@nurix/etna 0.19.40 → 0.19.42

package/data/harnesses/discover/codex/agents/distribution-artifacts.toml

@@ -55,4 +55,28 @@ Then write the per-type raw files that apply — `raw/{extension-manifest.md, el
 Return a **one-paragraph headline** of what the shipped client reveals — its declared hosts, permission surface, embedded vendor SDKs, and how much richer (or not) it is than the web app. The API paths and feature flags you appended feed the **seam rule (§6)**: your `dimensions/_shared/api-path-catalog.md` rows carry `source: distribution` so evaluation can run its adaptive diff against `deployed-client-bundle` (`source: bundle`), `session`, and `wire-capture`. The hosts and SDK list cross-check **infra-backend-fingerprint** (a host the binary calls that DNS/CT didn't surface is a finding); the runtime hosts also seed **wire-capture** targets (the binary says what it _might_ call; the capture says what it _does_).
 
 **Discipline:** EULA before extraction — restrictive license ⇒ Open question, never silent decompile. Manifests/strings only; no protected-logic reversing. Redact every embedded key before any `raw/` write (§7 rule 3). Digest large bundles (§5). A missing artifact type is a recorded gap, not a guess.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+the manifests verbatim (`manifest.json`, `Info.plist`, `AndroidManifest.xml`) and the extracted file listing.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
 '''

package/data/harnesses/discover/codex/agents/docs.toml

@@ -32,5 +32,31 @@ You collect a target's docs into `research/<target>/dimensions/docs/`. Read `.cl
 
 A one-paragraph headline (the product's documented surface in a sentence) + the section list with page counts + anything that overlaps the api/packages dimensions (flag it so synthesis can cross-check).
 
-**Discipline:** distill, don't mirror. A 300-page docs site becomes a navigable map + the load-bearing content, not 300 verbatim files. Record paywalled/gated docs as Open questions.
+**Discipline:** distill **in addition to**, never instead of. Save each captured page's source form under `raw/pages/` (md or html, as fetched) — that is the evidence layer — and build the navigable map + load-bearing content as the narrative layer beside it (ingestion §5.1). A distilled section whose source page was never saved is unverifiable by the next run (§5.3).
+
+**Do not mirror the SITE:** A 300-page docs site becomes a navigable map + the load-bearing content, not 300 verbatim files. Record paywalled/gated docs as Open questions.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+each captured page's source form under `raw/pages/` (md or html, as fetched).
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
 '''

package/data/harnesses/discover/codex/agents/infra-backend-fingerprint.toml

@@ -61,4 +61,28 @@ Return a **one-paragraph headline** of the reconstructed backend — origin clou
 **Discipline:** read-only public recon — never scan, brute-force, or hit a non-public host; respect crt.sh/WHOIS rate limits and back off; redact any credential-shaped value before writing (contract §7 rule 3 — DSNs, verification tokens, embedded keys surfaced in headers/TXT). An unresolvable host or a missing sub-processor page is an Open question, not a guess.
 
 **Verify an absence before asserting it (contract §7 rule 10).** Before reporting any host as **NXDOMAIN / absent** — especially `api.<target>` or any host a sibling dimension (bundle / session / wire) reports actually hitting — re-resolve via a **second resolver** (`dig @1.1.1.1 <host>`, `dig @8.8.8.8 <host>`) **and** a direct `curl -sI https://<host>`. An HTTP 4xx/5xx response means the host **resolves** → report it **present**, not absent. A single negative resolver result is recorded as _tentative_, never stated as fact and never used to instruct sibling dimensions. Likewise, an _inferred_ backend engine (e.g. "looks like Hasura/Supabase from the CNAME") is phrased as an inference and yields to a runtime-observing dimension (`session`/`deployed-client-bundle`) that names the engine directly.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+the raw `dig`, `crt.sh` and HTTP header responses as returned; the decoded table is the narrative layer beside them.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
 '''

package/data/harnesses/discover/codex/agents/packages.toml

@@ -27,4 +27,28 @@ You collect a target's installable artifacts into `research/<target>/dimensions/
 A one-paragraph headline (the package family in a sentence) + the package table + which packages are the core public API (so synthesis can anchor the SDK surface). Flag overlap with the docs api-reference and the codebase dimensions for cross-check.
 
