ADR-055: the close of an adoption runs before the pull request, and the consent it needs is asked at the intake

← Decision records

status: Accepted, revised by ADR-061 (adopt.evidence gets a recommended default) and ADR-062 (record.participation retired) date: 2026-08-19

Context

An align-to-standards wave ends with three steps that produce everything this project learns from an adoption: the upstream friction review (ADR-021), the record-run offer that asks whether the session may be kept as evidence (ADR-045), and the anonymous adoption ping that feeds the count the site publishes (ADR-047). All three were written as the last items of the phase file, after "open one focused PR".

Measured on 2026-08-19, on the largest field adoption this standard has had - 25 commits, 253 files, 20 questions asked, a run good enough that its own guard findings were adopted back into this repository - none of the three ran. No friction issue was offered, no transcript was mentioned, and GET https://stats.repositorystandards.workers.dev still returned {"count":12} after the run finished. It is not a lapse of that run: across every adoption ever performed, no adopter has been asked for a transcript even once, which is the single piece of evidence the human-prompting corpus's own README names as the thing it lacks.

The cause is placement, and it is the same defect ADR-054 was written about one level up. A step at the tail of a phase file competes with a user waiting for a pull request, at the point in a long run where context is thinnest and the work feels done. The pull request is the visible finish line, so everything printed after it is read as optional.

Two fixes were considered and rejected. Gating the push itself - denying git push until the consent question has fired - puts the elicitation guard on an action rather than an artifact and, because consent is work-scoped and no committed row can settle it, would demand a telemetry-consent answer before every push in every adopting repository forever; that is the unlivable-guard failure the same branch had already learned to avoid. Leaving the steps where they are and instructing harder is what the standard has been doing, and this record exists because it does not work.

Decision

The close comes before the pull request. The phase file's steps are renumbered so that reading order is execution order: 7 upstream friction, 8 record-run, 9 the ping, 10 the pull request, which is stated to be the final action of the wave.

The consent the close needs is asked in the intake round, as a declared point. adopt.evidence is added to .claude/elicitation/points.json, gating docs/adoption-intake.md - the artifact an adoption writes early and cannot skip - and asked in the intake block beside adopt.intent:

This session can be kept as evidence for the standard - the questions it asked, the answers you gave, what it produced. May it be kept, and may an excerpt be sent upstream after you have read it?

Options, in order: keep it, and I will read the excerpt before anything is sent / keep it local - assemble it, send nothing / record nothing. recommended is null and allowed_provenance is human alone: consent is never nudged, never inferred and never stubbed. This adds no new way for an adoption to stall, because adopt.intent already gates the same file on a human answer.

The answer governs step 8 and nothing else. record-run still shows the assembled file and takes a per-item yes before anything leaves the machine (ADR-045) - the intake answer is permission to assemble and to come back, not permission to send. It does not govern the ping, which carries no session content and is disclosed rather than asked (ADR-047).

The ping reports the count it produced. The step reads the counter before and after its POST and states both numbers. A count that did not move is reported as a ping that did not land.

Consequences

One more question in the intake round, in exchange for the only evidence this project cannot generate for itself.
record-run's own record.participation point is unchanged. It gates the write of the run record, a path that exists in this repository and not in an adopted one, which is why it could never have fired during an adoption - adopt.evidence is what fires there.
An adoption that answers record nothing says so in the ledger and skips step 8 out loud. A skipped step that announces itself is the outcome this record is buying; a silently skipped one is what it is replacing.
The site's live badge becomes a check on this record: if it stops moving while adoptions complete, step 9 has drifted again.

Compliance

skills/align-to-standards/steps.md orders the wave 7-8-9-10 with the pull request last; adopt.evidence appears in standard/.claude/elicitation/points.json with recommended: null and allowed_provenance: ["human"], in the shipped ledger template, and as a real AskUserQuestion block in skills/align-to-standards/intake.md - which tools/elicitation-points-check.mjs verifies for every declared point.

Revisit when

An adoption completes and the counter still does not move, or a run reaches step 10 with step 8 unmentioned. Ordering would then have failed the same way instruction did, and the next lever is the guard on the action rather than on the artifact - with the livability problem above solved first.
Consent at the intake proves to be worth less than consent at the close - an adopter who said yes early and objects to what was assembled from it. The two-stage design (permission to assemble now, per-item consent to send later) is what this record bets against that.