design`) | — | package | Positioning, archetype, voice, messaging; challenges weak briefs. |
-| `design-director` | Design Director (`design`) | — | package | Art-direction and brand-aligned visual judgment; composition, hierarchy, vector-vs-raster. |
-| `cmo` | CMO | 3 | package | Positioning anchored, messaging stacked, launches sequenced. |
-| `growth-pm` | Growth PM | 3 | package | Leaky-bucket vs growth-loop classified, activation correlated. |
-| `customer-success-lead` | Customer Success Lead | 3 | package | TTFV falsifiable, churn cause split, expansion pulled, NRR. |
-| `revops` | RevOps | 3 | package | Stage exit criteria, MEDDIC slots, forecast falsifiable. |
-| `engineering-manager` | Engineering Manager | 4 | package | 1:1 cadence, hiring loop, throughput-vs-morale tradeoff. |
-| `people-strategist` | People Strategist | 4 | package | Team shape, comp bands, ramp definitions, feedback craft. |
-| `finance-partner` | Finance Partner | 4 | package | Unit economics, runway, scenarios, the next 18 months. |
-| `strategist` | Strategist | 4 | package | Build-vs-buy, market entry, moat, vision, contracts, privacy. |
-
-## Per-domain count (persona-governance cap: max 2 specialists per domain)
-
-| Domain | Active specialists | Within cap |
-|---|---|---|
-| `ai-video` | `hollywood-director`, `ai-video-technical-director` | ✅ 2/2 |
-| `design` | `brand-strategist`, `design-director` | ✅ 2/2 |
-| GTM (Wing 3) | `cmo`, `growth-pm`, `customer-success-lead`, `revops` | ✅ (wing, not domain) |
-| Ops / Money (Wing 4) | `engineering-manager`, `people-strategist`, `finance-partner`, `strategist` | ✅ (wing, not domain) |
-| Backend | `backend-architect`, `eloquent-tamer` | ✅ 2/2 |
-| `performance` | `performance-engineer` | ✅ 1/2 |
-| `data` | `data-integrity` | ✅ 1/2 |
-
-Cross-cutting quality lenses (`qa`, `tech-writer`, `revops-maintainer`,
-`discovery-lead`, `production-validator`) are uncapped — they are not bound to a
-single content domain.
-
-Per-domain count is enforced by `src/scripts/lint_persona_governance.ts`
-(wired into `task ci`), which scans `src/agent-src/personas/` (the source of
-truth) and `src/skills/` + `src/rules/` for `personas:` citations.
-
-## Advisors (`personas/advisors/`, 5)
-
-Adversarial / second-opinion voices invoked by `ai-council` debates,
-not cited by skills directly:
-
-| ID | Lens |
-|---|---|
-| `contrarian` | Opposing the consensus position |
-| `executor` | Shipping bias — what's the minimum to land it? |
-| `expansionist` | Scope-widening — what else does this unlock? |
-| `first-principles` | Refuses analogy — derive from physics of the problem |
-| `outsider` | Naive reader — does this hold without prior context? |
-
-## Tier rules
-
-- **Core (≤ 6)** — always-loaded cast. 5 sections (Focus · Mindset · Unique Questions · Output Expectations · Anti-Patterns), ≤ 120 lines.
-- **Specialist** — opt-in lenses. 7 sections (Core-5 + Critical Rules + Workflows). Line cap is wing-scoped: ≤ 100 (no wing / Wings 1–2), ≤ 140 (Wings 3–4) per [`persona-schema § 4`](contracts/persona-schema.md).
-
-Schema contract: [`docs/contracts/persona-schema.md`](contracts/persona-schema.md). Governance rule: [`persona-governance`](../dist/agent-src/rules/persona-governance.md).
-
-## When to invoke a persona
-
-- **A skill cites it.** Skills declare `personas:` in frontmatter; the
-  agent adopts the listed personas while running the skill. No user
-  action needed. Example: `judge-security-auditor` cites
-  `security-engineer` — invoking the skill applies the lens.
-- **You want a second pass on existing work.** Pick the specialist
-  whose checklist matches the diff: `eloquent-tamer` for ORM heavy
-  changes, `frontend-engineer` for component changes,
-  `backend-architect` for service-seam changes.
-- **Adversarial review.** `critical-challenger` for the devil's-advocate
-  voice; `security-engineer` for abuse-case framing.
-
-Invocation example (when supported by the host tool):
-
-```text
---personas=backend-architect,security-engineer
-```
-
-The host agent reads the persona's `Critical Rules` and `Workflows`
-sections (specialist tier) or `Mindset` and `Unique Questions` (core
-tier) and merges them into its review pass.