 **Discipline:** the exports map + type defs are the load-bearing artifact — prioritize them over download-count trivia. Digest huge `.d.ts` to signatures. Redact any registry tokens.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+per package: the registry metadata JSON (`npm view <pkg> --json > raw/<pkg>.registry.json`, PyPI `/pypi/<pkg>/json`) and the `.d.ts` / exports surface as fetched.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
 '''

package/data/harnesses/discover/codex/agents/session.toml

@@ -38,4 +38,44 @@ Under `research/<target>/dimensions/session/`:
 A one-paragraph headline (the auth model + transport + the single most load-bearing wire finding) + the phase coverage + which passes ran. Flag drift between the _observed_ wire surface and the _published_ api/packages surface — that drift is the highest-value synthesis input. If Pass 2 was needed but not authorized, list the open write-side questions; never fabricate request bodies.
 
 **Discipline:** ethics first, always. Redact before writing. Confirm before any state change. No destructive actions. No exfiltration. A gap is an Open question, never a guess.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+≥3 verbatim envelopes as `raw/*.json` — the identity read, one canonical list read, and the primary write (request + response) — per ingestion §8.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
+
+## Visual capture (ingestion §5.5) — screenshot generously
+
+You drive a real browser, so you carry the visual-evidence obligation. **Capture continuously, not
+one-image-per-surface:** every surface, plus every materially different state within it — empty vs
+populated, modal/drawer open, dropdown expanded, error and validation states, permission/plan walls,
+loading/in-flight, and the before/after pair around any authorized state change.
+
+This is **not** scored — `screenshot_coverage` is a narrow Mode-5 floor. The beneficiary is the human
+reading this corpus weeks later for a purpose nobody anticipated. The product's past state is the one thing
+a later run genuinely cannot reproduce.
+
+**Redact in the DOM before the shutter** (`tradecraft.md` §1c) — a leaked pixel is undetectable afterwards.
+**Check the corners of the frame**, not just the content: support-chat widgets, toasts, presence avatars and
+member lists routinely carry *another human's* name and message. Then look at the saved image, and record
+`redacted: <what>` in `screens/_index.md` rather than a blanket zero-PII claim.
 '''

package/data/harnesses/discover/codex/agents/website.toml

@@ -33,4 +33,28 @@ You collect a target's public-facing site into `research/<target>/dimensions/web
 A one-paragraph headline (the positioning in a sentence) + the pricing tiers + the headline feature claims. Flag claims that should be cross-checked against the codebase/docs (marketing often over-claims) so synthesis can reconcile.
 
 **Discipline:** quote positioning verbatim; capture pricing precisely (numbers + limits); separate _claims_ from verified capability — the reconciliation happens in synthesis. Respect robots.txt and rate limits; this is read-only public crawling, not scraping at scale.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+the raw HTML or extracted text per captured page, and the pricing table as structured data.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
 '''

package/data/harnesses/discover/codex/agents/wire-capture.toml

@@ -42,4 +42,44 @@ Then the captures + raw artifacts (per `discovery` Part C (wire-capture)):
 Return a **one-paragraph headline** of what the wire reveals — the real hosts contacted, the envelope/auth scheme, and the rung used (so the reader knows how much to trust it). Then a compact table of observed endpoints. **Flag the seam rule (§6):** every `api-path-catalog.md` row is `source: wire`, so evaluation can diff _actually-called_ (wire) against _statically-declared_ (`bundle` / `distribution-artifacts`) and _browser-observed_ (`session`) — a host or path the binary declares but the wire never hits is a finding, not noise; the inverse (called but undeclared) is a stronger one. If only rung 1 was authorized, say so and name what rungs 2/3 would unlock so the main session can request authorization.
 
