Proposal (rationale)

Acts on claim 16 of the 2026-08-16 external review.

The review’s headline strategic recommendation is that the durable advantage here is not the modules, the CLI, or the templates — all copyable — but a catalogue of failure patterns keyed by how often each is actually observed:

P-026 unenforced imperative
observed across: 312 repositories
confidence: high

That is refused by an accepted decision. ADR-0005 Tier C rules out “any network transmission or cross-repo aggregation”, explicitly “including opt-in”, with the stated reason that an opt-in network path “would make every other guarantee here conditional.”

So this is not a feature request with a scheduling problem. It is an outside reader identifying the most defensible version of this product and an accepted ADR refusing it on trust grounds. Both positions are coherent, and the collision deserves an answer on the record — because it will be proposed again, by someone who has not read Tier C, roughly every time the project is discussed.

Two things make this more than a formality. First, the ADR’s stated reason is about transmission, while the review’s counts could in principle come from repos the operator reads themselves — a distinction Tier C does not currently draw, and possibly should. Second, this repo’s own biggest known weakness is that its four source repos share an author, so every convergence in the pattern catalogue is one operator agreeing with themselves — which is an argument for wider observation that the review did not make and that already has an accepted answer of its own (WI-018: read public repos, do not collect from users).

Decision

accepted — 2026-08-16, directed by the user. Opened proposed deliberately, because a spike that assumes its own outcome is not a spike; the outcome below was not the one expected when it was opened.

Plan

Filled on acceptance. What follows is the question and its boundaries, which are what a decision needs; the method is one document and a conversation.

Requirements

• Produce one outcome, dated, in the decision record: ADR-0005 amended with a new trust posture and the conditions attached, or ADR-0005 reaffirmed with the refusal restated and this item’s argument named as considered-and-declined.
• Distinguish the three things the review’s single mockup bundles, since they have different costs and different answers: (a) counts collected from users’ repos, (b) counts derived from public repos the operator reads, (c) a per-repo count that never leaves the machine and is already Tier A.
• State the cost of each honestly — trust posture, privacy surface, the infrastructure it implies, and what it does to every other guarantee the tool makes.
• Whichever way it goes, external-review-2026-08-16.md §4.2 links to the outcome, so the adjudication does not read as unresolved forever.

Impacts

• ADR-0005 — amended or annotated. Tier C is cited by product-brief.md §8 and by README.md’s design commitments, so an amendment touches both.
• No code. A decision that lands as an implementation is a decision nobody made.
• If the answer is (b) — public repos only — it is closer to WI-018’s existing method than to anything new, and may need no ADR change at all. Establish that before proposing one.

Approach

To be written on acceptance. The shape: one document that puts the review’s argument at its strongest next to Tier C’s stated reason, separates (a)/(b)/(c), and ends in a recommendation a person accepts or rejects. The strongest form of the argument for is written by whoever is inclined to refuse it — an ADR that beats a weak version of its alternative is the same defect as a gate that has never fired.

Timebox it. This is a decision, not a research programme, and it must not block the release.

Acceptance criteria / tests

1. ADR-0005 carries a dated outcome — amendment or reaffirmation — with the reasoning.
2. (a), (b) and (c) are separately answered; a single verdict covering all three is a failed spike.
3. The argument for is stated at its strongest before it is answered.
4. §4.2 of the adjudication links to the outcome.
5. If reaffirmed, the refusal is stated somewhere a future proposer will actually meet — not only inside a 250-line ADR.

Out of scope

• Building anything. No collection, no endpoint, no opt-in flag, no schema — including a disabled one. Nothing deferred: if the answer is yes, the implementation is a new item with its own plan.
• The per-repo ledger. Already accepted as Tier A and unaffected either way.
• Widening the research corpus. WI-018 owns that and is already accepted.
• Reopening the rest of Tier C — rework rate, session outcomes, instruction attribution, composite health scores. This item is the aggregation bullet only. The score bullet is separately reaffirmed in §4.1 of the adjudication.

Execution

Branch spike/WI-041-decide-cross-repo-evidence, cut from main at 4e2b8f1. The argument is docs/design/cross-repo-evidence-2026-08-16.md; the outcome is in ADR-0005 itself, which is what this item existed to produce.

The expected outcome was “reaffirm”. It was not.

Separating (a), (b) and (c) — the requirement that looked like bookkeeping when this was planned — is what produced the result. Tier C’s wording forbade research this repo was already doing.

The bullet said “any network transmission or cross-repo aggregation”. Read plainly, “cross-repo aggregation” covers counting things across the fourteen public frameworks extracted under WI-009 and WI-018 — each pinned to a commit SHA with its licence recorded, precisely so the counts are reproducible. One sentence in the ADR forbade the method another part of the repo required.

That is the decayed-rule failure this whole project argues against, sitting inside the ADR that argues it, and it had been there since 2026-08-14 with nobody noticing — because nobody had had cause to read the bullet against the research until an outside reviewer proposed something it appeared to forbid.

Review

Verified 2026-08-16.

1 · ADR-0005 carries a dated outcome with the reasoning. Tier C’s second bullet is amended in place, with the old wording and the reason for the change quoted beneath it rather than replaced — an amended-away error reads as an error that never happened. Met.

2 · (a), (b) and (c) separately answered; a single verdict would be a failed spike.

Met, and the split is the finding.

3 · The argument for is stated at its strongest before it is answered. The document opens with the reviewer’s case and their strongest supporting fact — that this repo’s own worst weakness is one operator agreeing with themselves, so an argument for breadth is not obviously wrong here. Met.

4 · The adjudication links to the outcome. Both the first review’s §4.2 and this item. Met.

5 · A future proposer meets the refusal somewhere they will actually read it. The item worried this would be buried in a 250-line ADR. It is now a revisit trigger — the section a reader checks when they want to reopen something — stated as three conditions rather than a wall, because the proposal is reasonable and has already arrived twice. Met.

What this deliberately did not do

No collection, endpoint, opt-in flag or schema was built, including a disabled one — the item’s own Out of scope. And the census’s 0% false-positive result is not treated as evidence that breadth is unnecessary: it was measured across one operator’s repos, which is the limit at issue, and it appears in the reopening conditions as the experiment to run properly rather than as an answer.