-
-## `personas:` vs `/mode` — two different axes
-
-| Axis | What it controls | Example |
-|---|---|---|
-| `/mode` (role-mode) | Which **contract** the agent operates under for a turn — what outputs are mandatory, what work is refused | `/mode reviewer` refuses to write production code |
-| `personas:` (lens) | Which **review checklists** the agent overlays on a skill's procedure | `personas: [security-engineer]` adds OWASP-shaped checks to whatever skill is running |
-
-The two compose. `/mode reviewer` + `personas: [security-engineer,
-backend-architect]` = a review-only turn that applies both the
-security-abuse-case lens and the service-seam lens on top of the
-review contract.
-
-Personas are **not** a replacement for `/mode`. A persona cannot
-refuse work or change output contracts; that is `/mode`'s job.
-
-## Skill citations
-
-Specialist personas need at least one skill citation to earn their
-slot. Current cite map:
-
-| Persona | Cited by |
-|---|---|
-| `qa` | `judge-test-coverage` |
-| `critical-challenger` | `adversarial-review`, `requesting-code-review`, `receiving-code-review` |
-| `developer` · `senior-engineer` · `product-owner` · `ai-agent` | `refine-ticket`, `refine-prompt`, `estimate-ticket` |
-| `backend-architect` | `authz-review`, `api-design`, `blast-radius-analyzer` |
-| `eloquent-tamer` | `eloquent`, `database` |
-| `security-engineer` | `judge-security-auditor`, `threat-modeling`, `authz-review` |
-| `frontend-engineer` | `existing-ui-audit`, `fe-design` |
-| `brand-strategist` | `brand`, `brand-strategy` |
-| `design-director` | `brand-identity`, `logo-generation`, `brand-asset-generation` |
-
-Add a citation by appending `personas: [<id>]` to the skill's
-frontmatter — never overwrite an existing list, append.
-
-## Override pattern
-
-Consumers can override any package-shipped persona via the standard
-override mechanism:
-
-- **Frontmatter override** — change `tier`, `mode`, `description`
-  while keeping the package body. Frontmatter wins.
-- **Body override** — full replace. Drop a same-named file under
-  `agents/personas/<id>.md`; the package version is shadowed.
-
-See [`override-management`](../dist/agent-src/skills/override-management/SKILL.md)
-for the mechanics.
-
-## Anti-patterns
-
-- **Personas as agents.** Personas are lenses, not sub-agents. They do
-  not spawn processes or run tools independently.
-- **Personas as modes.** They do not change the host's contract or
-  refuse work. Use `/mode` for that.
-- **Specialist sprawl.** Cap is "≤ 4 specialists per plate" by the
-  council A-OQ2 verdict. New specialists need ≥ 1 skill citation
-  before merging.
-- **Core-tier mutation.** Core personas are versioned (`version` in
-  frontmatter); breaking changes need a new ID, not an in-place
-  rewrite.

package/docs/positioning-evidence.md

@@ -1,86 +0,0 @@
-# Positioning evidence & decision (B0)
-
-> Market-readiness roadmap B0. The evidenced differentiators the positioning
-> rests on, the H1 options that were weighed, and the decision. Every
-> differentiator below is **shipped and machine-checked**, not aspirational —
-> each names the artifact that backs it.
-
-## The problem
-
-External adoption is ≈0. The prior README H1 — *"Governed skills, rules & work
-journeys for AI coding agents"* — was abstract and interchangeable: no single
-uncopyable hook. This document fixes the one sentence that carries the
-differentiator.
-
-## Evidenced differentiators
-
-1. **Machine-checked honest marketing (the Claims-Ledger).** Every public claim
-   that carries a `<!-- claim:ID -->` marker must bind to a resolvable evidence
-   pointer in [`CLAIMS.md`](CLAIMS.md), or CI fails. Unbacked claims are logged
-   as visible debt rather than hidden. Evidence:
-   [`src/scripts/check_claims.ts`](../src/scripts/check_claims.ts),
-   [`CLAIMS.md`](CLAIMS.md).
-2. **A self-verifying proof page.** [`proof.md`](proof.md) is generated from the
-   ledger (claim → evidence table) plus the honest-null benchmark runs
-   (*including the nulls*) plus a "verify it yourself" command block a skeptic
-   runs on a fresh checkout. Drift-gated in CI. Evidence:
-   [`src/scripts/build_proof.ts`](../src/scripts/build_proof.ts),
-   [`proof.md`](proof.md).
-3. **Published honest-nulls.** Measured null results ship rather than hide (e.g.
-   a recursive-verification wrapper measured for null lift; an orchestration
-   auto-flip held at a documented honest-null). Evidence:
-   [`benchmark.md`](benchmark.md).
-4. **No-runtime, host-agnostic governance.** Rules compile into each tool's
-   native format at projection time; the same governance rides on 7+ hosts with
-   no daemon. Surgical uninstall (JSON-pointer + SHA-256, never a neighbour
-   tool's keys). Evidence:
-   [`contracts/no-runtime-boundary.md`](contracts/no-runtime-boundary.md),
-   [`contracts/install-layout.md`](contracts/install-layout.md).
