# Cejel rubric changelog

Every change to how Cejel scores a repository is recorded here, with a full before/after
delta across the published corpus — score, verdict, and rank for every repository,
"no repository moved" stated explicitly when that is the result. A rubric version bump
that ships without an entry here is a bug, not a release.

**Two controls enforce that rule, and they see different things.** The first is the
rubric-rescore-protocol regression guard in the source monorepo's test suite, which fails the
build if `WITAN_RUBRIC_VERSION` changes without a matching entry below. It is keyed on an
identifier this project's own authors control, so it cannot see a scoring change made *under an
unchanged identifier* — and that is not a hypothetical, it happened five times in 0.4.9 under the
calibrated public default (see the entry below it, and the 2026-09-16 entry that records the
miss). The second, added on 2026-09-16, does not read the identifier at all:
`src/witan/__tests__/rubric-behaviour-fingerprint.test.ts` scores a committed corpus of synthetic
repositories under every selectable rubric and fails when the measured result moves away from a
committed pin, naming the rubric and the criteria. Every fingerprint it pins is cited in this file,
so re-pinning one requires writing here.

This changelog exists because a public leaderboard that can silently re-score another
repository is not a standard, it is a rumor with a number attached — see this repository's
README, "The public leaderboard: what we redact, what we exclude, and where we were wrong"
section, which this changelog continues.

## 2026-09-16 — the enforcement named above could not see the change it was written to catch

**Status.** No scoring change. This entry records a MISS in this changelog's own enforcement, and
the control added to close it. No repository's score, verdict, or rank moves because of anything
recorded here; the published corpus is untouched and was not re-scored. `WITAN_RUBRIC_VERSION`
does not change. What changes is what can be detected, and what a certificate discloses.

**The miss, stated plainly.** Until today the single enforcement this file named was a guard that
fails the build when `WITAN_RUBRIC_VERSION` changes without a matching entry here. That guard is
keyed on an identifier this project's own authors control. **A change to how Cejel scores a
repository, made under an unchanged identifier, is invisible to it — not by accident, but by
construction.** The guard is correct on its own terms and will never fire for that case.

This is not a theoretical gap. On 2026-09-15, five changes in the unreleased 0.4.9 altered scoring
under `witan-rubric-v17-2026-07-24` — the calibrated public default — with the identifier
unchanged. Two raised scores, one lowered them, two more widened detectors. The guard did not
fire. Nothing else noticed until a human compared two arms by hand. The entry immediately below
this one records that delta in full; this entry records that the control which was supposed to
compel that entry had no way of knowing it was needed.

The consequence is the one this product exists to prevent. Two certificates can both say
`witan-rubric-v17-2026-07-24`, be produced at the same revision by different tool versions, carry
different scores, and neither artifact nor the build record says why. The recorded defence against
a well-resourced clone is that the public calibration record gives certificates their meaning. A
record that cannot tell when behaviour changed is not a defence.

**What now detects it.** `src/witan/__tests__/rubric-behaviour-fingerprint.test.ts`, which never
reads `rubricVersion` to decide whether anything moved. It builds a committed corpus of thirteen
synthetic repositories (`src/witan/__tests__/fixtures/behaviour-corpus.ts`), scores every one of
them under every selectable rubric, projects each report down to its scoring-relevant output, and
compares that against committed pins. The failure names the rubric that moved and the criteria
that moved, because the pins are decomposed per criterion and per fixture before they are
digested. It runs in CI on every pull request (about nine seconds).

Per rubric, never aggregate: a combined digest would hide a change that moves a prospective rubric
and leaves the calibrated default alone. **The v17 digest moving is the serious case, and it is
separately visible.**

**What the digest covers**: criterion scores and statuses (which is where `insufficient_data`
surfaces), per-criterion metric values, per-criterion finding counts by severity, the repo
archetype, `contentReadSummary`, the report verdict, the composite scores, and the
insufficient-source abstention reason. **What it deliberately excludes**: everything
per-invocation (product identity, repository path, `generatedAt`, `toolVersion`, and the
fingerprint field itself), and all prose (finding summaries, metric labels, evidence labels,
criterion notes, `scanLimitations`) — a wording improvement is not a scoring change, and a guard
that fires on copy edits is one people learn to re-pin without reading. A finding appearing or
disappearing still moves its severity count, so the detector change is caught while the sentence
it produces stays editable.

**What it does not claim.** A matching fingerprint says no *corpus-visible* scoring behaviour
changed. It is not a coverage claim and must never be quoted as one. The corpus measures A1, A2,
A3, A4, A5, B2, B3, B4 and B6. It does not measure **B1 (dispatch trace completeness)** or **B5
(verified learning trace)**, and no fixture could: both lost their collectors in
`witan-rubric-v3-2026-07-13` because they were reachable only from the source monorepo's own file
paths, and no public repository scan has emitted a signal for either since. The guard asserts that
this gap is exactly `{B1, B5}`, so it can neither widen nor quietly close.

**A determinism limitation this corpus made visible.** Up to and including
`witan-rubric-v18-prospective-2026-07-25`, B4's audit-freshness metric rates an audit artifact
against the *scan* year rather than anything in the repository — `witan-rubric-v19` is the
correction, and the `stale-audit-trail` fixture is where the corpus shows the difference. The
fingerprints below are therefore measured at a pinned `generatedAt` of `2026-09-15`. For v17 and
v18 the fingerprint is an identity for scoring *at that scan year*; a scan run in a later calendar
year can score a repository with a dated audit trail differently under those two rubrics without
any code changing. That has been true since v17 shipped and is not introduced here. It is written
down here because a behaviour fingerprint that quietly depended on a clock would be the same class
of defect this entry is about.

**Emitted in the artifact.** `report.json` now carries `rubricBehaviourFingerprint` beside
`rubricVersion`, so rubric identity and behavioural identity are separately checkable by whoever
receives the certificate rather than only by this project's CI. That bumps the report contract to
`reportFormatVersion` 1.2 (`docs/format-stability.md`); the field is additive-optional, so a
consumer following the v1 contract to ignore unknown optional fields remains compatible.
Strict field-allowlist readers require an update; release verification found the
published-package board reader rejecting this new key until its schema was updated. **Reports produced by earlier versions do not carry
the field and their existing attestations remain valid** — absence means "produced before report
format 1.2", never "behaviour unknown and therefore suspect". The value is a committed constant
rather than something recomputed per scan (recomputing it costs thirteen git repositories and
ninety-one scans); what makes it true is the guard above, which re-derives it from the corpus on
every CI run and fails if it has drifted.

**This does not weaken report reproducibility.** `report.json` remains byte-identical for the same
tool version scanning the same revision under the same rubric. The new field is a property of the
(build, rubric) pair, derived from nothing but `rubricVersion`; it contains no timestamp, no path,
and no per-run state. It changes report bytes across *versions*, exactly as `toolVersion` already
does, and not across *runs*.

**The fingerprints, as of this entry.** Measured on the corpus at its pinned commits, at
`generatedAt` `2026-09-15T00:00:00.000Z`. Changing any of these values requires a new entry in
this file quoting the new digest — that is the mechanism that stops a re-pin from being a silent
repair, and it is checked mechanically.

| Rubric | Behaviour fingerprint |
| --- | --- |
| `witan-rubric-v17-2026-07-24` (calibrated public default) | `sha256:ff0f01abe8c12daa60375d0e18c1aca4a1135a55ab2a80b15f4c037bfc18b4b1` |
| `witan-rubric-v18-prospective-2026-07-25` | `sha256:abc35df0fdc0749aa941f47b295059d06c96dd9800af7f79cfef2bc8dae917da` |
| `witan-rubric-v19-prospective-2026-08-09` | `sha256:f670fa2cc44beef066fa38767bd587e1a16fb28b03bc904756daaea5a6465e96` |
| `witan-rubric-v20-prospective-2026-08-10` | `sha256:9276ce2ce865c2d50c52882ee45bbf1df5fe3c8b6038858b4a147f7a4cedfa98` |
| `witan-rubric-v21-prospective-2026-08-10` | `sha256:d86411a7fe108bba4d2b8fe7ff9d4f7f69a60f11ca0c5c3fef8e88f66a9af53a` |
| `witan-rubric-v22-prospective-2026-08-10` | `sha256:1bb57dd98d29b43b170f3a4f1c8b3b379b433a2aa9df0821ab9d5115f48cfb2f` |
| `witan-rubric-v23-prospective-2026-09-06` | `sha256:b558c0aa4b2f85b734bed2e077298fd55851a1086234d678f10a71cf913ec3cf` |

These seven values are pairwise distinct, and the guard asserts it: the corpus can tell every
selectable rubric apart from every other, which is the minimum evidence that it reaches what
separates them. Each prospective rubric moves at least one criterion on at least one fixture
relative to its predecessor — v18 on A2 (native RLS policy reading), v19 on B4 (freshness year),
v20 on A3 (direct HTTP entrypoint), v21 on B6 (authored administrative SQL), v22 on A3
(start-declared entrypoint), v23 on A1 (coverage-capable command flags) and A5 (withheld-path
abstention).

**What this control still cannot do.** It cannot tell whether the changelog entry a re-pin cites
is *truthful* — only that one exists and names the new digest. It cannot see a scoring change on a
repository shape the corpus does not contain; recall gaps are a known limitation of every static
control here and are priced in. And it cannot detect a change made by an author willing to edit
the guard, the corpus and this file together. It closes the specific hole that a change made in
good faith, under an unchanged identifier, ships with nobody knowing.

**Scope note.** The fixture corpus is a *detector*, not a calibration record. It is synthetic,
invented for this purpose, and carries no precision, recall, or false-positive claim. The
published corpus (`leaderboard/corpus.json`) remains the only thing a before/after delta in this
file is measured on.

## 0.4.9 — behaviour change under witan-rubric-v17-2026-07-24 (no identifier bump)

**Status.** Disclosed behaviour change under the existing calibrated public default. **Not a
recalibration** — the v17 calibration frame is retired and this entry does not touch it. The
rubric identifier does not change. `WITAN_RUBRIC_VERSION` was deliberately not bumped by the
author of this entry; whether it should have been is handed to the operator below. Unlike the
two entries below it, this delta was **preregistered**: expected values were committed before
the comparison ran, and the result matched them in full.

**What changed.** Five changes in CLI release 0.4.9 change how a repository scores under
`witan-rubric-v17-2026-07-24` with no change to the identifier. None is gated on rubric
version; each applies to every rubric, the calibrated default included.

