# E-016 — A typed round can lack a verifiable closing criterion

- Date: 2026-08-03
- Status: protocol limitation confirmed during dogfooding.

## Observation

Round `R-HOOK-DOGFOOD` declares its objective, scope, and available checks, but the transition command does not allow declaring evidence for `AC-PRIMARY`. By default the criterion has no linked checks, and `round-end` cannot turn a successful suite run into proof of the criterion.

Attempting to add `package-test` and `docs` after implementing telemetry would have changed the contract against the baseline of the same round. This was reverted: that change would have retroactively accommodated the work, exactly the pattern Objective Guard is meant to expose.

## Impact and follow-up

The hooks and the telemetry do have technical evidence: seeded E2E for both harnesses, a global inventory per tool, and a full suite. But the round cannot claim a certified closure until the acceptance intent is declared before the work.

The next evolution of the protocol must allow `start` and `next-round` to receive acceptance criteria with check references, or generate an initial round that already contains explicit criteria. The improvement must prove that a criterion configured before editing reaches `proven` in `round-end`, and that adding one afterward remains visible as contract drift.

## Later fix

The CLI now supports `--criterion-id`, `--criterion`, and `--criterion-checks` for `start` and `next-round`. The regression creates the criterion before starting the session and shows that the declared check leaves it at `proven`. This fix does not alter `R-HOOK-DOGFOOD`: its historical result remains `UNKNOWN`, accompanied by the E2E evidence and the manual outcome that corresponds to an explicit human review.