-5. **A 30-second wedge.** `production-validator` — a single self-contained Claude
-   Code subagent (one `curl` into `.claude/agents/`) that runs the last gate
-   before "done": no mock/stub on the shipped path, validated against real
-   systems. Evidence: [`wedge/production-validator/`](wedge/production-validator/).
-
-## Competitive frame
-
-The dominant competitor wins on reach behind a loud, **unbacked** headline
-performance number (present in several files, no reproducible methodology). The
-only defensible, uncopyable inverse is **reproducible honesty**: they cannot
-adopt a Claims-Ledger + proof page without exposing their own unbacked numbers.
-"Honesty" alone can read as merely defensive, so the wedge and the governed
-depth carry the offensive substance.
-
-## H1 options weighed
-
-- **Option 1 — verifiability as frame.** *"Agent Config — every claim
-  machine-checked, including 'zero runtime'."* Subline: 261 skills, 93 rules,
-  7 hosts; run the proof page yourself; no daemon, no lock-in, no unverifiable
-  promises.
-- **Option 2 — architecture first.** *"Governed AI agents with zero runtime —
-  and a proof page you verify in 30 seconds."*
-- **Option 3 — host-agnostic.** *"Write your agent governance once — run it in
-  7 hosts, with a proof page that can't lie."*
-
-## Decision
-
-**Adopted: Option 1** (verifiability as frame).
-
-Council (claude-sonnet-4-5 + gpt-4o, 2026-07-04) split only on headline vs.
-subline placement, and converged on the substance: the stateless / host-agnostic
-/ no-runtime architecture is the technical moat, and the Claims-Ledger + proof
-page is the maximally-uncopyable trust bootstrap. The decisive argument: in a
-low-trust market, "zero runtime" on its own is just another unverifiable
-promise — making **verifiability the frame** and the architectural claims *what
-you verify* is what makes "zero runtime" credible. The maintainer made the final
-call per the B0 "human picks" gate.
-
-**Biggest risk (tracked):** "machine-checked" can read as compliance-boring
-rather than a competitive weapon — mitigated by framing it offensively
-("built to be checked, in a market that runs on unbacked headlines"), not
-defensively.
-
-## Applied to
-
-- README H1 + hero line (this repo).
-- `distribution/launch-story-draft.md` headline placeholder (resolved).
-- `distribution/awesome-list-submission.md` (already Option-1-aligned).

package/docs/profiles.md

@@ -1,108 +0,0 @@
-# Six entry paths — by `profile.id`
-
-Each block below is the first-screen for one shipped profile. The
-`profile.id` written by `agent-config setup` selects the anchor; the block names
-the audience, the first three things the agent does for that role,
-and the exact commands and skills wired into the profile YAML at
-[`src/agent-src/profiles/<id>.yml`](../src/agent-src/profiles/).
-
-The summary table at the top of [`README.md`](../README.md) is the
-one-page index; the prose below is the deep version.
-
-→ Each profile's **experience page** (packs · flows · what is *not* loaded ·
-examples) lives under [`docs/experiences/`](experiences/); the multi-command
-**work journeys** those profiles run are in [`docs/flows.md`](flows.md).
-
-**Switching experience.** `agent-config setup` writes `profile.id` on first run
-(it is the wizard's first question). To switch later, run
-`agent-config use --profile=<id>` (one of `developer` · `content_creator` ·
-`founder` · `agency` · `finance` · `ops`). In 6.0.0-A this records the choice
-only; pack-scoped surfacing (projection-time filtering,
-[`ADR-040`](decisions/ADR-040-execution-model-projection-time-filtering.md))
-activates in 6.0.0-B behind a staged, opt-in rollout.
-
-<a id="profile-developer"></a>
-## 👩‍💻 `developer` — IC engineer
-
-Implement a ticket end-to-end, fix CI red, run a self-review before
-the PR. `/implement-ticket` refines the ticket, plans, edits, tests,
-and verifies; `/work` is the free-form sibling; `/review-changes`
-dispatches five judges (bug, security, tests, quality, architecture)
-on the local diff. Stack-aware skills cover Laravel · Symfony ·
-Next.js · React · Node. **Preset default: `balanced`.**
-[Profile YAML](../src/agent-src/profiles/developer.yml) ·
-[Role guide](getting-started-by-role.md#developer-the-original-audience).
-
-<a id="profile-content_creator"></a>
-## ✍️ `content_creator` — writers, ghostwriters, marketers
-
-Draft in someone else's voice, plan a quarter of content, ship a
-launch announcement, render a cinematic AI video. `/ghostwriter`
-fetches and writes against a public-figure voice profile; `/post-as`
-is the same primitive for your own voice (`.agent-user.md`);
-`voice-and-tone-design` and `messaging-architecture` lock the brand
-frame before any copy ships; `/video:from-script` and `/video:storyboard`
-drive the AI video pipeline (script → character-locked image →
-motion+audio prompt → provider render → stitched clip), with
-`character-consistency` locking identity tokens across scenes and
-`AIV_DRYRUN=true` as the cost-safety default.
