External review, 2026-08-16 — record and adjudication

Authoritative for: what an outside reviewer said about rungs on 2026-08-16, and which of those claims survived checking against this checkout. Not authoritative for: what rungs is (that is product-brief.md), any decision it recommends (those are ADRs), or the work that acts on it (WI-037).

An unsolicited review of the public repo and the docs site arrived on 2026-08-16. It is recorded here rather than summarized into a backlog item, because the parts that were wrong are as useful as the parts that were right — and a review that only survives as the items it produced cannot be re-read when a later decision contradicts it.

What this document treats the review as

A claim about this repo, not evidence about it. The reviewer read the public README, the docs site, and a handful of linked documents from outside; every statement below was re-derived against the working tree on 2026-08-16 before being retained. Where the review and the checkout disagree, the checkout is what shipped.

Two limits on the source, stated once so no row has to repeat them:

• It could not run the tool. Every claim about behaviour is inferred from prose. This is why three of its sharpest observations are directionally right and mechanically wrong — it describes capabilities as missing that exist but are unreachable, which is a different defect with a much cheaper fix.
• Its adoption evidence is a snapshot with no date and no command. “0 stars and 0 forks” is the kind of number CLAUDE.md requires a date and a command for. It is not retained as a measurement; it is retained as the reviewer’s read of earliness, which is not in dispute — first publication was 2026-08-14.

The reviewer’s scores

Reproduced because they are the review’s own summary of itself. These are opinion, unmeasured, and this document does not adopt them — there is no instrument here that produces a 6.5.

The scores are also, notably, the thing this repo’s own ADR-0005 refuses to emit about anyone else’s repo. That symmetry is not an argument against the reviewer — it is the reason §“Where the review argues against an accepted decision” exists below.

1. Adjudication

Verdicts: retained (accurate and actionable) · partly (the observation holds, the proposed remedy does not) · rejected (checked and false, or blocked by an accepted decision) · noted (opinion, plausible, not acted on).

2. The finding worth the whole review

The review says doctor should tell an unfamiliar repo what is wrong with it, and proposes two surfaces for that — doctor --explain and rungs lint agents. Both describe capability that exists and cannot currently be reached.

What doctor does today: src/cli.ts:77-183 scans the repo once and maps fifteen module signatures over the result, reporting each as ours / theirs / different paradigm / absent, then names one next command. It answers “which of our modules does this repo already have an equivalent of?” — a presence question, and it says so:

This reports presence, never quality. It cannot tell whether an adopted system is good, complete, or working — only that files are where a module’s files would be. — src/cli.ts:152-154

What the review asks for is a defect question, and the detectors for it are already written — as gates, in the modules, reachable only after installation and only over rungs-managed content:

So the gap is not “rungs cannot analyse a repo”. It is the analysis is gated behind installing the thing the analysis is supposed to justify installing — which is exactly backwards for an adoption wedge, and is a much smaller fix than the review assumes. This is WI-038.

One caveat the review does not raise and the item must carry: several of these gates read rungs-shaped inputs. Running them over an arbitrary repo will produce findings whose evidence is sound and whose framing assumes conventions the repo never adopted. Under-report deliberately — ADR-0004 already made that choice for detection signatures and the reason carries over unchanged.

3. Where the observation holds and the remedy does not

3.1 The backlog and external trackers

The review’s question — “why isn’t this Linear/Jira/GitHub Issues?” — is fair, and its proposed answer (a provider-neutral work model with four adapters) breaks two accepted decisions:

• ADR-0002: a scaffolded repo acquires no new language runtime. A Linear or Jira adapter needs credentials, network access and a client in the consumer repo. That is not a module; it is a different product with a different trust posture.
• ADR-0004 already answers the ownership worry a different way: adoption is a mapping, not a migration. add records where a repo’s equivalent lives and never rewrites it.

What is genuinely missing is one link in that chain. The backlog module declares exactly one [[detect.paradigm]] — milestones (modules/backlog/module.toml:103) — so a repo running its work in GitHub Issues today is detected as having no backlog, and doctor proposes installing a Markdown one beside it. That is the review’s fear, and it is real; the adapter layer is not the fix. A paradigm signature is, and it is the mechanism ADR-0004 already shipped for this exact case. WI-039.

3.2 Vocabulary

Retained narrowly: a public glossary and one pass over the first screen of each public surface. Not retained: renaming internal concepts. The nouns the review objects to are mostly load-bearing in design documents, where precision beats approachability, and the repo’s own one-definition-per-concept rule means a rename is a repo-wide edit for a readership that never sees the term.

4. Where the review argues against an accepted decision

Both belong to ADR-0005, and neither is a near-miss. A person has to decide these; an epic must not adopt them by writing an item.

4.1 The maturity bars

The review is explicit that it does not want a health score —

Not some fake AI-generated “health score.” Actual evidence. Actual files. Actual invariants.

— and then draws one:

Agent instructions       ████████████ mature
Validation enforcement   ███░░░░░░░░  weak
Decision authority       ████░░░░░░░░ fragmented

mature / partial / weak / fragmented over incommensurable signals is the composite score ADR-0005 Tier C refuses permanently, and the ADR names the failure mode precisely: “a probe encoding a guess is a gate that is confidently wrong.” The half of the mockup underneath it — HIGH: 17 imperative rules have no mechanical enforcement, with the files — is judgement-free, countable, and exactly what §2 retains.

Disposition: take the risk register, refuse the bars. No revisit trigger fires here; ADR-0005’s triggers are about the ledger, not about grading a foreign repo.

4.2 Cross-repo pattern counts

The review’s proposed moat is a catalogue keyed by observation frequency:

P-026 unenforced imperative
observed across: 312 repositories

ADR-0005 Tier C refuses “any network transmission or cross-repo aggregation, including opt-in”, with the stated reason that an opt-in network path “would make every other guarantee here conditional.”

This is the review’s headline strategic recommendation colliding with an accepted decision. It is not resolvable by an implementation item, and it should not be quietly declined either — the review is describing the one thing here that would be hard to copy. Disposition: open it as a decision, not as work. WI-041 puts the question, its cost, and the ADR’s stated reason in front of a person; the honest outcomes are an ADR amendment with a new trust posture, or a written refusal that stops the idea returning every quarter.

Answered 2026-08-16. Both, in different halves. The phrase covered three things and only one is refused: collection from users stays refused permanently, reading public sources was always permitted and the ADR’s wording had accidentally forbidden research this repo was already doing, and the local ledger is unchanged. ADR-0005 is amended, with a three-part test for reopening the refused half rather than a wall. cross-repo-evidence-2026-08-16.md.

5. What the review did not look at

Recorded so the next reader does not mistake this document for a full assessment:

• Nothing about the modules’ contents. Fifteen modules, and the review evaluates the catalogue as a count. The claim under test in module-catalog.md — that each module’s provenance is a real incident — is the repo’s central claim and went unexamined.
• Nothing about the research. ~80 patterns and four autopsies, cited as volume. The single most consequential weakness in this repo is already on record and the review did not find it: the four source repos share an author, so every convergence is one operator agreeing with themselves (BACKLOG.md, WI-009’s rationale). An outside reviewer landing on that would have been worth more than the messaging notes.
• No execution. No doctor run, no init, no gate output — see the limits stated at the top.

6. What this produced

WI-037, an epic with four children, targeted at the next release. The epic’s own constraint is the review’s ninth claim: it adds no module and no pattern. Everything in it either makes an existing capability reachable, corrects a surface that contradicts another surface, or puts a decision in front of a person.