 **Discipline:** lowest privilege first; never self-escalate the ladder; never install a CA or tap third-party traffic; never run Wireshark; redact every body before disk and log it in `_meta`. An unauthorized higher rung is an Open question, not an excuse to guess what the wire would show.
+
+---
+
+## Dump manifest (ingestion §5.3–5.4) — evidence before narrative
+
+**MUST land in `raw/` as verbatim, redacted files before this dimension may be `status: complete`:**
+the flow log itself (HAR or the tap's JSON export), not only the decoded narrative.
+
+Three rules bind this, and they override any instinct to summarise:
+
+1. **If you cite it, dump it (§5.3).** Every value that reaches your `_summary.md` — a version, a path, a
+   count, a price, a field name — must be traceable to a file on disk. **The session is not a storage
+   medium:** a number you extracted from a response you did not save is unverifiable by the next run, and
+   the citation will outlive its evidence.
+2. **Digest IN ADDITION, never INSTEAD (§5.1).** `raw/` is the evidence layer and may be unreadable;
+   `_summary.md` is the narrative layer. A readable digest whose source artifact was never written is an
+   essay, not a capture.
+3. **Size is a routing problem (§5.2).** Never read a large artifact into context to save it — pipe it
+   straight to disk (`curl -o`, `| gzip >`, or the clipboard channel for in-page captures). If you must
+   bound an unbounded source, record the bound in `_meta.sampling`. Context budget is never a reason to
+   discard evidence.
+
+**Self-check before returning:** count the verbatim files in `raw/`. Zero ⇒ the dimension is `partial`,
+regardless of how good the prose is.
+
+## Visual capture (ingestion §5.5) — screenshot generously
+
+You drive a real browser, so you carry the visual-evidence obligation. **Capture continuously, not
+one-image-per-surface:** every surface, plus every materially different state within it — empty vs
+populated, modal/drawer open, dropdown expanded, error and validation states, permission/plan walls,
+loading/in-flight, and the before/after pair around any authorized state change.
+
+This is **not** scored — `screenshot_coverage` is a narrow Mode-5 floor. The beneficiary is the human
+reading this corpus weeks later for a purpose nobody anticipated. The product's past state is the one thing
+a later run genuinely cannot reproduce.
+
+**Redact in the DOM before the shutter** (`tradecraft.md` §1c) — a leaked pixel is undetectable afterwards.
+**Check the corners of the frame**, not just the content: support-chat widgets, toasts, presence avatars and
+member lists routinely carry *another human's* name and message. Then look at the saved image, and record
+`redacted: <what>` in `screens/_index.md` rather than a blanket zero-PII claim.
 '''

package/data/harnesses/discover/harness.json

@@ -1,12 +1,11 @@
 {
   "harness": "discover",
   "name": "discover",
-  "version": "0.1.12",
+  "version": "0.1.14",
   "files": [
     ".prettierignore",
     ".prettierrc.json",
-    ".vscode/settings.json",
-    "tooling-recommendations.md"
+    ".vscode/settings.json"
   ],
-  "digest": "48356bdb7940"
+  "digest": "4ab9f67f0d67"
 }