1. **`test`-script content check (direction: down).** A `package.json` `test` script was
   credited by key presence alone, so npm's default placeholder (`echo "Error: no test
   specified" && exit 1`) scored identically to a real runner in both A1's
   `verification_script_ratio` and B3's `ci_script_depth`. Both now require the script's
   content to match a known test-runner invocation.
2. **A1 authenticated-absence credits (direction: up).** On a revision with package-level
   `lint`/`typecheck` scripts but no test files, A1 zeroed its whole verification-script signal
   while B3 credited the same scripts. A1 now credits lint/typecheck/coverage on that path;
   only the test-file-dependent metrics stay at zero.
3. **Pull-request-template directory form (direction: up).** `.github/PULL_REQUEST_TEMPLATE/<name>.md`
   read as no template; it is now recognised alongside the single-file form.
4. **A3 observability-depth pattern widening (direction: up).** The file-content pattern behind
   `A3.observability_depth` matched a vendor-product list (`sentry`, `otel`, `opentelemetry`,
   `datadog`, `prometheus`, `logtail`) plus the generic substrings `logger` and `metrics`, and so
   missed the common Node structured-logging libraries (`pino`, `winston`, `bunyan`), the Express
   request-logging convention (`morgan`) and the request-correlation idiom (`correlation`,
   `request` or `trace` plus an optional separator and `id`, and `AsyncLocalStorage`). All of
   those now match, each word-bounded.
5. **A3 `prod_readiness_primitives` error boundary (direction: up).** The "error boundary"
   component matched a frontend filename convention only (`error-boundary.*`,
   `*.error.(tsx|jsx|ts|js)`) and read no file content, so an Express error-handling middleware
   layer — which can live in any file under any name — never moved it. It now also recognises the
   canonical four-argument `(err, req, res, next)` signature by content, across implementation
   files, checked only when the filename convention finds nothing.

Two further changes in this release are **not capable of moving a score under the calibrated
default** and are recorded so that the list above is not mistaken for the whole release: the
health/readiness route widening (gated on `useV20ExplicitGaps`, prospective `witan-rubric-v20`+,
and feeding an info-severity absence finding rather than a scored metric), and the `too_large`
abstention-disclosure split (gated to prospective `witan-rubric-v23`). Both were measured rather
than assumed inert: they moved nothing, as predicted.

**Why the cited guard did not fire, stated plainly.** The rubric-rescore-protocol guard this
changelog's preamble cites lives in the source monorepo (BargLabs/alfred,
`packages/witan/src/__tests__/rubric-rescore-protocol.test.ts`). It has two halves: a check
that `WITAN_RUBRIC_VERSION` has a changelog entry, and golden fixture scores keyed by rubric
version that *would* catch a behaviour change under an unchanged identifier — if the fixture
exercised the changed behaviour. Three things made it inert here. The identifier did not
change, so the first half had nothing to key on. The golden fixture carries
`test: 'vitest run'` with test files present and no PR template, so none of the three shapes
above exist in it. And it runs against alfred's copy of the scanner, which has diverged from
this repository's (6,522 lines against 8,913 at the time of writing); this repository carried
no equivalent guard of its own. A control keyed to a version string, with a fixture that does
not contain the changed shape, in a different repository, is the same defect class the fixes
above were closing: a check that reports on something it did not examine. This release adds
the cejel-native guard, `src/witan/__tests__/v17-scoring-surface-golden.test.ts`, whose four
fixtures are exactly the four shapes in this entry; it fails 4/4 against the v0.4.8 source and
passes 4/4 against 0.4.9, naming the metric and direction each time.

**Measurement.** All 24 corpus rows at their pinned commits (`leaderboard/corpus.json`,
sha256 `dc723f53…`, byte-identical to the corpus the v19 protocol froze), scored twice from
source with the calibrated default, `generatedAt` fixed, on one machine within one hour:
baseline `7606392` (the `v0.4.8` commit) and `9df6f31`, the final measured scoring tree.
The immutable `v0.4.9` tag resolves to `cd75fc435d33bccc84293c8d1ce1a5a99685d6c0`.
A comparison of these two candidate commits finds only the CLI changelog, this rubric
changelog, and the paired-result JSON/Markdown record changed; no scorer or package source
changed. The candidate arm was re-scored after every 0.4.9 code change had landed. The scoring work it carries
arrived as `dad7841` (A3 runtime-pattern coverage), `90371ea` (certificate scope disclosure) and
`cfd9b16` (the behaviour fingerprint and report format 1.2). The published `@cejel/cejel@0.4.8` npm artifact reproduces the baseline arm
exactly on the row that moved most (django: 3.2 / 2.6 / 3.8, B3 3.6, `ci_script_depth` 3), so
the baseline is what a customer has, not an assumption about it. Canonical evidence:
`docs/experiments/v17-behaviour-delta-0.4.9-2026-09-15/paired-result.json`; the harness that
produced it is beside it. Raw per-row reports were retained locally and not committed (the
private row's report is not public; the public rows' reports are reproducible from the
harness).

**Preregistered, and the prediction held.** This repository's convention for a measurement whose
result matters is a preregistration commit that is a strict ancestor of the result commit
(`docs/standing-constraints.md`, Guard 5). The two entries below this one could not claim that
and said so. This one can, twice: `PREREGISTRATION.md` was committed as `0f80959`
before the first `compare.mjs` run, and `PREREGISTRATION-2.md` as `680b9d3` before the re-run
against both PRs' post-review heads. Each names both arms, the score-capable changes, the rows
expected to move and the exact metric values expected. Each measured result matched its
prediction in full; the second is byte-identical to the first on all 24 rows once `toolVersion`
is excluded, so the review fixes changed no figure here. What the first prediction named is what
both runs measured: the same four rows, the same two metrics, the same single headline change,
the same 20 byte-identical reports. No gap between prediction and result had to be explained
away, and none was closed by adjusting anything. Preregistration 2 also withdrew an expectation
before its run rather than after: an earlier, looser approximation of #308's widened
error-middleware pattern matched one file in vite, and the pattern as committed matches none, so
the prediction of a vite movement was retracted in the preregistration commit itself.

**Expected values, committed before the comparison ran** (`PREREGISTRATION.md`, `0f80959`; the
re-run's are in `PREREGISTRATION-2.md`, `680b9d3`).
For change 1, the rows expected to move are those whose `test` script contains no runner the
content check names: from the corpus `package.json` files, django (`grunt test --verbose`), vite
(`pnpm test-unit && pnpm test-serve && pnpm test-build`) and the private alfred row — **3 rows,
all down**. Every other JavaScript row names a runner (`vitest`, `mocha`, `jest-cli.js`,
`npm run test:vitest`) and was expected not to move. For change 2, the only row on A1's
authenticated-absence path is carddemo, which has no `package.json` — **0 rows**. For change 3,
no corpus checkout contains a `.github/PULL_REQUEST_TEMPLATE/` directory — **0 rows**. For
change 4, A3 is `not_applicable` on 17 of 24 rows, leaving seven that can move on any A3 signal
at all; of those, the widened pattern selects a file the old one did not on react and alfred
only — **2 rows, both up**. vue's single candidate file matches `requestIdleCallback`, which the
shipped word-bounded `\b(?:correlation|request|trace)[-_]?id\b` correctly rejects; axios's three
candidates already matched the existing terms and two are test files the implementation-file
filter excludes; vite, scorecard and cejel contain none; svelte contains two matching files but
its A3 is `not_applicable`. For change 5, no file in any of the seven A3-applicable rows matches
the four-argument `(err, req, res, next)` pattern — **0 rows**. An empty result for changes 2, 3
and 5 is therefore the expected value, checked against the checkouts, not a convenient absence.

**Result.** 24 of 24 rows completed in both arms. **0 reports byte-identical — see the note below, which is not a scoring result. 4 rows moved,
exactly the 4 predicted, on exactly the 2 predicted metrics: `B3.ci_script_depth` down on
django, vite and alfred, and `A3.observability_depth` up on react (68 → 108) and alfred
(64 → 73).** One headline changed: django 3.2 → 3.1 overall (process 3.8 → 3.6, B3 3.6 → 3.1),
verdict and placement unchanged (unranked, low-confidence coverage). vite and alfred lost one
`ci_script_depth` point each with no score change (the metric saturates), and neither
observability movement changed a criterion score or status (react's A3 stays 2.3/warning,
alfred's 3.6/verified). No verdict changed. No placement changed. No coverage figure changed.
Changes 2, 3 and 5 moved nothing on this corpus, which is what the checkouts predicted — not
evidence that they move nothing in general: the guard fixtures shipped with each change show
them moving on the shapes they were written for. The four pinned shapes in
`src/witan/__tests__/v17-scoring-surface-golden.test.ts` stay 4/4 green on the candidate, as
predicted, and the two guards shipped with changes 4 and 5 pass on it (20 assertions across the
three files).

**Finding: the content check over-reaches, disclosed rather than adjusted.** None of the three
rows that moved on `ci_script_depth` carries the placeholder that fix was written for. All three
run a real test entrypoint that delegates to a runner the check's allowlist does not name. The
fix converted "credits a placeholder" into "does not credit a delegating script" — a new false
assertion of the absence of a test capability, in the direction that lowers a score. Per the
constraint that governs this entry, nothing was adjusted to shrink the delta: the check ships as
written, the three rows move, and this paragraph says why. **Operator decision:** whether the
check should recognise delegation (a `test` script that is not the npm placeholder, or one that
invokes `npm run`/`pnpm`/`turbo`/`grunt`/`make`), which would need its own entry here and a
re-pin of the guard; and separately whether a behaviour change of this kind under a fixed
identifier should have forced a rubric version bump. Neither is decided by this entry.

**Why no report is byte-identical, and why that is not a scoring statement.** Every one of the
24 reports differs from its 0.4.8 counterpart at byte level, because 0.4.9 adds
`rubricBehaviourFingerprint` beside `rubricVersion` and moves the report format from 1.1 to 1.2.
All 24 carry the new field. An earlier draft of this entry recorded 20 of 24 byte-identical,
measured before that field existed; the number was true of the tree measured and false of the
tree that ships, so it is corrected here rather than left to read as a scoring claim.

The scoring result is unchanged by it: the same four rows move, on the same two metrics, by the
same amounts, and the baseline still reproduces the published board on 20 of 24 rows. That split
is the point — the metric comparison is the scoring assertion and the byte comparison is a
format assertion, and a format change moves only the second. From 0.4.9 the fingerprint is the
better instrument for the question the byte count was standing in for: whether two reports were
produced by the same scoring, rather than by the same bytes.

**Note on the two upward metric movements.** `observability_depth` is a raw count of matching
files, so react's 68 → 108 and alfred's 64 → 73 are the same repositories at the same commits
counted by a wider pattern, not repositories that became more observable. The count is not
comparable across releases as a quantity, and neither movement crossed a score band. A board
republished on 0.4.9 will show the new counts on those two rows with unchanged scores.