-**Preset default: `balanced`.**
-[Profile YAML](../src/agent-src/profiles/content_creator.yml) ·
-[Role guide](getting-started-by-role.md#creator-writer-marketer-indie-content-shop).
-
-<a id="profile-founder"></a>
-## 🚀 `founder` — solo / early-stage founder
-
-Sharpen a fuzzy idea, rank what to build, write the why-now slide.
-`/challenge-me` runs a grill-style interview that turns a vague plan
-into a copyable pitch; `/council` polls external AIs for a neutral
-second opinion; `rice-prioritization` ranks the backlog;
-`vision-articulation` and `fundraising-narrative` shape the
-internal-vs-external story. **Preset default: `fast`.**
-[Profile YAML](../src/agent-src/profiles/founder.yml) ·
-[Role guide](getting-started-by-role.md#founder-early-stage-operator-wearing-every-hat).
-
-<a id="profile-agency"></a>
-## 🏛 `agency` — multi-client delivery shop
-
-Refine a fuzzy client ask into an estimated, AC-tight ticket; turn a
-phase into a roadmap; ship per client without losing decision
-provenance. `/refine-ticket` rewrites the ticket and surfaces top-5
-risks; `estimate-ticket` sizes and splits; `decision-record` anchors
-the trade-off in an ADR before code starts. **Preset default:
-`strict`.**
-[Profile YAML](../src/agent-src/profiles/agency.yml) ·
-[Role guide](getting-started-by-role.md#consultant-advisory-freelance-fractional).
-
-<a id="profile-finance"></a>
-## 💼 `finance` — CFO / fractional finance / FP&A
-
-Build a DCF, stress-test the plan, frame the runway call. `dcf-modeling`
-walks the WACC / terminal-value / 5-year-hold reasoning; `forecasting`
-reconciles top-down vs bottom-up; `scenario-modeling` produces the
-base / upside / downside cuts; `runway-cognition` frames the
-fundraise-vs-cut-vs-grow decision. **Preset default: `strict`.**
-[Profile YAML](../src/agent-src/profiles/finance.yml) ·
-[Role guide](getting-started-by-role.md#finance--ops-cfo-controller-ops-lead-founder-finance).
-
-<a id="profile-ops"></a>
-## 🛡 `ops` — RevOps, support, SRE-adjacent
-
-Threat-model a change before it ships, command the incident when it
-breaks, build the dashboard that catches it next time.
-`/threat-model` enumerates abuse cases and trust boundaries before
-the first line of code; `incident-commander` frames severity and
-post-mortem; `dashboard-design` chooses the right RED / USE / Golden
-Signal panel. **Preset default: `strict`.**
-[Profile YAML](../src/agent-src/profiles/ops.yml) ·
-[Role guide](getting-started-by-role.md#finance--ops-cfo-controller-ops-lead-founder-finance).
-
-> **Universal AI Agent OS, not "for developers only".** The same
-> orchestration core also drives non-software trades. Worked-example
-> user types ship for [galabau](../src/agent-src/user-types/galabau-field-crew.md),
-> [metalworking](../src/agent-src/user-types/metalworking-shop.md),
-> [truck driving](../src/agent-src/user-types/truck-driver.md) —
-> with a [scaffold](../src/agent-src/user-types/_template/) for
-> contributing your own.
-
-→ [Public catalog](catalog.md) (all rules, skills, commands, guidelines) · [Skills only](skills-catalog.md) · [llms.txt](../llms.txt)

package/docs/proof.md

@@ -1,176 +0,0 @@
-<!-- GENERATED by src/scripts/build_proof.ts — do NOT hand-edit.
-     Drift-checked in CI (`task build-proof-check`). Regenerate with
-     `./scripts-run src/scripts/build_proof` after editing docs/CLAIMS.md. -->
-
-# Proof — verify our claims yourself
-
-We sell falsifiability, so the selling is machine-checked. Every
-public claim binds to resolvable evidence; a skeptic can reproduce
-every check below on a fresh checkout. This page is itself generated
-from those sources and fails CI if it drifts.
-
-## See it run (< 60s, real output)
-
-![The trust surface running green — every "verify it yourself" command below, recorded from a real run](media/proof-demo.gif)
-
-The recording is of the exact commands in § 5, nothing staged. Its
-commands are re-executed in CI (`.github/workflows/proof-demo.yml`), so a
-demo that showed something broken would turn CI red — the recording
-cannot drift from current behavior.
-
-## 1. Every public claim binds to evidence
-
-Rendered from [`docs/CLAIMS.md`](CLAIMS.md). A `<!-- claim:ID -->` marker in
-README/docs must resolve to a `backed` entry here with a resolving
-evidence pointer, or `task check-claims` fails the build.