package/data/harnesses/discover/rules/calibration.md

@@ -0,0 +1,219 @@
+# Calibration Reference — Blind Re-runs · Disposition · The Fleet Ledger
+
+The reference for the harness's **only external signal**. Everything else in this harness is self-assessment:
+collectors grade their own `confidence` and `completeness_pct`, Evaluation weights those self-grades, and
+Mode 5 scores the result against a checklist the same system authored. That loop can converge on a corpus
+that is confidently, well-provenanced, and wrong.
+
+**Calibration is the one procedure that can tell you whether any of it is true.** It re-derives claims a
+prior run published, blind, and records what survived.
+
+> **The measured position, as of 2026-07-28.** One target has ever been re-run (`emergent`, 2026-06-25 →
+> 2026-07-28). Of 16 claims re-tested: **7 refuted, 4 corrected, 5 confirmed** — a 44% flat-falsification
+> rate — and **4 of the 7 refutations came from lanes the prior run had graded `confidence: high`.** The
+> prior run had self-scored **88/100 "strong"** and `stopped_on: converged`.
+>
+> That is **n=1**. It is the entire empirical basis of this harness. Every rule in `.claude/rules/` is
+> currently a hypothesis with n=0 on any target it did not come from. The purpose of this file is to make
+> n grow.
+
+---
+
+## Two tiers — pick the one that will actually happen
+
+A metric too expensive to run leaves `n` at 1 forever. The cheap tier is the important one.
+
+| Tier | Cadence | Cost | What it measures |
+| --- | --- | --- | --- |
+| **A — Claim spot-check** | monthly | ~1 session, no full ingestion | the refutation rate, and whether `confidence` tracks truth |
+| **B — Full blind re-run** | quarterly, or when a rule set changes materially | a complete run | everything Tier A does, plus coverage drift and whether new rules changed behaviour |
+
+**Tier A is the default. Do not skip it waiting for capacity to do Tier B** — a monthly spot-check of 15–20
+claims across one target produces the regression that matters, and produces it twelve times a year.
+
+---
+
+## Selection — rotate, never choose
+
+**The target is selected by rotation, not by judgment.** Take the least-recently-calibrated target in the
+corpus root; break ties by oldest `captured_at`. Record the selection rule alongside the result.
+
+A human (or a run) choosing which target to re-check will choose the one it feels good about. That single
+substitution destroys the metric, because the sample stops being representative of the fleet and starts
+being representative of the harness's confidence — which is the very thing under test.
+
+---
+
+## THE BLIND RULE (this is what makes it evidence)
+
+**Capture first. Read the prior corpus only after your own observation is written down.**
+
+1. Select the target. **Do not open its corpus.**
+2. Draw the claim sample (below) — a list of *questions*, not answers. Someone or something other than the
+   re-deriving agent extracts the questions, or the questions are extracted mechanically from claim
+   anchors, so the sample carries no prior conclusions.
+3. **Re-derive each claim from the live product**, from scratch, recording what you observe and the method.
+4. **Only then** open the prior corpus and diff.
+
+> **This deliberately overrides `discovery.md` Step 00 for calibration runs.** Step 00 tells a run to find a
+> prior corpus, restore it, and reconcile against it — correct for a normal re-run, where the prior corpus
+> is a genuine corroboration lane and re-deriving everything from zero is waste. **It is wrong here.** A run
+> that reads the prior claim first will anchor on it: it will look for confirmation, find it, and report a
+> confirmation rate that measures its own suggestibility. Anchoring is the specific failure this protocol
+> exists to avoid, so for a calibration run Step 00 is deferred until after step 3, and the run records
+> `blind: true`.
+>
+> If a run cannot honestly attest that it re-derived before reading, it must record `blind: false` — and
+> the result is then a **consistency check, not a calibration**, and does not enter the ledger's refutation