**Prior undisclosed movement found while measuring.** Checking the baseline arm against the
board published on cejel.dev at scoring level (headline, per-criterion score/status, per-metric
value) showed 20 of 24 rows reproduce and the board predates 0.4.8 on four: fastapi A4
`lockfile_coverage` 1 → abstained, A4 3.6 → 3.4, headline overall 3.1 → 3.0 and code trust
3.0 → 2.8; biomejs B4 `audit_artifact_depth` 16 → 15 and `audit_freshness_depth` 2 → 1;
fmt A1 `non_hollow_test_share` 29 → 28 and `test_to_source_ratio` 55 → 54; alfred A3
`rollback_safety_depth` 806 → 753 — the last three at metric level with no score change. All
are consistent with 0.4.8's abstention-scoring change, which shipped with a CHANGELOG line and
no entry here. So the published board already disagrees with the shipped 0.4.8 binary on one
headline score, and this entry is the second undisclosed v17 movement in two releases, not the
first. Republishing the board is release execution and is not done here. (The first version
of this check compared whole report objects and reported a difference on all 24 rows regardless
of input, because 0.4.8 added a `derivation` field to every finding; the four rows above were
found by hand and the check was then rewritten to compare at scoring level, which reproduces
exactly those four — review finding on #306.)


**Measurement limits.** Public rows were fetched with `--depth=1`; both arms score the same
shallow checkouts, so the base → candidate delta is unaffected, but history-dependent signals
(A2's recent-history secret scan, B4's commit-year freshness) saw one commit of history, as in
the v19 protocol. `harness/RUN.md` records the exact invocation and a verification that the
committed harness reproduces the candidate arm from its committed location.


**Full v0.4.8 → 0.4.9 delta under witan-rubric-v17-2026-07-24 (all 24 rows, candidate = main + both remaining 0.4.9 scoring branches):**

| Repository | Overall | Code trust | Process trust | Verdict | Coverage | Board placement | Criteria that moved (score/status) | Metrics that moved |
|---|---:|---:|---:|---|---|---|---|---|
| react | 3 | 2.1 | 3.9 | Conditional | code_trust 5/5; process_trust 3/6 | 9 | none (report differs elsewhere) | A3.observability_depth 68 to 108 |
| vue | 2.9 | 2.4 | 3.4 | Conditional | code_trust 4/5; process_trust 3/6 | 11 | identical | none |
| svelte | 3.1 | 2.9 | 3.3 | Conditional | code_trust 4/5; process_trust 3/6 | 4 | identical | none |
| django | 3.2 to 3.1 | 2.6 | 3.8 to 3.6 | Conditional | code_trust 3/5; process_trust 2/6 | unranked | B3 3.6/verified to 3.1/verified | B3.ci_script_depth 3 to 2 |
| flask | 2.9 | 2.7 | 3 | Conditional | code_trust 4/5; process_trust 3/6 | 8 | identical | none |
| fastapi | 3 | 2.8 | 3.2 | Conditional | code_trust 2/5; process_trust 3/6 | unranked | identical | none |
| express | 3 | 2.8 | 3.2 | Conditional | code_trust 2/5; process_trust 3/6 | unranked | identical | none |
| vite | 3.4 | 2.8 | 4 | Conditional | code_trust 5/5; process_trust 3/6 | 1 | none (report differs elsewhere) | B3.ci_script_depth 5 to 4 |
| esbuild | 2.5 | 2.6 | 2.4 | Conditional | code_trust 3/5; process_trust 3/6 | 13 | identical | none |
| biomejs | 3 | 2.9 | 3 | Conditional | code_trust 3/5; process_trust 4/6 | 6 | identical | none |
| requests | 2.9 | 2.4 | 3.4 | Conditional | code_trust 3/5; process_trust 4/6 | 7 | identical | none |
| pydantic | 3.2 | 2.9 | 3.5 | Conditional | code_trust 3/5; process_trust 3/6 | 3 | identical | none |
| axios | 3.3 | 2.6 | 3.9 | Conditional | code_trust 5/5; process_trust 4/6 | 2 | identical | none |
| zod | 3.2 | 3.1 | 3.2 | Conditional | code_trust 3/5; process_trust 3/6 | 5 | identical | none |
| scorecard | 2.9 | 2.2 | 3.6 | Conditional | code_trust 4/5; process_trust 3/6 | 10 | identical | none |
| ripgrep | 2.1 | 2.1 | 2 | At risk | code_trust 3/5; process_trust 3/6 | 14 | identical | none |
| guava | 1.9 | 1.6 | 2.2 | At risk | code_trust 3/5; process_trust 2/6 | unranked | identical | none |
| cobra | 2.5 | 2.6 | 2.3 | Conditional | code_trust 2/5; process_trust 2/6 | unranked | identical | none |
| sinatra | 2.4 | 2 | 2.8 | At risk | code_trust 2/5; process_trust 4/6 | unranked | identical | none |
| automapper | 2.2 | 2 | 2.3 | At risk | code_trust 3/5; process_trust 2/6 | unranked | identical | none |
| fmt | 2.6 | 2 | 3.2 | Conditional | code_trust 3/5; process_trust 4/6 | 12 | identical | none |
| carddemo | scoreless | scoreless | scoreless | Insufficient source | code_trust 0/5; process_trust 0/6 | unrated | identical | none |
| alfred | 3.2 | 3.1 | 3.3 | Conditional | code_trust 5/5; process_trust 4/6 | transparency | none (report differs elsewhere) | A3.observability_depth 64 to 73; B3.ci_script_depth 5 to 4 |
| cejel | 2.8 | 2.3 | 3.2 | Conditional | code_trust 5/5; process_trust 3/6 | transparency | identical | none |

## witan-rubric-v19-prospective-2026-08-09 — recovery GO

**Status.** Prospective only; recovery protocol **GO**. V19 is available only by explicit
selection. The calibrated public default remains `witan-rubric-v17-2026-07-24`; this result does
not authorize default promotion or rewrite the preserved first-run NO-GO below.

**What changed.** V19 inherits prospective v18 and changes only B4's numeric freshness marker:
the scanned `HEAD` commit's committer year replaces the scan wall clock. Static
`recent|latest|current` markers remain. V17 and v18 retain their historical generated-at-year
behavior.

**Result.** The separately preregistered recovery measured only the missing private Alfred row at
its frozen commit and combined it with the original run's 23 successful rows byte-for-byte. All
24 rows completed. No row changed raw B4 freshness, B4 score or status, headline score, verdict,
coverage, board placement, or any non-B4 criterion. The recovery preregistration merged at
`8f937a6d326d9aacc1c0fd607f79f815e148c165`; the frozen recovery harness merged later at
`7c498ef3c7ac4fefcc08b8f03d2be31bde3476eb`, which is the execution commit.

Canonical recovery evidence:
`docs/experiments/b4-commit-year-v19-recovery-2026-08-09/alfred-recovery-result.json`; canonical
combined evidence:
`docs/experiments/b4-commit-year-v19-recovery-2026-08-09/combined-result.json`; human renderings
are beside those files.

**Full prospective-v18 to prospective-v19 combined delta (all 24 rows):**

| Repository | Raw B4 freshness | B4 score/status | Overall | Code trust | Process trust | Verdict | Coverage | Board placement | Non-B4 criteria |
|---|---:|---|---:|---:|---:|---|---|---|---|
| react | 4 to 4 | 3.8/verified to 3.8/verified | 3.0 to 3.0 | 2.1 to 2.1 | 3.9 to 3.9 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | 9 to 9 | identical |
| vue | 2 to 2 | 2.9/verified to 2.9/verified | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 11 to 11 | identical |
| svelte | 1 to 1 | 1.9/verified to 1.9/verified | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 4 to 4 | identical |
| django | n/a to n/a | 0/not applicable to 0/not applicable | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| flask | 2 to 2 | 3.7/verified to 3.7/verified | 2.9 to 2.9 | 2.7 to 2.7 | 3.0 to 3.0 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 8 to 8 | identical |
| fastapi | 48 to 48 | 3.6/verified to 3.6/verified | 3.1 to 3.1 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | code 2/5; process 3/6 to code 2/5; process 3/6 | unranked to unranked | identical |
| express | 1 to 1 | 1.9/verified to 1.9/verified | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | Conditional to Conditional | code 2/5; process 3/6 to code 2/5; process 3/6 | unranked to unranked | identical |
| vite | 5 to 5 | 4.0/verified to 4.0/verified | 3.4 to 3.4 | 2.8 to 2.8 | 4.0 to 4.0 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | 1 to 1 | identical |
| esbuild | 1 to 1 | 1.9/verified to 1.9/verified | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 13 to 13 | identical |
| biomejs | 2 to 2 | 3.3/verified to 3.3/verified | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 6 to 6 | identical |
| requests | 2 to 2 | 2.9/verified to 2.9/verified | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 7 to 7 | identical |
| pydantic | 2 to 2 | 2.9/verified to 2.9/verified | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 3 to 3 | identical |
| axios | 3 to 3 | 3.6/verified to 3.6/verified | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | code 5/5; process 4/6 to code 5/5; process 4/6 | 2 to 2 | identical |
| zod | 2 to 2 | 2.9/verified to 2.9/verified | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 5 to 5 | identical |
| scorecard | n/a to n/a | 0/not applicable to 0/not applicable | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 10 to 10 | identical |
| ripgrep | 1 to 1 | 1.9/verified to 1.9/verified | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | code 3/5; process 3/6 to code 3/5; process 3/6 | 14 to 14 | identical |
| guava | n/a to n/a | 0/not applicable to 0/not applicable | 1.9 to 1.9 | 1.6 to 1.6 | 2.2 to 2.2 | At risk to At risk | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| cobra | n/a to n/a | 0/not applicable to 0/not applicable | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | code 2/5; process 2/6 to code 2/5; process 2/6 | unranked to unranked | identical |
| sinatra | 1 to 1 | 2.5/verified to 2.5/verified | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | code 2/5; process 4/6 to code 2/5; process 4/6 | unranked to unranked | identical |
| automapper | n/a to n/a | 0/not applicable to 0/not applicable | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| fmt | 1 to 1 | 2.5/verified to 2.5/verified | 2.6 to 2.6 | 2.0 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 12 to 12 | identical |
| carddemo | n/a to n/a | 0/not applicable to 0/not applicable | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | code 0/5; process 0/6 to code 0/5; process 0/6 | unrated to unrated | identical |
| alfred | 73 to 73 | 3.9/verified to 3.9/verified | 3.4 to 3.4 | 2.8 to 2.8 | 3.9 to 3.9 | Conditional to Conditional | code 5/5; process 4/6 to code 5/5; process 4/6 | transparency to transparency | identical |
| cejel | n/a to n/a | 0/not applicable to 0/not applicable | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | transparency to transparency | identical |

## witan-rubric-v19-prospective-2026-08-09 — first paired run NO-GO

**Status.** Prospective only; protocol **NO-GO**. The calibrated public default remains
`witan-rubric-v17-2026-07-24`. This failed run does not authorize default promotion, rewrite any
historical report, or change the published board.

