Comparison execution record

Dated 2026-09-19. Persona: main-session. GHI #1048.

The operator authorized testing the proposed remedy and corrected two turns that ended with promises instead of execution. The correction was captured through the insights command before this comparison proceeded. No further permission was required to run the approved experiment.

Independent curation selected two prospective tasks without reading earlier raw answers. A separate reviewer, without candidate text, challenged the criteria. The incorporated clarifications and read limits are in cases.json. The freeze in inputs.json occurred before any subject dispatch.

Corpus construction used a Git archive of f4a0395a. It retained regular files under src/, tests/, data/, .gzkit/rules/ and docs/user/; primary canonical SKILL.md bodies; root and docs instruction files; .gzkit.json, the manifest and pyproject.toml. No symlinks were supplied. corpus.json enumerates included files with hashes and omitted regular-file paths. Required oracle witnesses are in the supplied corpus. Omitted doctrine/history limits conclusions about whole-repository authority and ownership; subjects must name uncertainty rather than invent missing authority. Current source may itself contain historical comments, so this is exclusion of evaluation artifacts, not removal of history from all code.

The harness rejected a third concurrent subject with agent thread limit reached. Execution therefore uses two active subjects at a time, with the same frozen prompts and model settings. No failed spawn produced an answer; no returned answer is replaced or discarded for its score. Dispatch order interleaves the two tasks and conditions as recorded in inputs.json.

Assessment tasks are stimuli, not repair authorizations. Observations outside their requested behavior remain leads recorded in cases/raw answers and the final disposition, rather than silently expanding this investigation.