Every completion now says whether anything checked it independently

[qwerfunch]

Every completion now says whether anything checked it independently
2026-07-26 · feature/role-contract-architecture

When one AI writes both the code and its tests, the tests get shaped around the code it just wrote — the bug is there and everything still passes. This adds a plain record of that: every finished feature is marked according to whether the building and the checking were done by different hands. It also stops cladding from telling you how to arrange your agents, and rewrites the Multi-Agent section of the README around why any of this matters.

Heads-up: nothing is blocked by default — the mark is only a note on the record. One honest limitation worth knowing before you rely on it: today almost every completion will read self-certified, because the only way to record a genuinely blind check is through your AI tool, and a person's own sign-off can't be recorded at all yet. Both gaps are written down in the backlog rather than papered over.

Also carries the "spec entry, not shard" wording fix, which was pushed earlier but never opened as its own pull request and is still missing from develop. It rides along here rather than being lost.

Added

▸ A mark on every finished feature — independent when something checked the work without being able to see the code, self-certified otherwise. It describes the evidence, never which agents did the work or how many.
▸ An optional stricter setting — teams that want it can refuse to count unchecked work as finished, in which case a self-certified feature is put back to unfinished exactly like a failed check.
▸ A diagram in all four languages showing how a finished feature gets its mark — one question, two outcomes, and the refusal the stricter setting adds.

Changed

▸ cladding no longer tells you how to arrange your agents. The coordinator prompt dropped its routing table, numbered principles, named agent chain and work-in-progress table, and now states only the conditions a feature must satisfy. How many agents run, which models, and in what order is the host's call.
▸ The specialist prompts are selectable briefs, not a fixed cast — any agent may take up any of them, and one that never touches cladding needs none.
▸ The README's Multi-Agent section leads with the stake rather than the vocabulary: it explains the self-testing problem first, then the three ways a feature can end up marked, each with its own reason.
▸ Version moves to 0.9.2 across every version site and all four published READMEs.

Fixed

▸ Numbers that had drifted — the feature and test-file counts on the front page and in the status table were months out of date.
▸ A broken layout in the HTML READMEs — the "before / after" pair rendered as a single narrow card with two-thirds of the row empty, because one card sat alone in a three-column grid.
▸ A false claim about human sign-off. The section said a person approving a review earns independent. No code path records that, so the claim was removed rather than softened.

What we measured, and what we didn't

▸ The choreography removal was tested against a control — six agents, three per arm, differing only in the coordinator prompt, scored by a hidden check against each arm's final code. The planted bug was caught 3/3 in both arms: removing it did not weaken anything.
▸ No benefit is claimed. Effort favoured the new prompt but the ranges overlapped at this sample size, so the decision rule written before the run forbids calling it a win. The change is recorded as safe and neutral.
▸ Not measured: discipline across a long run of consecutive features. That phase was cancelled once it was clear the risk it targeted is already enforced by the gate, and it is recorded as not measured rather than answered.

Gate green, 2736 tests passing.

🤖 Generated with Claude Code