+> rate.
+
+**Sampling the claims.** 15–20 per target, drawn to make the regression computable, not to be flattering:
+
+- **Stratify by the prior run's stated `confidence`** — roughly equal numbers of `high`, `medium`, `low`.
+  Without this the regression cannot be computed, and the regression is the point.
+- **Include the load-bearing ones** — claims a README headline rests on, and claims a reader would act on
+  (pricing, limits, guarantees, auth model, what the product can and cannot do).
+- **Include at least 3 negative claims** ("no public SDK", "no realtime layer", "not supported"). Negatives
+  are the harness's demonstrated weak spot: **5 of the 7 refutations in the one data point were negatives.**
+- **Exclude** anything whose truth cannot change and cannot be re-checked (a historical funding round).
+
+---
+
+## Disposition — five values, and the fifth is the one the old schema lacked
+
+Every sampled claim resolves to exactly one:
+
+| Disposition | Meaning | What it says about the harness |
+| --- | --- | --- |
+| **confirmed** | independently re-derived, same conclusion | the method worked |
+| **corrected** | right direction, wrong detail (a count, a version, a name) | the method worked; the reading was imprecise |
+| **refuted** | the claim was **wrong when it was made** | **the method failed — this is the signal** |
+| **stale** | the claim was **true when made; the product changed** | the method worked; the corpus has a shelf life |
+| **unverifiable** | cannot be re-checked now (surface gone, plan changed, gated) | no signal either way; excluded from the rate |
+
+> **`refuted` and `stale` are the distinction the original artifact had no slot for, and they mean opposite
+> things.** A high `refuted` rate means *the instrument is unreliable* and the rules need work. A high
+> `stale` rate means *the instrument is fine and the corpus decays* — which calls for shorter shelf-lives
+> and `volatility` tagging, not methodology changes. Conflating them makes the harness fix the wrong
+> problem. When you cannot tell which it is — no dated evidence either way — record **`refuted`**. The
+> conservative reading is the one that keeps the harness honest about itself.
+>
+> (Origin: `emergent/00-prior-run-reconciliation.md` — a genuinely good artifact whose section A had to be
+> titled *"REFUTED (prior claim was wrong, not merely stale)"* precisely because the schema gave it nowhere
+> else to put the distinction.)
+
+**The refutation rate** = `refuted ÷ (confirmed + corrected + refuted + stale)`. `unverifiable` is excluded
+from the denominator, and its count is reported so a shrinking sample is visible.
+
+---
+
+## The regression — the actual question
+
+For each calibration, record the disposition **against the prior run's stated confidence for that claim**:
+
+| prior confidence | n | confirmed | corrected | refuted | stale | refutation rate |
+| --- | --- | --- | --- | --- | --- | --- |
+| high | | | | | | |
+| medium | | | | | | |
+| low | | | | | | |
+
+**If `high` claims are refuted at a similar rate to `low` claims, the `confidence` field carries no signal
+and `ingestion.md` §3's method→confidence table is decorative.** That is the single most consequential thing
+this harness could learn about itself, and it is unknowable without this table.
+
+Also record refutation rate **by `method`** once n is large enough — that tells you which of the 17 methods
+in §3 are over-graded, which is directly actionable (a method's band gets lowered).
+
+---
+
+## Artifacts
+
+### Per calibration — `research/<target>/00-prior-run-reconciliation.md`
+
+Formalizes the shape `emergent` invented. **This file is part of the output contract** whenever a prior
+corpus exists.
+
+```markdown
+---
+target: <slug>
+reconciled_at: <YYYY-MM-DD>
+prior_run_date: <YYYY-MM-DD>
+prior_run_commit: <sha>
+prior_run_score: <n>
+tier: A | B                     # spot-check or full blind re-run
+blind: true | false             # false ⇒ consistency check only, excluded from the rate
