Proposal (rationale)

The same outside reviewer re-assessed rungs after WI-037 shipped. Recorded and adjudicated in external-review-2026-08-16b.md.

Most of it agrees with decisions already made, including one the reviewer had argued against first time — they withdrew their own recommendation for tracker adapters after reading why it was refused. Agreement produces no work. Five things do:

1. The most differentiated capability is invisible from the entry point. doctor prints no occurrence of the string explain, so the analysis is discoverable only from --help. The reviewer’s shape — a short count and the command, not the findings — respects WI-005’s one-next-command rule. → WI-049
2. The board contradicts fourteen item files. Nine rows under Proposed and five under Planned name items whose files read status: done; nine now link into archive/, so the board says proposed about a document filed under “cannot change any more”. This is the finding the reviewer should have made and did not — they checked the corpus, got the right answer, and our board would have told them otherwise. → WI-050
3. The site’s status line is stale — “20 gates register” against 21 since WI-044. Typed into site.config.ts, whose own comment calls this the thing this repo has the most scar tissue about, next to a TODO (generate-derivable). → WI-051
4. Applicability is not first-class. The reviewer’s strongest technical recommendation across both reviews: can this detector legitimately interpret this repository must be asked before did the condition fire. It exists — as two hard-coded sets of engine names in one file, which is the shape that goes stale. → WI-052
5. Generalisation is now the main risk, and it is unmeasured beyond the four source repos. → WI-053

Why now. The reviewer’s own priority order is make doctor excellent on arbitrary repos, measure false positives brutally, make public claims reproducible, get outsiders running it — and the rift-forge incident is our argument for it, not theirs. Nothing here broadens the product.

Decision

accepted — 2026-08-16. The user supplied the review and directed the same process as the first. Claims verdicted noted produce no work and their reasons are in the adjudication rather than restated per item.

Plan

Requirements

• No new module and no new pattern. Same first requirement as WI-037, and the reviewer’s explicit advice.
• Each child names the review claim it acts on, so the adjudication stays checkable.
• No child weakens ADR-0005 Tier C. The fleet product the review speculates about is the same cross-repo boundary WI-041 still holds, deliberately unopened.
• Every count written to a public surface is derived or dated-and-gated; none is typed and left.

Impacts

• CLI: doctor gains a summary line (WI-049). No new command.
• Gates: one new declared gate reconciling the board (WI-050). 21 → 22.
• Site: site.config.ts numbers become derived or checked (WI-051).
• Architecture: detector applicability moves from hard-coded sets to a declared property (WI-052). This is the only structural change and it may want an ADR.
• Evidence: a false-positive census across every repo available locally (WI-053).
• Release state is not impacted: the 0.1.2/0.1.3 ambiguity the review flags is already WI-035’s, which is planned. Opening a second item for it would be the duplicate-tracking failure this repo exists to argue against.

Approach

Sequence: 050 → 049 → 051, then 052 → 053. WI-050 first because it is the smallest and because a board that lies makes every later claim about “what is done” unverifiable. WI-052 before WI-053 so the census measures the abstraction rather than the current hard-coded sets — otherwise the census has to be re-run.

Where the review is followed against instinct: claim 13 asks for analysis to be advertised from plain doctor, and WI-038 deliberately put it behind a flag after measuring 114 findings on hexguard. Both are right — the fix is to advertise the existence and the count, never the findings. If WI-049’s output grows past a few lines it has failed.

Open: whether WI-052 warrants an ADR. It changes how every detector declares itself, which is ADR-0003 territory. WI-052 decides against its own diff.

Acceptance criteria / tests

1. All five children reach done, or one is closed with a written reason naming the claim it drops.
2. rungs doctor names the analysis and its size in no more than three lines, and still ends with exactly one recommended command (WI-005 not regressed).
3. No board row’s group disagrees with its item’s status field, and a gate refuses it.
4. No number on a public surface is both typed by hand and unchecked.
5. Detector applicability is declared per detector, not listed centrally by engine name.
6. A false-positive rate is published per repo across every repo available, with the method stated.
7. rungs check, npm test, and the site build pass at the epic boundary.

