Testing and Evaluations

What the evidence establishes

The v7.2.0 release is published. Its publication receipt records the exact source, designated review, required jobs, artifacts, restored tag protection, and publication checks. This is the starting point for current release status. The v7.1.0 release retains its own records.

The v7 implementation evidence packet is a pre-publication snapshot. Its conditional verdict records what remained open at that time. Publication did not rewrite that history into a different judgment.

Read each claim at its evidence level

Level	Useful evidence	Limit
Source and contract integrity	Schema validation, inventories, refusal controls, template tests	Does not establish model application
Delivery and discovery	Observed installed identity, discovery response, loaded body	Does not establish use or benefit
Exercised behavior	A task or runtime observation at a named revision	Covers the observed case and environment
Comparative benefit	Valid controlled comparisons with preserved failures	Cannot be inferred from structural checks

What passed at release

The exact release gate receipt lists the dispatched runs on the exact v7.2.0 merge commit. Every required job passed: the five dispatchable gate workflows, the required Ubuntu custody contract job, and CodeQL. The dispatch-only macOS custody diagnostic failed the two known case-insensitive filesystem assertions and is disclosed separately. After tagging, the original version-tag creation rule was restored and a seeded Git push was rejected. At v7.0.0, a live Codex CLI check in an isolated workspace listed all 17 skills; that live host check was not repeated for v7.2.0.

The designated review was done by the Codex agent that prepared the release, as designated in the pre-authorization. It used the same task context, so the review was neither independent nor blind. The four new research-gap fixtures check declared structure; they have no scorer or live agent run and do not establish coverage-map quality in use.

Important limits

• macOS custody: required Ubuntu contract checks passed. The dispatch-only macOS diagnostic failed the two known case-insensitive filesystem assertions in issue #162; subsequent macOS steps were skipped. The aggregate workflow therefore remains red. No case-insensitive POSIX custody guarantee is claimed.
• Research-gap behavior: the v7.2.0 literature contract and structural fixtures were checked; no live coverage-map trial or comparative outcome test ran for this release.
• v6 to v7 comparison: one baseline attempt ran, but the environment violated the intended provider isolation and blocked necessary reads. The remaining 31 slots were not dispatched. There are zero valid pairs and no v7 superiority claim. See the preserved pilot result.
• Earlier four-arm experiment (August 2026): a preregistered, blinded run of 72 trials compared four versions of the instructions, one of them an empty baseline, on an earlier version of the methods. It found no separation between the four (p = 0.875 on the preregistered main test). Its size and a scoring adapter added after the run make it exploratory, and it supports no superiority claim in any direction. See the preserved four-arm result.
• Host application: the host coverage record distinguishes observed discovery from unexercised loading and workflows.
• Acceptance calibration: deterministic UAT record judging does not establish that all observations are correct or empirically calibrated.

For reproducing checks after a change, see the maintainer guide. Retain historical failures and scope; add new evidence instead of silently overwriting them.