status: Accepted date: 2026-09-16

ADR-062: a contributed run is the full run, anonymised, or nothing

Context

ADR-045 gave record-run two consent levels: Level 1, the user's literal turns with three summary yes/nos and one result line, and Level 2, the same plus the agent's own text, the tool log and the repository slug. ADR-061 then gave the intake point adopt.evidence three answers: send it, send it, once I have read it, send nothing.

Two external records have arrived since. anon-r3k (2026-09-04) and anon-x8v (PR #143, 2026-09-16, closed unmerged) both chose Level 1. Both carry provenance: unverified and transcript: null, and not one of their verdicts can be replayed by anybody who was not in the session: the scoring was done from memory of the run by the session that ran it. ADR-045's own revisit clause named this outcome - "Level 1 alone proves to still collect too little to explain a finding" - and record-run already stated that every finding the human-prompting method has produced so far needed the agent's own text. The cheaper branch was offered so that a smaller yes would beat a large no; what it collected was a record that is not evidence.

The owner's decision, 2026-09-16: the full run is always what is wanted, and that decision is not handed to the adopter - either they contribute something real that can be checked, or they do not contribute. If they agree, everything is anonymised.

Options considered

A - Keep both levels; recommend Level 2. Rejected: a choice the adopter can take the cheaper branch of is still a choice, and the two records so far show which branch gets taken.
B - Level 1.5: agent text without the slug, as a third option beside the two. The middle ground - what the corpus needs to explain a finding, minus the name. Rejected for the same reason as A: three branches is still a menu with a cheapest item on it, and the point of this decision is that there is no cheaper branch. What Level 1.5 describes is not a third option; it is the one shape, made mandatory.
C - One shape: the full run, fully anonymised, or nothing (chosen). The record carries the literal user turns, the agent's text verbatim, the tools that ran in order (names only, never their input or output), the scoring and the result line, and nothing that identifies the repository: an opaque code not derived from its name, no owner, no paths, no hosts, no usernames. Consent is one intake question with two answers.

Two further variants were rejected on the way to C:

A held-for-review answer (ADR-061's send it, once I have read it). The pull request record-run opens from the adopter's own account is the review: it is theirs to edit or close before anyone upstream reads it, so a second look in the session adds a step without adding a safeguard. And the 2026-09-03 measurement behind ADR-061 showed adopters take the last answer on a list; a middle answer only moves where the run stalls.
A named, not anonymised, variant. Both external records so far were anonymised and private, so the name buys the evidence nothing - a finding is replayed from the turns, not from the slug. A list of adopters willing to be named is a marketing surface, kept separate from the evidence record if it is ever kept at all.

Decision

Option C.

adopt.evidence has two answers, in order: send it (recommended) / send nothing. Its asks says what goes out and that it is anonymised. allowed_provenance stays human alone.
record-run asks no consent question at the close. It reads the intake answer from the Evidence line of docs/adoption-intake.md; under send it it assembles the full run, scrubs it, shows the assembled record once in the session as a courtesy, and opens the pull request from the adopter's account; under send nothing it records nothing and says so. record.participation (whose run this is, and may the excerpt be kept) is retired. Its second half is consent, settled at intake. Its first half is answered without asking: the pull request opens from the adopter's own account, so real-adoption is read off the author - a maintainer's own run keeps the validation scope - and whether a person answered the questions or the agent answered for itself is read off the agent turns the one shape always carries. The ADR-054 failure, a maintainer's run recorded as an external adopter's, cannot recur from an answer typed at the close because no answer is typed at the close. The write gate on docs/validation/**/runs/*.json moves to adopt.evidence, so a run record still needs a human intake answer in the transcript before it is written.
Every record gets an opaque code; the subject line feat(real-adoption): <code>, <stack> - what the run showed never carries a repository name.
The step-3 scrub is unchanged and now also replaces the repository's name, owner and slug wherever they occur in the record.
tools/human-prompting.mjs refuses a run record dated after 2026-09-16 that carries no agent turn and no transcript. Records before that date are history and still pass.

This revises ADR-045 (decision point 2, the two-level split), ADR-061 (the three-answer set) and ADR-054 (the record.participation point it declared; the provenance states on run records stand). Everything else in the three stands: the corpus record-run feeds, the scrub, the intake placement, the once-per-run consent, the recommended default, the guard.

Consequences

Positive: every record that arrives from now on can be replayed - the verdicts sit next to the text they were scored against.
Negative, stated plainly: some adopters who would have sent prompts only will now send nothing. That is accepted; the prompts-only record was not paying for the row it took.
record-run asks nothing at the close, and the ledger template loses one row. What the retired question used to attest - that a person, not the agent, stood behind the record - is now attested by the pull request author and by the transcript, neither of which the agent can type for itself.
Negative: the scrub carries more weight, since the agent's text is now always in the record and may quote what a tool read. Step 3 drops raw tool input and output outright, and the pull request from the adopter's own account is the human read in front of the merge.

Compliance

standard/.claude/elicitation/points.json: adopt.evidence.asks and why rewritten, recommended stays "send it", and the point gates docs/validation/**/runs/*.json as well as the intake file; record.participation removed. tools/elicitation-guard-test.mjs asserts the record is still refused without the intake answer and allowed with it.
skills/align-to-standards/intake.md: the [adopt.evidence] block lists two options.
skills/align-to-standards/steps.md step 8, standard/docs/adoption-intake.md, CONTRIBUTING.md, docs/validation/human-prompting/reporting.md, services/adoption-stats/README.md: the two-level and three-answer wording is gone.
skills/record-run/SKILL.md: one shape, no consent question at the close, opaque code on every record.
tools/human-prompting.mjs and its test: the dated refusal above.

Revisit when

A pull request opened under send it carries something the scrub missed and the adopter did not catch before merge - the pull request is the review, and this decision removed the in-session read that used to sit in front of it.
Real adoptions show send nothing taken so consistently that the corpus stops growing at all - the trade here is fewer records for checkable ones, not none.

ADR-045 - the two-level split this revises; its corpus destination, scrub and consent gate stand.
ADR-061 - the three-answer set this narrows to two; its once-per-run consent and recommended default stand.
ADR-054 - declared record.participation; the point is retired here, the mechanism and the provenance states stand.