+selection_rule: rotation        # or: named-by-user (record it either way)
+sampled: <n>
+confirmed: <n>
+corrected: <n>
+refuted: <n>
+stale: <n>
+unverifiable: <n>
+refutation_rate: <0-100>
+by_confidence:                  # the regression — REQUIRED
+  high:   { n: <n>, refuted: <n> }
+  medium: { n: <n>, refuted: <n> }
+  low:    { n: <n>, refuted: <n> }
+---
+
+## A. REFUTED — wrong when made        (claim · what was observed · lane that refuted it · prior confidence)
+## B. STALE — true when made, changed  (claim · what changed · evidence of the change being NEW)
+## C. CORRECTED — right, imprecise
+## D. CONFIRMED — independently re-derived
+## E. UNVERIFIABLE — and why
+## F. What this says about the METHOD  (which lanes/methods produced the refutations, and whether a
+                                        §3 confidence band should be lowered)
+```
+
+### Fleet-level — `research/_calibration/ledger.md`
+
+**The metric itself.** One row per calibration, accumulating across targets and time. This is the file to
+read when asking "does this harness work?" It is never rewritten, only appended to.
+
+---
+
+## What a result CHANGES (otherwise this is just bookkeeping)
+
+A calibration that produces a number and no action is theatre. Each outcome routes:
+
+| Finding | Action |
+| --- | --- |
+| A **method** shows a high refutation rate | lower its band in `ingestion.md` §3, and record the calibration as the origin |
+| **Negative claims** refute disproportionately | tighten the absence rules (`ingestion.md` §7 rule 10, `discovery.md` Part B) — the `blind_to:` direction |
+| `high` refutes ≈ as often as `low` | **the confidence field is not carrying signal** — this is a structural finding about the harness, and it goes to the user, not into a rule tweak |
+| A **rule added after the prior run** would have prevented a refutation | evidence the improvement loop works — record it, it is the only such evidence that exists |
+| A refutation recurs **after** a rule was added to prevent it | evidence a rule did **not** change behaviour; escalate to the user rather than adding another rule |
+| High **stale** rate | shorten shelf-life expectations and tag volatile claims; **not** a methodology problem |
+
+---
+
+## How to run one (Tier A, the monthly default)
+
+1. **Select** by rotation from `research/_calibration/ledger.md`. Record the target and the rule.
+2. **Extract the sample** mechanically — pull claim anchors from the target's `evaluation/*.md`, stratify by
+   the cited confidence, and write them as **questions with the answers stripped**.
+3. **Re-derive blind.** Do not open the corpus. Use whatever lanes the access vector allows; record method
+   and observation per claim.
+4. **Diff and disposition** each claim into the five values.
+5. **Write** `00-prior-run-reconciliation.md` (append a new dated section if one already exists — never
+   overwrite a prior calibration).
+6. **Append one row** to the fleet ledger, and update its running totals.
+7. **Route** every finding through the table above. Anything that becomes a rule change goes through the
+   normal Mode-5 approval gate.
+
+**Cost/state discipline is unchanged:** a calibration is read-only unless the user authorizes otherwise; a
+claim that can only be re-derived by spending money or mutating state is recorded `unverifiable` with the
+reason, never quietly exercised.
+
+---
+
+## How the harness uses this reference
+
+Calibration is a **cross-run process**, not a mode of a single run — nothing in Modes 1–5 triggers it, and it
+must not be folded into them (a run cannot calibrate itself; that is the circularity it exists to break).
+Mode 5 reads the fleet ledger as *context* when weighing a proposal — a rule contradicted by calibration
+data should not be promoted — but Mode 5 never writes to it.
+
+> **Do not attach calibration to the Mode-5 score.** The moment `refutation_rate` becomes a rubric input,
+> the run acquires an interest in the number, and the one uncorrupted signal in the system is gone. It
+> measures the **harness**, not the run. Publish it; never grade with it.