**What changed.** V19 inherits prospective v18 and changes only B4's numeric freshness marker:
the scanned `HEAD` commit's committer year replaces the scan wall clock. Static
`recent|latest|current` markers remain. V17 and v18 retain their historical generated-at-year
behavior. The implementation and one-shot harness merged at
`ce6af76376264540a4d12494a8ac8d4ab92082ee`, strictly after preregistration commit
`9eefecbc1c7f83ec2ba795ea823a3edb43b12bf1`.

**Result.** The one permitted run completed 23 of 24 frozen rows. All 23 completed rows had
byte-identical non-B4 criteria and zero B4-score, status, headline-score, verdict, coverage, or
placement changes. No completed row changed the raw freshness numerator. The private Alfred row
failed before checkout or scanning: the harness correctly denied file transport globally, but
then attempted a local clone that required that same denied transport. The failure is preserved;
the run was not repaired or retried. A recovery measurement requires a separately merged
preregistration and a new one-shot protocol.

Canonical evidence:
`docs/experiments/b4-commit-year-v19-2026-08-09/paired-result.json`; human rendering:
`docs/experiments/b4-commit-year-v19-2026-08-09/result.md`.

**Full prospective-v18 to prospective-v19 first-run delta (all 24 rows):**

| Repository | Raw B4 freshness | B4 score/status | Overall | Code trust | Process trust | Verdict | Coverage | Board placement | Non-B4 criteria |
|---|---:|---|---:|---:|---:|---|---|---|---|
| react | 4 to 4 | 3.8/verified to 3.8/verified | 3.0 to 3.0 | 2.1 to 2.1 | 3.9 to 3.9 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | 9 to 9 | identical |
| vue | 2 to 2 | 2.9/verified to 2.9/verified | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 11 to 11 | identical |
| svelte | 1 to 1 | 1.9/verified to 1.9/verified | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 4 to 4 | identical |
| django | n/a to n/a | 0/not applicable to 0/not applicable | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| flask | 2 to 2 | 3.7/verified to 3.7/verified | 2.9 to 2.9 | 2.7 to 2.7 | 3.0 to 3.0 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 8 to 8 | identical |
| fastapi | 48 to 48 | 3.6/verified to 3.6/verified | 3.1 to 3.1 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | code 2/5; process 3/6 to code 2/5; process 3/6 | unranked to unranked | identical |
| express | 1 to 1 | 1.9/verified to 1.9/verified | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | Conditional to Conditional | code 2/5; process 3/6 to code 2/5; process 3/6 | unranked to unranked | identical |
| vite | 5 to 5 | 4.0/verified to 4.0/verified | 3.4 to 3.4 | 2.8 to 2.8 | 4.0 to 4.0 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | 1 to 1 | identical |
| esbuild | 1 to 1 | 1.9/verified to 1.9/verified | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 13 to 13 | identical |
| biomejs | 2 to 2 | 3.3/verified to 3.3/verified | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 6 to 6 | identical |
| requests | 2 to 2 | 2.9/verified to 2.9/verified | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 7 to 7 | identical |
| pydantic | 2 to 2 | 2.9/verified to 2.9/verified | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 3 to 3 | identical |
| axios | 3 to 3 | 3.6/verified to 3.6/verified | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | code 5/5; process 4/6 to code 5/5; process 4/6 | 2 to 2 | identical |
| zod | 2 to 2 | 2.9/verified to 2.9/verified | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | code 3/5; process 3/6 to code 3/5; process 3/6 | 5 to 5 | identical |
| scorecard | n/a to n/a | 0/not applicable to 0/not applicable | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | code 4/5; process 3/6 to code 4/5; process 3/6 | 10 to 10 | identical |
| ripgrep | 1 to 1 | 1.9/verified to 1.9/verified | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | code 3/5; process 3/6 to code 3/5; process 3/6 | 14 to 14 | identical |
| guava | n/a to n/a | 0/not applicable to 0/not applicable | 1.9 to 1.9 | 1.6 to 1.6 | 2.2 to 2.2 | At risk to At risk | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| cobra | n/a to n/a | 0/not applicable to 0/not applicable | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | code 2/5; process 2/6 to code 2/5; process 2/6 | unranked to unranked | identical |
| sinatra | 1 to 1 | 2.5/verified to 2.5/verified | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | code 2/5; process 4/6 to code 2/5; process 4/6 | unranked to unranked | identical |
| automapper | n/a to n/a | 0/not applicable to 0/not applicable | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | code 3/5; process 2/6 to code 3/5; process 2/6 | unranked to unranked | identical |
| fmt | 1 to 1 | 2.5/verified to 2.5/verified | 2.6 to 2.6 | 2.0 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | code 3/5; process 4/6 to code 3/5; process 4/6 | 12 to 12 | identical |
| carddemo | n/a to n/a | 0/not applicable to 0/not applicable | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | code 0/5; process 0/6 to code 0/5; process 0/6 | unrated to unrated | identical |
| alfred | error | error | error | error | error | error | error | error | not compared |
| cejel | n/a to n/a | 0/not applicable to 0/not applicable | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | code 5/5; process 3/6 to code 5/5; process 3/6 | transparency to transparency | identical |

## witan-rubric-v18-prospective-2026-07-25

**Status.** Prospective only. The public `npx @cejel/cejel@latest` default remains the
holdout-calibrated `witan-rubric-v17-2026-07-24`. V18 must be selected explicitly and is not
eligible to replace that default until a fresh authenticated untouched v51 holdout clears every
preregistered gate. The public leaderboard selects v18 explicitly, with verdict bands withheld,
so it can disclose the correction without representing the changed detector as calibrated.

**What changed.** V18 inherits the complete v17 detector and scoring closure except for D7
multi-tenancy evidence. Positive isolation credit is now derived from each repository's own
`CREATE POLICY ... USING (...)` and `WITH CHECK (...)` clauses. Identifiers parsed from those
clauses define the repository-native scope vocabulary and are traced across the data-layer
surface; a fixed tenant-token dictionary cannot create positive isolation credit. The
`rls_policy_count` denominator is derived from that same scoped surface rather than saturating at
a hand-typed three policies.

The negative rule remains fail-closed. A tenant-shaped schema with zero RLS policies still emits
`CORE-A2-TENANT-WITHOUT-RLS`; it is never silently relabeled single-tenant. The conservative
non-policy schema premise exists only to emit that gap finding and cannot award isolation metrics.
A regression fixture now exercises this exact zero-policy case under calibrated v17 and
prospective v18.

**Why.** The v0.2.0 board exposed two distinct protocol defects. Its headline used a 50/50 mean
of category means, allowing a thin four-criterion process bucket to receive half the headline;
the board now uses the pre-statable, uniform rule that thin buckets must not receive
disproportionate headline weight, implemented as equal weight per measured comparable criterion.
Separately, the attempted D7 correction initially changed v17 in place. That would have shipped a
different detector under the same identifier used by the v50 GO. Versioning D7 as prospective
v18 preserves the exact calibrated detector and its measured claim while making the correction
auditable. The finding-level v50 GO and `terminal-go.md` are unchanged.

All 24 pinned corpus rows were compared at their existing immutable commits. The only rubric-score
move is the private Alfred transparency snapshot; no third-party repository changes score,
verdict, coverage class, or board placement. Publisher-owned rows remain outside the calibrated
public population. Board verdict bands are withheld because the historical 3.5 cutoff has not
been calibrated for the comparative score.

**Full v17 to prospective-v18 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | Native verdict | Board placement |
|---|---:|---:|---:|---|---|
| react | 3.0 to 3.0 | 2.1 to 2.1 | 3.9 to 3.9 | Conditional to Conditional | 9 to 9 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 11 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | 4 to 4 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.9 to 2.9 | 2.7 to 2.7 | 3.0 to 3.0 | Conditional to Conditional | 8 to 8 |
| fastapi | 3.1 to 3.1 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | unranked to unranked |
| express | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | Conditional to Conditional | unranked to unranked |
| vite | 3.4 to 3.4 | 2.8 to 2.8 | 4.0 to 4.0 | Conditional to Conditional | 1 to 1 |
| esbuild | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | 13 to 13 |
| biomejs | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | 6 to 6 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 7 to 7 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | 3 to 3 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | 2 to 2 |
| zod | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | 5 to 5 |
| scorecard | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | 10 to 10 |
| ripgrep | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | 14 to 14 |
| guava | 1.9 to 1.9 | 1.6 to 1.6 | 2.2 to 2.2 | At risk to At risk | unranked to unranked |
| cobra | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | unranked to unranked |
| automapper | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | unranked to unranked |
| fmt | 2.6 to 2.6 | 2.0 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | 12 to 12 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | unrated to unrated |
| alfred | 3.5 to 3.3 | 3.0 to 2.6 | 3.9 to 3.9 | Verified to Conditional | transparency to transparency |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | transparency to transparency |

## witan-rubric-v17-2026-07-24

**What changed.** V17 is the first free-core rubric promoted after a preregistered untouched
holdout cleared every gate. It carries forward the v9 evidence boundary and adds the measured
v17 corrections: bounded reviewable-source proof for absence propositions, authenticated A1
configured-runner absence, full test-inventory discovery, and narrowly scoped structural rescue
only after a positive core signal. Blind abstention uses authored production source and the
configured-runner premise; it does not infer a healthy result from metadata alone. The public
default is promoted only after the v50 terminal GO, not from a development score or a
configured-runner premise without evidence. The commercial score remains bounded to observable
repository evidence; it is not a security guarantee.

**Why.** The v49 untouched holdout's only failure was inappropriate abstention (16 of 189,
upper envelope 12.70% against a 10% gate). V50 tested the v17 revision on a fresh 200-repository
cohort with three independent blind reviews and the frozen clustered estimator. It passed finding
precision, worst-case recall, FPR, criterion, missingness, and abstention gates. The measured
finding precision was 96.43% (95% lower bound 94.16%), worst-case recall 95.64% (lower bound
92.23%), and worst-case FPR 0.66% (upper bound 1.10%). The terminal claim's precision threshold
is 80%; this is a calibration result for the frozen free-core population, not universal recall.

All 24 pinned corpus rows were rescored at their existing immutable commits with zero scan or
score errors on 2026-07-25 UTC. CardDemo remains scoreless `insufficient_source`. The private
Alfred transparency row is labeled private and is not independently reproducible from public
source. The table below is the complete v9-to-v17 delta. The historical **Full v8 to v9 delta**
remains below; it is not silently replaced by this prospective rescore.

