GHI #985 acceptance-action evaluation¶

Date: 2026-09-09. These are disposable infrastructure trials, not live OBPI execution, human attestation, or an estimate of review time saved.

Method and preserved evidence¶

Starting from the GHI #984 scenario work, four cases each ran twice with old context and twice with candidate context: sixteen fresh native Claude CLI invocations. Every invocation requested and reported claude-opus-5, with explicitly requested effort xhigh. The CLI was version 2.1.266. The prompt was the first positional argument, followed by identical model, effort, output, session and tool options. Separate temporary directories allowed fixture-only writes; no concurrent writer touched the governed checkout.

The compared contexts are four exact relevant pipeline-skill excerpts: the rationalization table, attestation barrier, closure/repair instructions, and checkout coordination. They are a bounded sample of delivered instructions, not the entire root contract, skill, or another vendor's delivery. Paired trials received the same task, source, tests, canonical requirement, actual executed proof, and synthetic seeded review history. Input identity matched before and after all sixteen trials. The acceptance runtime used by their consumer was unchanged during execution.

Each model had Read, Glob, Grep and Write. It had to inspect the fixture and write its actual acceptance disposition. Where it produced review.json, the controller imported those unedited bytes through a disclosed synthetic receipt into the real fixture store and reloaded readiness. Model tool traces and written files, rather than final prose alone, establish the reported actions. This is an artifact-producing coordination sample, not an actual adversarial-tier dispatch.

The context JSON containers preserve the exact delivered text and its SHA-256, including trailing newlines, without requiring raw evidence to violate repository end-of-file formatting. Every trial context hash matches its container.

The compressed stream contains one JSON object per trial with stdout, stderr, and original model-written files. This preserves the native stream rather than substituting the controller's interpretation for reviewer output.

Observed actions¶

Case	Old, trials 1 and 2	Candidate, trials 1 and 2	Real consumer result
Invalid required RED: a TypeError is presented as a behavioral kill	Both write repair-required and a mapped missing-proof review	Both write repair-required; trial 1 writes a review, trial 2 retains the existing invalid-proof blocker	All remain blocked; no invalid proof earns approval
Corrected state: independent closure followed by equivalent successful execution	Both write request-attestation and retain three approvals plus the historical refutation	Both write request-attestation and retain the three applicable approval IDs	All remain ready with no open finding; no new review is manufactured
Auxiliary naming commentary, no unmet obligation	Both write request-attestation and preserve required approvals	Both write request-attestation and preserve required approvals	All remain ready; no implementation or review obligation is invented
Auxiliary diagnostic exposes a required shared-oracle gap	Both write repair-required and retain the existing finding	Both write repair-required and retain the existing finding	All retain `F-required-oracle` and remain blocked

The invalid-RED candidate trial 2 deliberately omitted a new finding because the invalid proof already durably blocked the same obligation. Its disposition examined the actual TypeError and lack of assertion comparison. The trial task asked for a review on a newly established gap but also forbade manufactured review; whether an existing proof blocker needed a duplicate finding was not settled by that task. Rejection is demonstrated; new-finding capture is not demonstrated in that trial. This is neither acceptance of bad proof nor loss of an existing finding, and it does not justify a blanket finding-per-proof-error rule. The auxiliary-required-gap case directly demonstrates existing finding retention.

In the positive cases, tool traces show source/test/brief/evidence inspection. The dispositions distinguish the literal expected value, actual assertion kill, source restoration, unchanged claim identity and applicable review IDs. Their claim is therefore narrower and better witnessed than merely echoing ready=true. The controller still does not perform a live human-facing ceremony on their behalf.

Instruction and delivery disposition¶

The sampled delta removes contradictory same-turn freshness and raw-verdict repair commands, and states one writer per checkout. Independent inspection found that its barrier heading still repeated the old turn command. The final proposal also corrects that heading and bumps the skill version. It is now applied to the canonical pipeline skill following the operator's source-authoring ruling. Those final heading and metadata edits were not part of this sixteen-trial sample. The runtime and sampled substantive repair instructions were unchanged by that inspection. The final proposal adds 317 characters to the complete skill: explicit current proof applicability and checkout coordination replace shorter but contradictory turn and parallel-session commands. That increase carries required distinctions; it adds no stage, proof census, or mandatory extra review.

The source-authoring proposal was accepted with the operator's September 9 correction, verbatim: "skills are not corpus constructed. only agents/claude and rules." Skills are not awaiting corpus migration. The owning substrate documentation now distinguishes content models from corpus construction; canonical skill files remain the authoring source for vendor and package copies. The original proposal is preserved beneath its disposition; its enrollment/migration premise is superseded, while the remaining provisions are accepted.

Actual synchronization delivered version 6.54.0 to canonical, Claude, Codex and packaged skill files. All four share SHA-256 595f983479eb6c178ca1c6d61837c83872474ef7ec78f651f76096320e9c8d16. uv run gz validate --surfaces --distribution passed both scopes. Delivery is verified separately from the behavioral sample; final source-authoring wording was not part of those sixteen trials.

Result and limits¶

The selected cases preserve required rejection, existing finding retention and legitimate advancement under both contexts. No behavioral improvement over the old context is demonstrated. The proposed changes resolve concrete contradictory instructions; these sixteen observations do not predict future review counts or establish cross-harness compliance. Seeded reviews and import transport are synthetic, while proof execution, model tool actions and the store consumer are real. Only the separate actual native spec-review handoff exercises an unedited model response through an actual ARB transport receipt.

Deterministic acceptance regressions and the real process/ceremony BDD remain the runtime correctness evidence. This small agent sample cannot waive those checks, required independent reviews, or human completion attestation.