-
-| Claim | Kind | Evidence | Resolves |
-|---|---|---|---|
-| The release process is documented as an inheritable runbook + succession doc, and the project's bus-factor (trailing-90-day distinct human reviewers) is tracked and reported truthfully — currently 1, not implied to be more. | qual | `docs/succession.md#trailing 90 days` | ✅ |
-| 190 commands. | quant | `src/scripts/check_artefact_count_messaging.ts#Artefact-count messaging gate` | ✅ |
-| A MEASURED-BUT-NOT-SHIPPED experiment — the thin rule projection reduced eager rule load 78,513 → 13,881 GPT-tokens (whole always-loaded projection 98,529 → 33,897, ~65.6%), but FAILED the quality gate (thin win-rate 36.2% vs required 48%) and does not ship; it un-defers only behind `discipline_profile: essential`. Shipped behavior does NOT include this reduction. Method: `agent-config benchmark` over the pinned token baseline; the baseline is the honest "what the user pays if everything loads eagerly", NOT a synthetic full-corpus strawman (council Q4); quality gate per the Phase-0 paired judge run. | quant | `internal/bench/reports/token-baseline.json#eager_rule_load` | ✅ |
-| The first cross-vendor parity pass (5 orchestration-corpus tasks × 2 vendors × 3 repeats, identical prompts via the council transport, $0.16) measured real per-host finding-count differences — claude-sonnet-4-5 surfaced ~2× the findings of gpt-4o on the multi-file analysis task (median 11 vs 5) while both vendors were identical on the planted hollow-implementation task (2 vs 2) and perfectly silent on the clean-code negative control (0 vs 0, no spurious findings). The per-task `finding_floor` values are calibrated from the cross-host lower envelope and the gate is armed. | quant | `internal/bench/reports/parity-count.json#min over hosts of median` | ✅ |
-| On a weak host (claude-haiku-4-5) the package produces a significant, placebo-controlled discipline lift on scope/downstream traps; on a strong host the same measurement is a published null — the package transplants discipline a weak model lacks, not model intelligence. | quant | `docs/benchmark.md#weak-host-specific` | ✅ |
-| The non-coding domain skills (finance/founder/ops/content) are forged on TS/PHP and labeled unvalidated until they pass a sourced domain-truth fixture; no public prose implies proven domain correctness, and the validated count is CI-ratcheted. | qual | `src/scripts/domain_soundness_status.ts#checkRatchet` | ✅ |
-| The validated non-coding domain-skill count is pinned and CI-ratcheted at a maintainer-set floor (9 of 20 default-surface skills carry a sourced `evals/domain-truth.json` fixture at pin time, 2026-07-11 — 5 deterministic, keys from cited formulas; 4 rubric, criteria matching a named external practice); the floor only rises via a maintainer `--write-floor` after a new sourced fixture lands. | quant | `internal/evals/domain-soundness-floor.json#validated` | ✅ |
-| On the READ-ONLY FAN-OUT slice family, tier-downshifted subagent dispatch (lite/haiku vs session-tier-proxy sonnet) nets a ≥30% USD-weighted token-cost reduction at held quality — measured 2026-07-08 (n=10 paired live dispatches, 20 telemetry lines): 10/10 exact-match on BOTH arms, 29.4% fewer raw tokens, 76.5% USD-weighted cost reduction at the 3x haiku↔sonnet price ratio. FAMILY-SCOPED — the mechanical-edit family is unmeasured and its downshift (incl. the deferred tier downgrades of existing units) stays gated. Negative control held: an open-ended synthesis/unknown slice never resolves below the session tier (inferSliceTier → medium/inherit, never lite). | quant | `internal/bench/routing-downshift/results-2026-07-08.md#FAMILY-SCOPED PROVE` | ✅ |
-| The lift-carrying essential cut (kernel + downstream-changes) keeps a significant weak-host discipline lift at a fraction of the full load's tokens, and the lift is FAMILY- and HOST-SCOPED — measured on three hosts: claude-haiku-4-5 (weak) shows the family-scoped lift (trapE 0.533→1.000, 7/7 discordant, corpus cost 1.71x); claude-sonnet-4-6 (strong) is a ceiling null; gpt-5-mini (non-Claude weak, codex prompt-prepend surface) FAILED replication with headroom (corpus Δ=+0.024 p=0.70, capability trend n.s. — no harm claimed, injection-surface confound documented). Therefore discipline_profile: auto enables the lift only where measured (vendor-granular unknown_defaults). Non-claims — the balanced router profile was removed after a NULL measurement (p=0.81, n=24); no full-tier recommendation exists; no cross-vendor lift is claimed. | quant | `docs/benchmark.md#REPLICATION FAILED` | ✅ |
-| Behavioural-eval coverage is measured per tier and CI-ratcheted so it can only rise; the current coverage and its gap are published, never implied as "264 evaluated skills". | qual | `src/scripts/skill_eval_coverage.ts#checkRatchet` | ✅ |
-| On the fixture corpus (n = 20 before/after pairs, 16 length-controlled within ±25%), the humanizer pass removes every mechanically detected AI-writing tell (mean hard hits 0.9 → 0, cluster score 53.97 → 0 per 500 words, dash density 9.22 → 0), and a blind judge (claude-sonnet-4-5, deterministic per-pair A/B seed) preferred the humanized text in 16/16 length-controlled pairs. Scope note — the "before" fixtures were deliberately tell-seeded, so this measures seeded-tell removal on a self-constructed corpus, NOT real-draft improvement; real-world lift is unmeasured until step 4b has processed real ghostwriter drafts (see the road-to-humanizer-hardening live-usage blocker). | quant | `internal/bench/reports/humanizer-v1.md#prefers the humanized text` | ✅ |