**Full v9 to v17 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | Verdict | Rank |
|---|---:|---:|---:|---|---:|
| react | 3.2 to 3.0 | 2.5 to 2.1 | 3.9 to 3.9 | Conditional to Conditional | 5 to 7 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 11 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | 6 to 5 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.9 to 2.9 | 2.7 to 2.7 | 3.0 to 3.0 | Conditional to Conditional | 13 to 13 |
| fastapi | 3.1 to 3.1 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | unranked to unranked |
| express | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | Conditional to Conditional | unranked to unranked |
| vite | 3.4 to 3.4 | 2.8 to 2.8 | 4.0 to 4.0 | Conditional to Conditional | 1 to 2 |
| esbuild | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | 15 to 15 |
| biomejs | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 9 to 9 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | 4 to 4 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | 3 to 3 |
| zod | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | 7 to 6 |
| scorecard | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | 10 to 10 |
| ripgrep | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | 16 to 16 |
| guava | 1.9 to 1.9 | 1.5 to 1.6 | 2.2 to 2.2 | At risk to At risk | unranked to unranked |
| cobra | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | unranked to unranked |
| automapper | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | unranked to unranked |
| fmt | 2.6 to 2.6 | 2.0 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | 14 to 14 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | 3.2 to 3.5 | 2.4 to 3.0 | 4.0 to 4.0 | Conditional to Verified | 2 to 1 |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | 12 to 12 |

## witan-rubric-v9-2026-07-22

**What changed.** V9 is the failure-derived remediation from the immutable free-core v32
NO-GO. A1 now requires the same configured-runner premise its no-coverage proposition claims,
and it never infers a missing-test finding from a coverage artifact alone. Scheduled-health
recognition covers common non-npm test commands, while claim packets carry a bounded excerpt for
each conjunct from the finding's own file. A2 excludes generated, vendored, test, fixture, and
example paths from current and historical production-secret claims; qualified environment
templates such as `.env.production.example` are templates; and non-secret environment-file
hygiene remains visible without forcing the whole criterion into warning.

Headline scoring now requires at least 80% of source-shaped files to belong to a published Cejel
source family when a competing unread language exists. Native web component/template formats and
non-source translations, certificates, snapshots, and data artifacts are classified explicitly so
the stricter boundary does not turn mature public projects into false abstentions. Blind abstention
packets sample authored production source across extension families and start excerpts at
substantive code rather than repeated license headers. Criterion states use evidence-derived
materiality floors for non-hollow tests, production readiness, PR outcomes, CI depth, audit depth,
dependency sanity, and explicit human/fail-closed privilege gates.

**Why.** The sealed v32 result passed recall but failed precision, FPR, criterion state-exactness,
and inappropriate-scoring gates. All 17 false positives for `CORE-A1-NO-COVERAGE-CONFIG` had zero
configured runner surfaces, while all 25 supported candidates had at least one. Five of six
committed-secret candidates were generated or fixture evidence. The criterion errors clustered at
repeatable materiality boundaries, and 16 repositories were scored when the blind packet did not
materially represent their source—principally because excerpts stopped in license headers or an
unread source family was substantial. These corrections are prospective; they do not reinterpret
v32 or claim calibration GO. V33 still requires fresh seeds, a fresh untouched cohort, blind
review, sealing, and the frozen GO estimator.

All 24 pinned corpus rows rescored with zero scan errors. No verdict band or rank changed.
CardDemo remains scoreless `insufficient_source`. React, Flask, Express, Vite, and Alfred gain small
score corrections from the evidence/state changes; every other headline score is unchanged.

**Full v8 to v9 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | Verdict | Rank |
|---|---:|---:|---:|---|---:|
| react | 3.1 to 3.2 | 2.3 to 2.5 | 3.9 to 3.9 | Conditional to Conditional | 5 to 5 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 11 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | 6 to 6 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.8 to 2.9 | 2.5 to 2.7 | 3.0 to 3.0 | Conditional to Conditional | 13 to 13 |
| fastapi | 3.1 to 3.1 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | unranked to unranked |
| express | 2.9 to 3.0 | 2.8 to 2.8 | 3.0 to 3.2 | Conditional to Conditional | unranked to unranked |
| vite | 3.3 to 3.4 | 2.6 to 2.8 | 4.0 to 4.0 | Conditional to Conditional | 1 to 1 |
| esbuild | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | 15 to 15 |
| biomejs | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 9 to 9 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | 4 to 4 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | 3 to 3 |
| zod | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | 7 to 7 |
| scorecard | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | 10 to 10 |
| ripgrep | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | 16 to 16 |
| guava | 1.9 to 1.9 | 1.5 to 1.5 | 2.2 to 2.2 | At risk to At risk | unranked to unranked |
| cobra | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | unranked to unranked |
| automapper | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | unranked to unranked |
| fmt | 2.6 to 2.6 | 2.0 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | 14 to 14 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | 3.2 to 3.2 | 2.4 to 2.4 | 3.9 to 4.0 | Conditional to Conditional | 2 to 2 |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | 12 to 12 |

## witan-rubric-v8-2026-07-21

**What changed.** V8 activates the failure-derived detector and evidence-boundary corrections
staged after the immutable free-core v26 NO-GO. A1 recognizes lean tests in nested workspaces while
excluding generated, vendored, fixture, and example manifests from the production toolchain. A2
reports current and historical credential evidence independently, including a current-tree hygiene
warning for a committed non-template `.env` without a confirmed secret. A3 evaluates each multi-stage
Dockerfile's effective final stage, recognizes bounded Rack services, and requires a runtime command
rather than a generic entrypoint name. A4 applies app-runtime expectations only to strongly evidenced
packaged Electron/Tauri software. A5 accepts only content-authenticated reconciliation artifacts and
rejects generated-source claims. The same authored-production boundary governs the V8 control and
applicability premises used by blind calibration packets.

**Why.** Free-core v26 failed because decisive controls and criterion labels showed that the prior
detectors conflated workspace metadata with production evidence, collapsed current and historical
credential states, treated container build stages as deployed services, under-modeled packaged apps,
and accepted generic reconciliation filenames without authenticating their content. V8 corrects
those observed failure families before any v27 cohort is selected or scanned. It does not reinterpret
v26, weaken any GO gate, or claim calibration GO.

All 24 pinned corpus rows rescored with zero scan errors. No verdict band changed, and CardDemo
remains scoreless `insufficient_source`. FastAPI gains corrected A1 evidence but becomes unranked
because A5 now abstains instead of scoring an unauthenticated claim artifact, leaving only five of
eleven dimensions measured. The remaining rank movement follows the score corrections, FastAPI
leaving the ranked table, and Alfred's refreshed private transparency snapshot. That Alfred
snapshot advances from `b608b99` to the main-reachable V8 activation commit `fecc4d3`; its source
remains explicitly non-public. The published row was corrected from the activation branch's
pre-squash source commit to that merge commit. The table below therefore reports V7 directly to
the current V8 board. Alfred's headline Overall (3.2), Code trust (2.4), Process trust (3.9), and
verdict are unchanged by the correction; its common-dimension ranked score rises from 3.1 to 3.2
and B2 rises from 3.5 to 3.7 as the recent-PR merge ratio moves from 5/12 to 8/12 at the reachable
merge snapshot.

**Full v7 to v8 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | Verdict | Rank |
|---|---:|---:|---:|---|---:|
| react | 3.2 to 3.1 | 2.4 to 2.3 | 3.9 to 3.9 | Conditional to Conditional | 4 to 5 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 12 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | 6 to 6 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.8 to 2.8 | 2.6 to 2.5 | 3.0 to 3.0 | Conditional to Conditional | 14 to 13 |
| fastapi | 2.9 to 3.1 | 2.6 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | 9 to unranked |
| express | 2.9 to 2.9 | 2.8 to 2.8 | 3.0 to 3.0 | Conditional to Conditional | unranked to unranked |
| vite | 3.3 to 3.3 | 2.6 to 2.6 | 4.0 to 4.0 | Conditional to Conditional | 2 to 1 |
| esbuild | 2.5 to 2.5 | 2.6 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | 16 to 15 |
| biomejs | 3.0 to 3.0 | 2.9 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 10 to 9 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | 3 to 4 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | 1 to 3 |
| zod | 3.2 to 3.2 | 3.1 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | 7 to 7 |
| scorecard | 2.9 to 2.9 | 2.2 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | 11 to 10 |
| ripgrep | 2.1 to 2.1 | 2.1 to 2.1 | 2.0 to 2.0 | At risk to At risk | 17 to 16 |
| guava | 1.9 to 1.9 | 1.5 to 1.5 | 2.2 to 2.2 | At risk to At risk | unranked to unranked |
| cobra | 2.5 to 2.5 | 2.6 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | unranked to unranked |
| automapper | 2.2 to 2.2 | 2.0 to 2.0 | 2.3 to 2.3 | At risk to At risk | unranked to unranked |
| fmt | 2.7 to 2.6 | 2.1 to 2.0 | 3.2 to 3.2 | Conditional to Conditional | 15 to 14 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | 3.1 to 3.2 | 2.3 to 2.4 | 3.9 to 3.9 | Conditional to Conditional | 5 to 2 |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | 13 to 12 |

## witan-rubric-v7-2026-07-21

**What changed.** V7 is the failure-derived free-core v26 implementation. Secret findings now
point to the exact matched line and expose only redacted value shape. Finding and control packets
for no-tests, no-coverage, no-CI/release-deploy, and no-lockfile propositions carry the same exact
tracked-inventory proof and applicability premises. Criterion packets publish unweighted raw
metric facts plus frozen label definitions without exposing scores, states, score thresholds, or
weights.

Repository evidence discovery also changes. Coverage files must contain coverage configuration;
Makefiles and CMake files must contain actual test targets before counting as runner configuration;
recognized source files under ecosystem-specific test directories count as tests. Nested deploy
targets count only outside tests, fixtures, examples, samples, and demos; a Dockerfile remains a
container-build signal, not release automation. B6 excludes test/fixture SQL rather than emitting a
logging-only finding, and A5 limits claims to root/product documentation or dedicated
claim-reconciliation artifacts. Cohesive Fortran, CUDA/HIP, and web-template/style trees are
recognized at explicitly unmodeled depth, while prospectively frozen size/ratio/remainder rules
abstain from generated/vendor-dominated and independent solution-catalog trees. CardDemo remains
scoreless and `insufficient_source`.

**Why.** Free-core v25 closed NO-GO. Its decisive error families and packet missingness showed that
reviewers could not verify absence propositions from positive excerpts, criterion labels lacked
the measured denominators, secret evidence could point at the wrong line, test and fixture paths
crossed production applicability boundaries, and structural source eligibility was incomplete.
These are detector and review-protocol corrections, not a retroactive v25 reinterpretation. Both
v25 waves are excluded from the future v26 cohort; v25 Wave 2 remains unreviewed and unlabeled.

All 24 sealed corpus rows rescored with zero scan errors. No verdict band changed. CardDemo remains
scoreless and unranked. Express becomes unranked because the corrected A5 scope removes an
unsupported measured dimension, leaving five of eleven dimensions measured; this is a coverage
classification change, not a score penalty. The other rank changes below follow the published
score/evidence corrections and Express leaving the ranked table. The Alfred transparency row also
advances from self snapshot `707b7b5` to the reachable v7 implementation snapshot `b608b99`.