package/data/harnesses/discover/rules/cartography-coverage.md

@@ -33,6 +33,40 @@ differently:
 | **edition/plan-gated** | not in this tenant's edition/plan (e.g. a trial) | a recorded gap — **not** re-walked (re-walking can't find what isn't shipped here) |
 | **roadmap-not-shipped** | marketed but not yet a shipped capability | a recorded gap |
 | **marketing over-claim** | the claim overstates what the product does | an `evaluation/product-features.md` **over-claim flag** |
+| **not-yet-reachable** | it exists and is shipped, but the surface **does not render at all** until a state change (deploy / connect / upgrade / first-run) that this run did not perform | the **gated re-walk queue** — NOT the read-only auto-loop |
+
+### `not-yet-reachable` vs `deeper-than-looked` — the distinction that matters
+
+These two look identical in a coverage matrix and route **oppositely**, so get it right:
+
+- **`deeper-than-looked`** — the surface **is in the DOM right now**; the run just didn't navigate to it.
+  More read-only walking finds it. Routes to the **read-only re-walk queue**, which Mode 5's auto-loop
+  works for free.
+- **`not-yet-reachable`** — the surface **is not in the product yet.** No amount of navigation reveals it,
+  because the affordance is created by a state change. Routes to a **separate gated queue** that the
+  auto-loop must **not** touch: closing it costs money, mutates state, or needs a third-party consent, so
+  it goes through the explicit-confirmation gate (ingestion §7 rule 4) or is recorded as a standing gap.
+
+**Mis-filing a `not-yet-reachable` row as `deeper-than-looked` is a Cartography defect**, because it sends
+the auto-loop hunting for something that does not exist, and it reads as "we were sloppy" when the honest
+statement is "the product's IA is state-dependent and this run stayed inside its authorization."
+
+**A run that never performs the state change cannot map a state-dependent IA — say so rather than
+scoring it as a coverage miss.** Note it beside the scorecard the same way `experiential_gated` annotates
+a policy deferral: it documents, it does not inflate.
+
+> (Origin: Emergent — a deployment executed mid-run **created** two whole surfaces: the entire *Manage
+> Publishing* panel with five sub-tabs, and the `db_tool` database-console affordance, which opened a
+> separately-hosted plugin on a second auth system. Neither exists in the toolbar of an undeployed project.
+> The pre-deploy coverage doc had filed the publish family under *deeper-than-looked* and queued it for a
+> read-only re-walk that could never have succeeded. Note also the **two-factor gate**: the cohort flag
+> `db_tool: true` was necessary but **not sufficient** — an entitlement flag reading `true` does not mean
+> the surface is reachable.)
+
+**Where this bites generally:** any PaaS or marketplace with deploy/publish-gated management surfaces; any
+product whose integrations render only after an OAuth connect; any tier-gated admin console; any
+first-run/onboarding-gated dashboard. Treat an entitlement flag as **necessary, not sufficient** — check
+whether the *state* it depends on has also been reached.
 
 ---
 
@@ -50,6 +84,7 @@ coverage_scorecard:
   ia_surfaces_mapped: <N>          # surface cards written in information-architecture.md
   ia_nav_complete: <true|false>    # every top-level nav destination has a surface card
   screenshot_coverage: <0-100>     # primary surfaces with a screenshot ÷ all primary surfaces
+  features_not_yet_reachable: <N>  # shipped, but behind a state change this run did not perform
 ---
 ## Coverage matrix
 | Feature (claimed) | Source | Located? | Route · depth | Walked? | Edition-gated? | Confidence | Note |
@@ -57,6 +92,9 @@ coverage_scorecard:
 | <feature> | website,docs | ✅ | #/…/x · D2 | ❌ not-walked | — | high | promoted on site, 2 levels deep, not exercised this run |
 | <feature> | website(AI) | ⚠ edition? | — | ❌ | maybe-not-in-trial | low | marketed flagship, not located in this tenant |
 ## Located-but-not-walked   (the within-run re-walk queue — Mode 5 walks these read-only, no cost)
+## Not-yet-reachable        (the GATED queue — shipped surfaces behind a deploy/connect/upgrade this run
+                             did not perform. Mode 5 must NOT auto-walk these; each names the exact state
+                             change that would unlock it and its cost/consent class)
 ## Claimed-but-not-located  (each row dispositioned per the taxonomy above — NEVER left blank)
 ## Buried flagships          (marketing-promoted features sitting ≥D3 — the positioning-vs-product gap,
                              drawn from information-architecture.md's promoted-vs-buried headline)
@@ -69,6 +107,10 @@ coverage_scorecard:
 experiential / IA coverage axis (`self-correction.md` rubric, axis 2, max 20) is reproducible:
 
