Benchmark authoring and evidence

• Treat context-control arguments as requests, not verified removal. For claims about skill exposure or instruction timing, inspect matching recorded initial messages when available: a later file read may repeat an already-injected body. Catalog mentions alone do not prove executable availability. Keep unavailable evidence unknown and never publish private instruction text as benchmark data.
• Before freezing a new QA fixture, exercise the supplied test support through both a passing check and a deliberate failing assertion. Verify that failure output preserves the relevant actual/expected values, rather than reporting a support-code exception. Nominal green tests alone are not a harness preflight.
• For native Python fixtures, also import the public package entrypoint in a fresh process using the same source path/interpreter as the model task. A unittest selector can retry a failed test-module import and use cached submodules while the package itself still cannot import. Check required generated metadata explicitly; a passing native test is not bootstrap proof.
• Define input, payload and failure contracts before calling a fixture clean. Do not count a reasonable finding under an underspecified contract as an agent false positive merely because the author did not intend it.
• Before freezing, ensure every scored obligation is explicit in the model-visible task or project instructions. For example, "disposable copies" alone is not an explicit end-of-task deletion requirement. Disclose discovered mismatches; do not silently rescore measured attempts.
• For project-only tasks, give fixture tests explicit project-local temporary directories and verify actual scratch creation paths during preflight. Do not rely on system temp defaults or a model changing global TMPDIR/launcher settings to repair the fixture. Preserve any already-measured version and disclose it.
• Preserve frozen fixtures, criteria and every scheduled cell. If a fixture defect is discovered during execution, disclose the effect on interpretation; correct a future version, not the input underlying an already reported run.
• Separate original model evidence, author replay and resource arithmetic. Lower recorded cost with unequal work or invalid controls is not by itself an accepted performance win. Keep existing featured results tied to their actual source revision and obey the repository's localization/chart synchronization rule.
• New runner tests must work from a source archive without repository history or local-run artifacts. Exercise scheduling controls with explicit synthetic resources; isolate actual pinned-resource comparisons behind require_history. Check the affected tests in a fresh archive, not only the Git checkout. Never treat synthetic resources or skipped historical checks as model evidence.
• For retained unittest code, scan_test_contracts.py can locate async overrides of synchronous runner methods without executing the project. Review inherited repairs and actual assertion paths; a candidate is not a scored failure, and no candidates is not proof of correct tests. Disclose new post-run findings without rewriting frozen observations or criteria.