-| A hand-rolled, dependency-free BM25 + trigram lexical index resolves the "recalls but does not rank" gap: on the retrieval-precision corpus (9 keyword-overlapping-confuser tasks) it drives the mean top tie-set from 3.333 (the `_score` bucket scorer) to 1.0 — every needed decision uniquely top-ranked — with precision@1 and precision@5 unchanged at 1.0. Method: deterministic, model-free re-ranking of the SAME retrieved entry set; both scorers measured over the identical store via `measure_lexical_ranking.ts`. | quant | `internal/bench/reports/lexical-ranking.json` | ✅ |
-| The whole layer is compiled into host agents with zero runtime daemon. | qual | `docs/contracts/no-runtime-boundary.md#file-first, no-runtime suite` | ✅ |
-| On a 12-fixture option-decision corpus (3 arms × 2 providers, blind rubric judge claude-opus-4-8, pre-registered hypotheses), famous-figure identity framing added nothing beyond the underlying method text (method 5.04 vs figure 4.88, Δ=0.17, sign-test p=0.607), and provider diversity moved judged quality ~15× more than persona identity (provider Δ=2.58 vs identity Δ=0.17); the whole persona layer lifted only +0.08 over bare prompts. Honest null — persona panel-mode stays CUT, evidence-closed. | quant | `internal/bench/reports/persona-placebo.json#honest-null` | ✅ |
-| The only agent layer that publishes the runs where it changed nothing. Deliberately falsifiable — if a reader finds a comparable agent layer publishing its own honest-null benchmark runs, this line updates; that is the point. | comparative | `docs/benchmark.md#honest` | ✅ |
-| On the live end-to-end retrieval benchmark (9 tasks × 3 arms × 3 seeds on claude-haiku, fixture store with keyword-overlapping confusers), the memory retrieval substrate scored precision@5 = 100% (9/9) with 100% poisoned-entry rejection, and the retrieval-on arm passed 27/27 model-scored tasks vs 2/27 with retrieval off and 4/27 with a placebo injection. Known limit stays published: mean tie-set 4.1 means top-k ties break by store order, not relevance (the ADR-116/FTS5 signal). | quant | `internal/bench/reports/second-brain-retrieval.json#retrieval-on` | ✅ |
-| 105 governed rules. | quant | `src/scripts/check_artefact_count_messaging.ts#Artefact-count messaging gate` | ✅ |
-| On a deterministic multi-session recall corpus, the memory substrate produces a measured, placebo-controlled recall lift — memory-on 27/27 vs no-memory 10/27 and vs equal-byte placebo 9/27 (claude-haiku-4-5, n=9 tasks x 3 seeds, sign test p=0.031 for BOTH pairings). Scoped honestly: this is the context-value upper bound (perfect retrieval on a one-fact-per-task corpus), not retrieval precision under a large store. | quant | `internal/bench/reports/second-brain-delta.json` | ✅ |
-| Removing the perfect-retrieval assumption, the substrate's REAL keyword retrieval recalls the needed decision into the top-5 under keyword-overlapping confusers (precision@5 9/9) and the model disambiguates it from the co-injected confusers — retrieval-on 27/27 vs no-memory 5/27 and vs equal-count placebo 5/27 (claude-haiku-4-5, 9 tasks x 3 seeds, sign test p=0.008 both). Named limit: retrieval RECALLS but does not RANK (mean tie-set 3.3, ties broken by store order, not relevance) — the discrimination gap that motivates the SQLite-FTS5 activation path (ADR-116) at larger scale. | quant | `internal/bench/reports/second-brain-retrieval.json` | ✅ |
-| Every artifact the package ships — source AND the condensed projection that reaches consumers — is machine-scanned in CI for hidden-Unicode, mixed-script-confusable, and instruction-smuggling payloads (the rules-file-backdoor class); a finding blocks the release before `npm publish`, not just the merge. | qual | `.github/workflows/publish-npm.yml#lint_agent_security` | ✅ |
-| 278 skills. | quant | `src/scripts/check_artefact_count_messaging.ts#Artefact-count messaging gate` | ✅ |
-| Removes only its own keys from a shared host config (matched by JSON-pointer + SHA-256), never a neighbour tool's entries. | qual | `docs/contracts/install-layout.md#JSON-pointer` | ✅ |
-| On the pre-registered 12-fixture defect-finding corpus (three arms, deterministic file-level recall, codex reviewer gpt-5.5), the cross-model team-review arm produced NO recall lift over single-model self-review — all three arms recalled every planted defect (Δ = 0, H1 not met). Honest null, ceiling-limited (recall 1.00 everywhere: the seeded defects are too obvious to discriminate the arms on recall); the only non-null signal is a single self-review false positive on the controversial-but-correct control vs 0 for team/council. No cross-model quality/lift claim binds; team mode stays workflow-value-only. Re-open: a judge-survivable-subtlety corpus or a new model generation. | quant | `internal/bench/reports/defect-finding.json#honest_null` | ✅ |
-| On the A3 orchestration eval (deterministic verify, measured token deltas), the production-validator subagent returned the correct verdict on both fixtures — NOT READY with the exact `file:line` citation on a planted hollow implementation, READY with zero spurious findings on the clean control — while consuming ~45k fewer tokens than the inline-host baseline on each task. Scope: two planted fixtures on a Claude Code host, not a broad hit-rate. | quant | `internal/bench/orchestration/pv-a3-results.md#token_delta_provenance: measured` | ✅ |
-
-**24 backed claim(s)** — all evidence pointers resolve in CI.