 - `feature_location_rate` = `100 × features_located ÷ features_claimed`.
+- **`not-yet-reachable` rows stay in the `features_claimed` denominator** — they are real, shipped,
+  unlocated capabilities and hiding them would inflate the rate. But they are **dispositioned**, so they
+  never draw the −2 un-dispositioned penalty, and the scorecard carries `features_not_yet_reachable` so a
+  reader can tell a *policy/authorization* boundary from a *thoroughness* failure.
 - `flow_coverage` = `100 × (primary journeys diagrammed in ux-flows.md ÷ journeys identified)`.
 - `screenshot_coverage` = `100 × (primary surfaces with a screenshot ÷ all primary surfaces)`.
 - `ia_nav_complete` = every D0 nav destination has a surface card.
@@ -81,6 +123,13 @@ un-dispositioned. The `located-but-not-walked` list is the re-walk queue Mode 5
 
 ## Screenshots — `dimensions/session/captures/screens/`
 
+> **The scorecard metric is a FLOOR, not the capture rule.** `screenshot_coverage` asks a narrow question —
+> did each *primary surface* get one image — because Mode 5 needs something countable. **The actual capture
+> obligation is `ingestion.md` §5.5: screenshot generously, including states within a surface, for the human
+> who reads this corpus later.** Satisfying this metric and stopping is the failure mode; the metric is
+> nearly always already met by a real walk. Nothing here caps how much you capture, and extra images never
+> hurt the score.
+
 One annotated screenshot per **primary** surface (each D0 nav destination + each materially distinct
 detail / editor surface). **Redact visible PII before saving** (user names, emails, customer data,
 partner identifiers). Name `<nn>-<surface>.png`; write a one-line `screens/_index.md` mapping each file
@@ -89,6 +138,25 @@ partner identifiers). Name `<nn>-<surface>.png`; write a one-line `screens/_inde
 **Screenshots are the Mode-4 deliverable that the session playbook's textual surface-map only
 _substitutes_ for** (ingestion §9.5) — capture them live here; fall back to the textual map (and record
 the `screenshot_coverage` gap) only when the runtime genuinely cannot screenshot (no image tool, or the
-app never reaches page-idle because of a persistent connection). The screenshots live under the
+app never reaches page-idle because of a persistent connection).
+
+> **Re-test the capability before recording `screenshot_coverage: 0` — an environment limitation claimed
+> by a PRIOR run is `verify:`, never inherited.** Take one screenshot with `save_to_disk: true` in the
+> current session and record the literal tool result. Only an actual failure in *this* session justifies
+> the textual fallback. (Origin: Emergent — a run recorded `screenshot_coverage: 0` because "`save_to_disk`
+> returned no persistable path in this environment", and the harness codified that as an accepted
+> fallback. In the same environment it works and returns a real path; the next run captured 13 images. An
+> unverified environment claim had become a standing excuse that suppressed the most heavily-weighted
+> rubric axis.)
+>
+> **Validate the coverage doc by READING THE FILE, not by trusting a delegated agent's return.** Because
+> Mode 5 scores this file's `coverage_scorecard` frontmatter directly, a **stale** `feature-coverage.md`
+> is worse than a missing one — it scores silently, against the *previous* run's numbers. Assert (a) the
+> file's mtime is newer than the dispatch and (b) its frontmatter matches this run's facts
+> (`write_side_observed`, `screenshot_coverage`, the credit figure). (Origin: Emergent, second run — a
+> sub-agent delivered its IA doc and silently dropped this one; the untouched Pass-1 file would have
+> scored a full write-side run at `screenshot_coverage: 0` / `write_side_observed: false`.)
+
+The screenshots live under the
 `session` dimension (it owns the browser captures); `information-architecture.md`'s surface cards
 reference them by path.