**Full v6 to v7 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | Verdict | Rank |
|---|---:|---:|---:|---|---:|
| react | 3.2 to 3.2 | 2.5 to 2.4 | 3.9 to 3.9 | Conditional to Conditional | 4 to 4 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 11 to 12 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | Conditional to Conditional | 5 to 6 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.8 to 2.8 | 2.5 to 2.6 | 3.0 to 3.0 | Conditional to Conditional | 15 to 14 |
| fastapi | 2.9 to 2.9 | 2.5 to 2.6 | 3.2 to 3.2 | Conditional to Conditional | 14 to 9 |
| express | 2.8 to 2.9 | 2.6 to 2.8 | 3.0 to 3.0 | Conditional to Conditional | 13 to unranked |
| vite | 3.3 to 3.3 | 2.6 to 2.6 | 4.0 to 4.0 | Conditional to Conditional | 2 to 2 |
| esbuild | 2.6 to 2.5 | 2.7 to 2.6 | 2.4 to 2.4 | Conditional to Conditional | 17 to 16 |
| biomejs | 2.9 to 3.0 | 2.8 to 2.9 | 3.0 to 3.0 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | Conditional to Conditional | 9 to 10 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | Conditional to Conditional | 3 to 3 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | Conditional to Conditional | 1 to 1 |
| zod | 3.0 to 3.2 | 2.8 to 3.1 | 3.2 to 3.2 | Conditional to Conditional | 7 to 7 |
| scorecard | 3.0 to 2.9 | 2.3 to 2.2 | 3.6 to 3.6 | Conditional to Conditional | 10 to 11 |
| ripgrep | 2.2 to 2.1 | 2.4 to 2.1 | 2.0 to 2.0 | At risk to At risk | 18 to 17 |
| guava | 1.9 to 1.9 | 1.5 to 1.5 | 2.2 to 2.2 | At risk to At risk | unranked to unranked |
| cobra | 2.6 to 2.5 | 2.8 to 2.6 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | At risk to At risk | unranked to unranked |
| automapper | 1.9 to 2.2 | 1.5 to 2.0 | 2.3 to 2.3 | At risk to At risk | unranked to unranked |
| fmt | 2.7 to 2.7 | 2.2 to 2.1 | 3.2 to 3.2 | Conditional to Conditional | 16 to 15 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | 3.0 to 3.1 | 2.2 to 2.3 | 3.8 to 3.9 | Conditional to Conditional | 6 to 5 |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | Conditional to Conditional | 12 to 13 |

## witan-rubric-v6-2026-07-21

**What changed.** Cejel now recognizes AVA's root-level `test.js` and `test-*.js`
conventions as concrete JavaScript test files. Version 0.1.7 recognized AVA as a configured
runner but missed those filenames, producing a false missing-tests finding on
`sindresorhus/slugify` even though its root `test.js` was present.

**Why.** This was a detection gap in Cejel, not evidence about the repository. The correction
adds the convention to both A1 test-file discovery and the non-hollow-test classifier, with a
regression fixture that reproduces the real repository shape. The fix does not execute tests
or broaden the scanner beyond tracked source evidence.

No externally pinned corpus score, verdict, or rank moved. None of the 22 external pinned
entries uses the newly recognized filename shape as previously missing evidence, and the two
rubric-protocol golden fixtures are unchanged. The Alfred transparency row advanced from the
previous self snapshot to implementation commit `707b7b5`: unrelated intervening source changes
move its Overall **3.1 to 3.0**, Code trust **2.4 to 2.2**, and rank **5 to 6**. Svelte moves rank
**6 to 5** only because Alfred moves below it; Svelte's score and evidence are unchanged. The
public Cejel row remains pinned to the same public commit and is unchanged.

**Full v5 to v6 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | A1 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| react | 3.2 to 3.2 | 2.5 to 2.5 | 3.9 to 3.9 | 2.3 to 2.3 | Conditional to Conditional | 4 to 4 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | 2.5 to 2.5 | Conditional to Conditional | 11 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | 2.6 to 2.6 | Conditional to Conditional | 6 to 5 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | 2.8 to 2.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.8 to 2.8 | 2.5 to 2.5 | 3.0 to 3.0 | 2.3 to 2.3 | Conditional to Conditional | 15 to 15 |
| fastapi | 2.9 to 2.9 | 2.5 to 2.5 | 3.2 to 3.2 | 2.0 to 2.0 | Conditional to Conditional | 14 to 14 |
| express | 2.8 to 2.8 | 2.6 to 2.6 | 3.0 to 3.0 | 2.3 to 2.3 | Conditional to Conditional | 13 to 13 |
| vite | 3.3 to 3.3 | 2.6 to 2.6 | 4.0 to 4.0 | 2.5 to 2.5 | Conditional to Conditional | 2 to 2 |
| esbuild | 2.6 to 2.6 | 2.7 to 2.7 | 2.4 to 2.4 | 2.5 to 2.5 | Conditional to Conditional | 17 to 17 |
| biomejs | 2.9 to 2.9 | 2.8 to 2.8 | 3.0 to 3.0 | 2.7 to 2.7 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | 2.5 to 2.5 | Conditional to Conditional | 9 to 9 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | 2.8 to 2.8 | Conditional to Conditional | 3 to 3 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | 2.8 to 2.8 | Conditional to Conditional | 1 to 1 |
| zod | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | 2.5 to 2.5 | Conditional to Conditional | 7 to 7 |
| scorecard | 3.0 to 3.0 | 2.3 to 2.3 | 3.6 to 3.6 | 2.4 to 2.4 | Conditional to Conditional | 10 to 10 |
| ripgrep | 2.2 to 2.2 | 2.4 to 2.4 | 2.0 to 2.0 | 1.9 to 1.9 | At risk to At risk | 18 to 18 |
| guava | 1.9 to 1.9 | 1.5 to 1.5 | 2.2 to 2.2 | 1.3 to 1.3 | At risk to At risk | unranked to unranked |
| cobra | 2.6 to 2.6 | 2.8 to 2.8 | 2.3 to 2.3 | 2.3 to 2.3 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | 1.8 to 1.8 | At risk to At risk | unranked to unranked |
| automapper | 1.9 to 1.9 | 1.5 to 1.5 | 2.3 to 2.3 | 0.3 to 0.3 | At risk to At risk | unranked to unranked |
| fmt | 2.7 to 2.7 | 2.2 to 2.2 | 3.2 to 3.2 | 2.8 to 2.8 | Conditional to Conditional | 16 to 16 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | 0.0 to 0.0 | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | **3.1 to 3.0** | **2.4 to 2.2** | 3.8 to 3.8 | 2.3 to 2.3 | Conditional to Conditional | **5 to 6** |
| cejel | 2.8 to 2.8 | 2.3 to 2.3 | 3.2 to 3.2 | 2.0 to 2.0 | Conditional to Conditional | 12 to 12 |

## witan-rubric-v5-2026-07-18

**What changed.** Git-history evidence is now read only from commits reachable from the
checked-out `HEAD`. Previous versions used `--all`, so unrelated local branches and
remote-tracking refs could alter A2 evidence, measured coverage, and the score of the same
pinned source commit. V5 binds the history scan to the immutable revision being certified.

**Why.** The required public-path reproduction job caught the same pinned Alfred/Cejel
snapshots producing different evidence in a developer worktree and a clean CI clone. The
clone's extra refs exposed a historical `.env` path that the generation environment did not
see. That was ambient clone state, not evidence for a different source snapshot. A regression
test now adds credential history on a non-HEAD branch and proves it cannot affect A2.

All 21 externally pinned rows are unchanged. Alfred's score is also unchanged, although its
complete evidence set is now reproducible. The Cejel self row gains the HEAD-ancestor history
signal that clean-clone reproduction had already exposed: A2 moves **0.0 to 3.2**, Code trust
**2.5 to 2.7**, and Overall **3.3 to 3.4**. It remains low-coverage and unranked; no ranked
position, verdict, or external score moves. The self snapshot moves from `d6248edd47f6` to
`25627e00c6eb`, the clean v5 implementation commit; external source pins do not move.

**Full v4 to v5 delta (all 24 rows):**

| Repository | Overall | Code trust | Process trust | A2 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| react | 3.2 to 3.2 | 2.5 to 2.5 | 3.9 to 3.9 | 2.4 to 2.4 | Conditional to Conditional | 4 to 4 |
| vue | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | 0.0 to 0.0 | Conditional to Conditional | 11 to 11 |
| svelte | 3.1 to 3.1 | 2.9 to 2.9 | 3.3 to 3.3 | 3.6 to 3.6 | Conditional to Conditional | 5 to 5 |
| django | 3.2 to 3.2 | 2.6 to 2.6 | 3.8 to 3.8 | 2.8 to 2.8 | Conditional to Conditional | unranked to unranked |
| flask | 2.8 to 2.8 | 2.5 to 2.5 | 3.0 to 3.0 | 2.4 to 2.4 | Conditional to Conditional | 14 to 14 |
| fastapi | 2.9 to 2.9 | 2.5 to 2.5 | 3.2 to 3.2 | 0.0 to 0.0 | Conditional to Conditional | 13 to 13 |
| express | 2.8 to 2.8 | 2.6 to 2.6 | 3.0 to 3.0 | 0.0 to 0.0 | Conditional to Conditional | 12 to 12 |
| vite | 3.3 to 3.3 | 2.6 to 2.6 | 4.0 to 4.0 | 2.4 to 2.4 | Conditional to Conditional | 2 to 2 |
| esbuild | 2.6 to 2.6 | 2.7 to 2.7 | 2.4 to 2.4 | 0.0 to 0.0 | Conditional to Conditional | 16 to 16 |
| biomejs | 2.9 to 2.9 | 2.8 to 2.8 | 3.0 to 3.0 | 3.2 to 3.2 | Conditional to Conditional | 8 to 8 |
| requests | 2.9 to 2.9 | 2.4 to 2.4 | 3.4 to 3.4 | 0.0 to 0.0 | Conditional to Conditional | 9 to 9 |
| pydantic | 3.2 to 3.2 | 2.9 to 2.9 | 3.5 to 3.5 | 0.0 to 0.0 | Conditional to Conditional | 3 to 3 |
| axios | 3.3 to 3.3 | 2.6 to 2.6 | 3.9 to 3.9 | 3.6 to 3.6 | Conditional to Conditional | 1 to 1 |
| zod | 3.0 to 3.0 | 2.8 to 2.8 | 3.2 to 3.2 | 3.6 to 3.6 | Conditional to Conditional | 7 to 7 |
| scorecard | 3.0 to 3.0 | 2.3 to 2.3 | 3.6 to 3.6 | 0.0 to 0.0 | Conditional to Conditional | 10 to 10 |
| ripgrep | 2.2 to 2.2 | 2.4 to 2.4 | 2.0 to 2.0 | 0.0 to 0.0 | At risk to At risk | 17 to 17 |
| guava | 1.9 to 1.9 | 1.5 to 1.5 | 2.2 to 2.2 | 0.0 to 0.0 | At risk to At risk | unranked to unranked |
| cobra | 2.6 to 2.6 | 2.8 to 2.8 | 2.3 to 2.3 | 0.0 to 0.0 | Conditional to Conditional | unranked to unranked |
| sinatra | 2.4 to 2.4 | 2.0 to 2.0 | 2.8 to 2.8 | 0.0 to 0.0 | At risk to At risk | unranked to unranked |
| automapper | 1.9 to 1.9 | 1.5 to 1.5 | 2.3 to 2.3 | 0.0 to 0.0 | At risk to At risk | unranked to unranked |
| fmt | 2.7 to 2.7 | 2.2 to 2.2 | 3.2 to 3.2 | 0.0 to 0.0 | Conditional to Conditional | 15 to 15 |
| carddemo | scoreless to scoreless | scoreless to scoreless | scoreless to scoreless | 0.0 to 0.0 | Insufficient source to Insufficient source | insufficient to insufficient |
| alfred | 3.0 to 3.0 | 2.4 to 2.4 | 3.6 to 3.6 | 2.4 to 2.4 | Conditional to Conditional | 6 to 6 |
| cejel | **3.3 to 3.4** | **2.5 to 2.7** | 4.0 to 4.0 | **0.0 to 3.2** | Conditional to Conditional | unranked to unranked |