-
-Artefact counts in public prose (skills, commands, governed rules,
-guidelines, personas) are **generated from source and CI-drift-checked**:
-`update_counts.ts` writes the numbers, `check_artefact_count_messaging.ts`
-fails the build on any count-shaped prose mention that drifts from the
-source count — or on two different numbers for the same artefact kind.
-
-We also publish our **debt**: 5 claim(s) are logged as
-`unbacked` inventory in the ledger — not yet bound, and therefore not
-allowed to carry a marker in public prose. Hiding them would be the
-opposite of the point.
-
-## 2. We publish honest nulls
-
-Benchmark results — **including the runs where the package changed nothing** —
-live in [`docs/benchmark.md`](benchmark.md). We do not delete a measured null to
-make a number look better; the null is the evidence of honesty.
-
-The freshest example: the **persona-placebo benchmark** <!-- claim:persona-identity-placebo-null -->
-(2026-07-12, 3 arms × 2 providers, blind rubric judge, pre-registered
-hypotheses) measured whether famous-figure persona framing improves decision
-answers over the bare method text it wraps. It does not (Δ=0.17, p=0.607) —
-while provider diversity moved judged quality ~15× more than persona identity.
-The measured null closed a planned feature (persona panel-mode) instead of
-shipping theater; the full per-cell data is committed at
-`internal/bench/reports/persona-placebo.json`.
-
-**Behavioural-eval coverage — the honest baseline.** Skill *quality* is only
-as good as its measurement. Today **42 of 278** skills carry a behavioural
-`evals.json`; the highest-traffic / highest-cost tiers (default-surface +
-`rich` + routers) are **fully covered (35 of 35)**, the long tail
-(7 of 243) is not. We publish that gap rather than imply
-"278 evaluated skills": coverage is measured per tier
-(`./scripts-run src/scripts/skill_eval_coverage`), **CI-ratcheted so it can
-only rise**, and the priority tiers carry a hard **tier floor**: every
-rich / default-surface / router skill MUST have a behavioural eval or an
-explicit exemption-with-reason in `internal/evals/tier-floor-exemptions.json`
-— no silent exemptions (the bright line deliberately replaces a
-weighted-coverage score). Authoring long-tail evals is gated on per-case
-human ratification (a generated assertion that checks the wrong property is
-worse than none), so the number grows deliberately, not overnight.
-
-**Non-coding domain soundness — scoped, not proven.** The `finance` /
-`founder` / `ops` / `content` profiles sell concrete domain value (DCF,
-runway, RICE, incident command, messaging), but the skills are forged on
-TS/PHP — "promising, not proven" off those stacks. A disclaimer floor
-bounds *liability*, not *correctness*: a skill can be format-correct,
-disclaimered, and still embed a wrong domain assumption. Today **9 of 20**
-default-surface domain skills carry a sourced `domain-truth` fixture
-(`./scripts-run src/scripts/domain_soundness_status`); the rest are labeled
-`unvalidated` and the validated count is CI-ratcheted. The fixtures landed
-so far are the **deterministic** targets (runway, unit-economics, DCF,
-forecasting, scenario band/sensitivity), whose answer keys are computed
-from cited standard formulas — never the skill's own output; the
-**rubric** targets (incident command, messaging, fundraising, editorial)
-need domain-competent grounding and remain unvalidated, so validation
-lands deliberately — the gap is published, never implied away.