Out of scope

• The fleet / multi-repo product. ADR-0005 Tier C; WI-041 holds the question and stays proposed. Nothing deferred here.
• Publishing 0.1.3 or resolving the release-state ambiguity. WI-035.
• New detectors. WI-053 may report that a class of problem is undetected; adding one is a separate item, and only once independent repos show it recurring — the review’s own rule.
• Anything verdicted noted. On record in the adjudication; produces no work.
• Whether --explain’s surviving findings are valuable — the question §4 of the adjudication says nobody has asked. It needs a repo owner, not a measurement, and there is no honest way to answer it from here.

Execution

Landed 2026-08-16 in the planned order — 050 → 049 → 051 → 052 → 053 — with one insertion.

WI-054 was inserted between 050 and 049, out of the epic, because WI-050 immediately exposed F-016: upgrade --apply never re-registered gates, so the gate WI-050 had just shipped could reach no existing install. It is not a child of this epic — it acts on no review claim — and it is recorded here because the sequence would otherwise look like it drifted.

The pattern across the five

Four of the five were changed by measurement, not by the plan. That is the epic’s most useful output and none of it was predicted:

• WI-049 was specified with a finding count. Measured, that took plain doctor on rift-forge from 1.6s to 16.8s — a 10× tax on the entry point to advertise a flag. It ships reporting the number detection already computed, and one acceptance criterion is recorded unmet rather than reworded.
• WI-050’s requirement 4 was wrong: reporting every undeclared board heading produced seven findings against a correct document. The typo case it was aimed at is caught precisely instead.
• WI-051 left the run result typed, reasoning it was a dated measurement like the research snapshots. Within minutes the page read 23 gates register beside 22 pass, same day — the incoherence the item existed to remove, reintroduced by the one part exempted from it.
• WI-053 found that 63 of 82 directories are one project’s worktrees, and that rift-forge’s count had moved by 63 findings in a few hours because it is somebody’s live repo.

Review

Verified 2026-08-16.

1 · All five children done. No claim was dropped. Met.

2 · doctor names the analysis in ≤ 3 lines and still ends with one command. Three lines including the heading; WI-005 not regressed. Met with a stated exception: a repo where the analysis would find nothing is still told it exists, because knowing otherwise costs the 15 seconds criterion 5 forbade. Recorded in WI-049 as unmet rather than redefined.

3 · No board row disagrees with its item’s status, and a gate refuses it. backlog-board-reconciled examines 12 rows and passes; the fourteen that disagreed when the review arrived are corrected. Met.

4 · No number on a public surface is both typed and unchecked. Structural counts derive from generated/claims.json and site-claims-current refuses drift; the run result is captured by npm run claims. phase.label remains prose, deliberately and on the record. Met.

5 · Applicability declared per detector, not listed centrally by engine. 41 gates, three values, explain.ts holds no engine names. Proven behaviour-preserving by byte-identical --explain output on all four source repos. Met.

6 · A false-positive rate published per repo, with the method stated. 2,291 findings, 6 repos, 0 wrong, 0 unclassified, classifier proven able to return every verdict first. Met.

7 · Gates, tests, site. rungs check → 23 pass · 0 fail · 0 unimplemented · 0 error; npm test → 20 pass; site builds with 0 broken links. Met.

What the epic did not resolve

• Generalisation. The census measures eleven project shapes built by one operator. The review’s actual test — twenty repositories nobody involved touched — cannot be run from here, and §4 of the results document says so rather than letting 0% imply otherwise.
• Whether findings are worth acting on. Of rift-forge’s 1,994 surviving findings, nobody knows how many its owner would fix. Low false positives and low value look identical from outside.
• WI-041 stays proposed, as directed — the cross-repo evidence question the review’s fleet-product speculation reopens, and ADR-0005 Tier C still refuses.
• F-017 stays open: upgrade does not update .ai/rungs.toml, and the obvious fix would stamp our hash onto user-diverged files.