No external score, verdict, evidence-derived rank, or source pin moved. No ranked row moved.

## witan-rubric-v4-2026-07-18

**What changed.** Repository evidence discovery now accepts only tracked regular files. It
does not follow tracked symlinks, because a symlink target can escape the immutable checkout
and make both evidence and scores depend on ambient host files. The scorer still records the
symlink in Git history, but never reads its target as repository evidence.

**Why.** The pinned Zod snapshot has a root `README.md` symlink. V3 followed that link while
also discovering the target `packages/docs/README.md`, which counted one claim source twice
and credited depth that did not exist. V4 reads only the regular target file. Zod A5 falls
from **2.2 to 2.0**; its rounded Code trust, Overall, verdict, and rank do not move. No other
externally pinned repository changes score, verdict, or rank.

The same regeneration moved the two self rows from source snapshot `75fa69511494` to the
clean v4 prerequisite commit recorded in their reports. Those are disclosed as a separate
corpus-pin act, not attributed to the symlink rubric change: Alfred falls **3.1 to 3.0** and
rank 5 to 6; Cejel falls **3.4 to 3.3** and remains unranked for low coverage. Svelte moves
rank 6 to 5 only because Alfred moves below it. CardDemo remains scoreless with
`insufficient_source`.

**Full v3→v4 delta (all 24 rows; external source pins unchanged):**

| Repository | Overall | Code trust | Process trust | A5 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| react | 3.2 → 3.2 | 2.5 → 2.5 | 3.9 → 3.9 | 2.2 → 2.2 | Conditional → Conditional | 4 → 4 |
| vue | 2.9 → 2.9 | 2.4 → 2.4 | 3.4 → 3.4 | 2.2 → 2.2 | Conditional → Conditional | 11 → 11 |
| svelte | 3.1 → 3.1 | 2.9 → 2.9 | 3.3 → 3.3 | 2.2 → 2.2 | Conditional → Conditional | 6 → 5 |
| django | 3.2 → 3.2 | 2.6 → 2.6 | 3.8 → 3.8 | 0.0 → 0.0 | Conditional → Conditional | unranked → unranked |
| flask | 2.8 → 2.8 | 2.5 → 2.5 | 3.0 → 3.0 | 2.2 → 2.2 | Conditional → Conditional | 14 → 14 |
| fastapi | 2.9 → 2.9 | 2.5 → 2.5 | 3.2 → 3.2 | 2.0 → 2.0 | Conditional → Conditional | 13 → 13 |
| express | 2.8 → 2.8 | 2.6 → 2.6 | 3.0 → 3.0 | 2.0 → 2.0 | Conditional → Conditional | 12 → 12 |
| vite | 3.3 → 3.3 | 2.6 → 2.6 | 4.0 → 4.0 | 2.7 → 2.7 | Conditional → Conditional | 2 → 2 |
| esbuild | 2.6 → 2.6 | 2.7 → 2.7 | 2.4 → 2.4 | 2.4 → 2.4 | Conditional → Conditional | 16 → 16 |
| biomejs | 2.9 → 2.9 | 2.8 → 2.8 | 3.0 → 3.0 | 2.0 → 2.0 | Conditional → Conditional | 8 → 8 |
| requests | 2.9 → 2.9 | 2.4 → 2.4 | 3.4 → 3.4 | 2.2 → 2.2 | Conditional → Conditional | 9 → 9 |
| pydantic | 3.2 → 3.2 | 2.9 → 2.9 | 3.5 → 3.5 | 2.4 → 2.4 | Conditional → Conditional | 3 → 3 |
| axios | 3.3 → 3.3 | 2.6 → 2.6 | 3.9 → 3.9 | 2.4 → 2.4 | Conditional → Conditional | 1 → 1 |
| zod | 3.0 → 3.0 | 2.8 → 2.8 | 3.2 → 3.2 | **2.2 → 2.0** | Conditional → Conditional | 7 → 7 |
| scorecard | 3.0 → 3.0 | 2.3 → 2.3 | 3.6 → 3.6 | 2.6 → 2.6 | Conditional → Conditional | 10 → 10 |
| ripgrep | 2.2 → 2.2 | 2.4 → 2.4 | 2.0 → 2.0 | 2.2 → 2.2 | At risk → At risk | 17 → 17 |
| guava | 1.9 → 1.9 | 1.5 → 1.5 | 2.2 → 2.2 | 2.2 → 2.2 | At risk → At risk | unranked → unranked |
| cobra | 2.6 → 2.6 | 2.8 → 2.8 | 2.3 → 2.3 | 0.0 → 0.0 | Conditional → Conditional | unranked → unranked |
| sinatra | 2.4 → 2.4 | 2.0 → 2.0 | 2.8 → 2.8 | 2.2 → 2.2 | At risk → At risk | unranked → unranked |
| automapper | 1.9 → 1.9 | 1.5 → 1.5 | 2.3 → 2.3 | 1.4 → 1.4 | At risk → At risk | unranked → unranked |
| fmt | 2.7 → 2.7 | 2.2 → 2.2 | 3.2 → 3.2 | 2.2 → 2.2 | Conditional → Conditional | 15 → 15 |
| carddemo | scoreless → scoreless | scoreless → scoreless | scoreless → scoreless | 0.0 → 0.0 | Insufficient source → Insufficient source | insufficient → insufficient |
| alfred | **3.1 → 3.0** | 2.4 → 2.4 | **3.8 → 3.6** | 2.4 → 2.4 | Conditional → Conditional | **5 → 6** |
| cejel | **3.4 → 3.3** | **2.7 → 2.5** | 4.0 → 4.0 | 2.2 → 2.2 | Conditional → Conditional | unranked → unranked |

No external source pin moved. Apart from Zod's A5 correction, no externally pinned score,
verdict, or rank moved. The self-source movement and its induced Svelte rank change are
identified separately above.

## witan-rubric-v3-2026-07-13

**What changed.** The leaderboard no longer has a privileged scoring route for internal
rows. The public CLI, batch scanner, and board generator now call one sealed public scorer;
the required board guard re-scores **every** corpus row through that function and compares
score, criterion status, verdict, measured coverage, and the complete evidence-pointer set
with the published report. This is a structural equivalence check, not a blacklist of known
collector names: adding any differently named board-only input changes the claims and makes
the required check RED.

Every external corpus entry now pins a 40-character source commit. Re-scoring at the same
pins is a rubric change and publishes the delta below. Moving a pin is a separate corpus
change that must be declared as such; upstream default-branch movement can no longer masquerade
as a rubric effect. Alfred and Cejel are likewise reproduced from the exact local commit
recorded in their published reports.

The repository scanner also states its B1/B5 behavior precisely: it evaluates neither
dimension for any repository, including Alfred. Both are always `not_applicable` for this
input type. They remain part of the wider rubric for
structured substrate evidence, and the board excludes them fail-closed if it encounters a
legacy or separately produced structured report.

**Why.** The previous board generator could append private, internal-only collectors after
the public scan. A collector-name blacklist asserted that known private inputs were absent,
but could not prove the published row was obtainable from the public product. Removing that
second path exposes the honest Alfred result: Code trust falls from **2.5 to 2.4**, driven by
A5 falling from **2.9 to 2.4**. The headline remains 3.1 because process evidence moves from
3.6 to 3.7 when B1/B5 stop contributing internal-only scores. The lower Code number is the
point of this release, not an artifact to explain away.

**Corpus-wide v2→v3 delta (all 17 repositories, same pinned source snapshots):**

| Repository | Overall | Code trust | Process trust | A5 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| react | 3.2 → 3.2 | 2.5 → 2.5 | 3.9 → 3.9 | 2.2 → 2.2 | Conditional → Conditional | 4 → 4 |
| vue | 2.9 → 2.9 | 2.4 → 2.4 | 3.4 → 3.4 | 2.2 → 2.2 | Conditional → Conditional | 11 → 11 |
| svelte | 3.1 → 3.1 | 2.9 → 2.9 | 3.3 → 3.3 | 2.2 → 2.2 | Conditional → Conditional | 6 → 6 |
| django | 3.2 → 3.2 | 2.6 → 2.6 | 3.8 → 3.8 | 0.0 → 0.0 | Conditional → Conditional | unranked → unranked |
| flask | 2.8 → 2.8 | 2.5 → 2.5 | 3.0 → 3.0 | 2.2 → 2.2 | Conditional → Conditional | 14 → 14 |
| fastapi | 2.9 → 2.9 | 2.5 → 2.5 | 3.2 → 3.2 | 2.0 → 2.0 | Conditional → Conditional | 13 → 13 |
| express | 2.8 → 2.8 | 2.6 → 2.6 | 3.0 → 3.0 | 2.0 → 2.0 | Conditional → Conditional | 12 → 12 |
| vite | 3.3 → 3.3 | 2.6 → 2.6 | 4.0 → 4.0 | 2.7 → 2.7 | Conditional → Conditional | 2 → 2 |
| esbuild | 2.6 → 2.6 | 2.7 → 2.7 | 2.4 → 2.4 | 2.4 → 2.4 | Conditional → Conditional | 15 → 15 |
| biomejs | 2.9 → 2.9 | 2.8 → 2.8 | 3.0 → 3.0 | 2.0 → 2.0 | Conditional → Conditional | 8 → 8 |
| requests | 2.9 → 2.9 | 2.4 → 2.4 | 3.4 → 3.4 | 2.2 → 2.2 | Conditional → Conditional | 9 → 9 |
| pydantic | 3.2 → 3.2 | 2.9 → 2.9 | 3.5 → 3.5 | 2.4 → 2.4 | Conditional → Conditional | 3 → 3 |
| axios | 3.3 → 3.3 | 2.6 → 2.6 | 3.9 → 3.9 | 2.4 → 2.4 | Conditional → Conditional | 1 → 1 |
| zod | 3.0 → 3.0 | 2.8 → 2.8 | 3.2 → 3.2 | 2.2 → 2.2 | Conditional → Conditional | 7 → 7 |
| scorecard | 3.0 → 3.0 | 2.3 → 2.3 | 3.6 → 3.6 | 2.6 → 2.6 | Conditional → Conditional | 10 → 10 |
| alfred (private) | 3.1 → 3.1 | **2.5 → 2.4** | 3.6 → 3.7 | **2.9 → 2.4** | Conditional → Conditional | 5 → 5 |
| cejel (private) | 3.4 → 3.4 | 2.7 → 2.7 | 4.0 → 4.0 | 2.2 → 2.2 | Conditional → Conditional | unranked → unranked |

**No external repository's score, verdict, evidence-derived rank, or A5 result moved.**
The only score changes are Alfred's disclosed correction above. The corpus pins are the
source commits already recorded by the v2 reports, so this table isolates the rubric/public-
path change from upstream repository movement.