-
-**Second-brain substrate — measured recall lift, honestly bounded.** On a
-deterministic multi-session recall corpus, the memory substrate beats a
-no-memory baseline AND an equal-byte placebo: memory-on **27/27** vs
-no-memory **10/27** vs placebo **9/27** (claude-haiku-4-5, 9 tasks × 3
-seeds, sign test **p = 0.031** for both pairings) — a real,
-placebo-controlled lift (`internal/bench/reports/second-brain-delta.json`).
-Scoped, not oversold: this is the context-value upper bound (perfect
-retrieval on a one-fact-per-task corpus), NOT retrieval precision under a
-large store; the lift concentrates exactly where memory is the only source
-and ties where the prompt self-contains the fact. Boundary vs a human PKM,
-and why the Obsidian export stays declined, are in
-[`docs/second-brain-scope.md`](second-brain-scope.md).
-
-A follow-up removed the perfect-retrieval assumption: against a store of
-keyword-overlapping confusers the REAL retrieval recalls the needed
-decision into the top-5 (9/9) and the model disambiguates it — retrieval-on
-**27/27** vs no-memory **5/27** and vs placebo **5/27** (p=0.008 both). The
-honest limit: the keyword scorer recalls but does not rank (mean tie-set
-3.3), which is the SQLite-FTS5 activation signal (ADR-116) at scale.
-
-## 3. Known limits (published, witness-tested)
-
-Each limitation below is stated by the skill itself and carries a witness
-test that reproduces it. If the limitation is ever fixed, its witness goes
-red — so a "Known limit" can never quietly become false.
-
-| skill | known limit | witness |
-|---|---|---|
-| `check-refs` | Only validates references to known-root paths (docs/, skills/, rules/, commands/, contexts/, personas/, …). A relative-path link such as `./sibling.md` or `../foo.md` is not matched, so a broken relative link is never reported. | [`tests/scripts/witness/check_refs_relative_gap.test.ts`](../tests/scripts/witness/check_refs_relative_gap.test.ts) |
-
-## 4. What is checkable — us vs. the category
-
-This is not a takedown — the point is the last column. For each claim,
-our evidence is a pointer you can resolve on a fresh checkout; the wider
-category is described only by what is publicly observable, never a named
-competitor and never a counter-claim to anyone's headline number. A claim
-is "checkable" only when its `our evidence` pointer resolves — CI enforces
-that (`task check-comparison`), so this column can never lie.
-
-| Claim | Our evidence | The category | Checkable? |
-|---|---|---|---|
-| No runtime — no background daemon, no state database, no auto-write memory. | [`docs/contracts/no-runtime-boundary.md#file-first`](../docs/contracts/no-runtime-boundary.md) | Swarm-runtime tools in this category ship a background process and/or a state store by design; that is an architectural fact of the runtime approach, not a defect. | ✅ |
-| Capability/benchmark results are published — including the runs where the package changed nothing. | [`docs/benchmark.md`](../docs/benchmark.md) | A category headline figure (e.g. an '84.8%' score) appears across marketing surfaces with no reproducible methodology published — so it cannot be verified either way. | ✅ |
-| Every public claim binds to machine-checked evidence. | [`docs/CLAIMS.md`](../docs/CLAIMS.md) | Marketing claims in this category are not bound to a machine-checked ledger a reader can reproduce. | ✅ |
-| Skills publish their own known limits, each with a witness test. | [`tests/scripts/witness/check_refs_relative_gap.test.ts`](../tests/scripts/witness/check_refs_relative_gap.test.ts) | Published known-limitation surfaces backed by reproducing tests are not a standard artifact in this category. | ✅ |
-| A 30-second single-file wedge exists — one read-only subagent, one command, verdict with file:line — and its promise is a scoped, published eval, not a feature list. | [`internal/bench/orchestration/pv-a3-results.md#token_delta_provenance: measured`](../internal/bench/orchestration/pv-a3-results.md) | Category entry points typically install the full platform (profiles, packs, runtime) before the first felt win; single-artifact entries with a published eval behind the promise are not a standard artifact. | ✅ |
-| The persona-identity question was tested and published as an honest null (identity swap ~zero effect, provider choice ~15x larger) instead of being shipped as theater. | [`internal/bench/reports/persona-placebo.json#honest-null`](../internal/bench/reports/persona-placebo.json) | Named-persona prompts are shipped as a feature across the category without a published A/B separating identity from method. | ✅ |
-
-## 5. Verify it yourself
-
-On a fresh checkout, reproduce the claims above:
-
-```bash
-task check-claims   # every markered public claim binds to resolvable evidence
-task check-refs     # no broken internal references
-task check-skill-gaps    # every logged known-limit cites a real witness test
-task check-comparison    # every comparison-table "our evidence" pointer resolves
-./scripts-run src/scripts/skill_eval_coverage         # behavioural-eval coverage, per tier
-./scripts-run src/scripts/skill_eval_coverage --check # the ratchet: coverage may not drop
-task build-proof-check   # this page is in sync with its sources
-```
-
-If a claim ever loses its binding, or this page drifts from the ledger,
-CI goes red. Reproducibility is the proof.