**Separate corpus act composed in the same release (17 → 22 repositories).** The language-
calibration work adds one pinned, healthy repository in each of five ecosystems. These rows
did not exist in the v2 corpus, so `new` is the honest before-state; their low results expose
documented B3 CI-depth and Maven A4 calibration gaps rather than being hidden or normalized.

| Repository | Overall | Code trust | Process trust | A5 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| ripgrep | new → 2.2 | new → 2.4 | new → 2.0 | new → 2.2 | new → At risk | new → 16 |
| guava | new → 1.9 | new → 1.5 | new → 2.2 | new → 2.2 | new → At risk | new → unranked |
| cobra | new → 2.6 | new → 2.8 | new → 2.3 | new → 0.0 | new → Conditional | new → unranked |
| sinatra | new → 2.4 | new → 2.0 | new → 2.8 | new → 2.2 | new → At risk | new → unranked |
| automapper | new → 1.9 | new → 1.5 | new → 2.3 | new → 1.4 | new → At risk | new → unranked |

**Further corpus-only addition (22 → 23 repositories, 2026-07-15, #488).** One dedicated
C++ reference row, the same "widen ecosystem coverage" act as the five rows above, composed
later in the same v3 release rather than in the original batch. No rubric behavior changed;
`new` is the honest before-state for this row alone.

| Repository | Overall | Code trust | Process trust | A5 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| fmt | new → 2.7 | new → 2.2 | new → 3.2 | new → 2.2 | new → Conditional | new → 15 |

**No repository's score, verdict, or A5 result moved. Two ranks did**, mechanically and only
because a new row sorted above them: **esbuild 15 → 16, ripgrep 16 → 17.** A rank is a position
in a list, not a property of a repository — adding a row necessarily moves the rows beneath it.
Nothing was re-scored to produce that shift.

**Archetype classification: dominance, not presence (2026-07-15).** A behavior change, published
here because it changes which repositories Cejel is willing to score at all. `classifyRepoArchetype`
previously short-circuited to `source` when **any** recognised-extension file existed
(`sourceFileCount > 0`). One recognised file in a thousand was enough. The rule is now a ratio:
recognised source must reach `SOURCE_DOMINANCE_RATIO_THRESHOLD` (0.2) of tracked files, or the
repository is `unrecognised_ecosystem` and gets no score, no rank, and no verdict.

**Why.** The previous rule was correct at the ratios it was tested at and catastrophic below them.
A 99%-COBOL mainframe repository carrying nine incidental `.sh` deploy scripts — 2.7% of its
tracked files, 3.6% of its source-shaped ones — classified as `source`, was scored on those nine
files, and drew a confident **0.0 / Unverified** on a healthy codebase. Every real legacy
repository has a deploy script, so the abstention path was unreachable for essentially all of them
while every test passed. The threshold was calibrated against a ratio golden set — eight cases
spanning the ratio space, each with an expected outcome recorded and committed **before** the
threshold number was chosen — and the choice is insensitive within a wide band: 15%, 20% or 25%
all abstain this repository and all preserve a 50/50 mixed-language tree as `source`. The number
was not reverse-engineered from the repository that exposed the bug.

**No existing corpus row changed archetype, score, verdict, or rank.** Every scored repository's
recognised source is dominant by a wide margin; the 0.2 threshold is nowhere near any of them. The
change is visible only in what Cejel now declines to score.

**Corpus addition composed with it (23 → 24 repositories).** The first repository on this board
that Cejel refuses to rate. It is published, not omitted.

| Repository | Overall | Code trust | Process trust | A5 | Verdict | Rank |
|---|---:|---:|---:|---:|---|---:|
| carddemo | new → unrated | new → unrated | new → unrated | new → unrated | new → Insufficient source | new → unranked |

`unrated` is the honest before/after state: there is no score to report, and a `0.0` here would be
a claim about AWS's COBOL that we have not earned. The row's published reason names the extensions
found (`.cpy`, `.jcl`, `.cbl`, `.bms`, `.ps`, `.ctl`) and the measured ratio (9 of 329 tracked
files, 2.7%) rather than a bare verdict.

## Board presentation correction — 2026-07-12

No rubric version changed and no repository was re-scored. The board's public
"Overall" column now prints the same own-certificate overall score shown in each
repository's certificate/report, instead of a comparative-ranking-only value with the
same label. Four repositories' displayed headline numbers moved by 0.1; the underlying
reports, verdicts, and ranking basis did not change.

## witan-rubric-v2-2026-07-12

**What changed.** A1's scheduled-product-health-workflow sub-signal — previously a literal
check for the filename of this repository's own internal QA workflow (a private-repo path,
withheld here per the redaction policy below) — is now detected by **shape**: a CI workflow
with a `schedule:`/`cron:` trigger that runs the verification suite. Our own internal nightly
QA workflow is one recognized instance of that shape, not its definition; a differently-named
nightly workflow with the same shape is detected identically, and a workflow merely sharing
the same filename without the shape (no schedule trigger, no test-run command) is no longer
flagged at all.

A workflow matching the shape is then classified on whether its results are durably
published (a public pages deploy, a commit back to the repo, a PR/issue comment) or handed
only to an ephemeral, access-gated CI artifact (`actions/upload-artifact` with no other
publish step). Only the latter earns a warning; a repository with no scheduled health
workflow is `not_applicable` (nothing to rate), and a repository whose workflow's
publication status cannot be determined from a static file-tree read is `insufficient_data`
— never a warning.

**Why.** The public OSS rubric hardcoding our own internal agent's filename as a detector
constant was a home-field bias: inert for every external repository by construction, and a
free shot to hand a critic in week one (see `cejel-process-rubric-homefield-bias` in
project memory). Generalizing it to the concept it actually measures is the fix; this
changelog — and the Guard 1 regression test that enforces it going forward — is what makes
future rubric edits on a public board a public, versioned, delta-reported act instead of a
silent re-score of someone else's reputation.

**Corpus-wide before/after delta (all 17 repositories, regenerated 2026-07-12 in
goal_cejel_generalize_homefield_rule_and_rescore_protocol_2026-07-12):**

| Repository | Overall before | Overall after | Verdict before | Verdict after | Rank change | A1 finding |
|---|---|---|---|---|---|---|
| react | 3.2 | 3.2 | Conditional | Conditional | none | Same score/status (2.3 warning); the pre-existing synthetic "no single finding drove this" placeholder is replaced by a real, specific finding — react's `.github/workflows/devtools_regression_tests.yml` (schedule-triggered, `actions/upload-artifact`-only) is now the named cause. |
| vue | 2.9 | 2.9 | Conditional | Conditional | none | unchanged |
| svelte | 3.1 | 3.1 | Conditional | Conditional | none | unchanged |
| django | 3.2 | 3.2 | Conditional | Conditional | none | unchanged |
| flask | 2.8 | 2.8 | Conditional | Conditional | none | unchanged |
| fastapi | 2.9 | 2.9 | Conditional | Conditional | none | unchanged |
| express | 2.8 | 2.8 | Conditional | Conditional | none | unchanged |
| vite | 3.3 | 3.3 | Conditional | Conditional | none | unchanged |
| esbuild | 2.6 | 2.6 | Conditional | Conditional | none | unchanged |
| biomejs | 2.9 | 2.9 | Conditional | Conditional | none | unchanged |
| requests | 2.9 | 2.9 | Conditional | Conditional | none | unchanged |
| pydantic | 3.2 | 3.2 | Conditional | Conditional | none | unchanged |
| axios | 3.3 | 3.3 | Conditional | Conditional | none | unchanged |
| zod | 3.0 | 3.0 | Conditional | Conditional | none | unchanged |
| scorecard | 3.0 | 3.0 | Conditional | Conditional | none | unchanged |
| alfred (private) | 3.1 | 3.1 | Conditional | Conditional | none | Same warning, now worded generically: "results are handed only to an ephemeral, access-gated CI artifact" instead of naming an internal component. A1's sub-score moved 2.2 -> 2.3 from ordinary repo growth between the two scan commits (more test files at HEAD), not from the rubric change — see A1's `test_to_source_ratio`/`non_hollow_test_share` metric inputs in the evidence report. |
| cejel (private) | 3.4 | 3.4 | Conditional | Conditional | none | unchanged |

**Erratum — 2026-08-23 (disclosure redaction):** The internal row above was edited to replace a
closed internal product codename with the generic phrase "an internal component." This is a
disclosure-only correction: it changes no score, verdict, rank, metric input, or rubric behavior.
The original published wording remains in immutable Git history; Cejel does not rewrite history
because certificates and attestations pin commits.

**No external repository's score, verdict, or rank moved.** One external repository
(react) gained a more specific, correct finding for an already-existing warning at an
unchanged score and status — not a new penalty. No repository gained a new warning or
critical finding it did not already have. Alfred (internal) kept its own honest warning
(Guard 4), reworded to the generalized concept.

## witan-rubric-v1-2026-06-24

Predates this changelog. Introduced metric-based scoring (continuous, weighted per-criterion
metrics replacing v0's presence/absence scoring) — see this package's `src/witan/scoring.ts`'s
`usesMetricScoring` and the golden-set calibration work done in the source monorepo at the
time. No corpus-wide delta was recorded for this version; the leaderboard did not yet exist
at the time of this bump.