Run an installed align skill against short, informal requests in English, Korean and Japanese. Positive requests do not name the skill. Negative controls cover ordinary fixes and other meanings of alignment. Ambiguous requests are repor 3DB8 ted separately so a vague complaint is not automatically treated as an alignment task.
Read the protocol before running. The actor uses fresh project folders and conversations with the authenticated local Codex CLI. Model calls consume the account's normal usage. Fixtures and expected labels are synthetic; the actor receives only the request and fixture files.
python3 tests/activation/run.py \
--output tests/activation/results/new-run \
--private-output /absolute/path/outside-the-repository/private-evidence
python3 tests/activation/run.py \
--cases tests/activation/holdout-cases.json \
--output tests/activation/results/new-additional-run \
--private-output /absolute/path/outside-the-repository/additional-private-evidenceThe runner refuses existing output directories. No trial is silently retried. The additional wording set is maintainer-authored and must not be described as independent validation. Both sets are frozen before their runs.
To check retained evidence without model calls:
PYTHONDONTWRITEBYTECODE=1 python3 -m unittest discover -s tests/activation -p 'test_*.py'
PYTHONDONTWRITEBYTECODE=1 python3 tests/activation/verify.py tests/activation/results/2026-10-17-baseline --require-complete
PYTHONDONTWRITEBYTECODE=1 python3 tests/activation/verify.py tests/activation/results/2026-10-17-additional --require-complete
PYTHONDONTWRITEBYTECODE=1 python3 tests/activation/verify.py tests/activation/results/2026-10-17-regression --require-complete --require-clean-filesThe replay checks frozen source hashes, requests, model context, skill catalog, completion and method-output evidence. It also checks narrow file invariants. It does not judge every response's usefulness or establish another client's runtime behavior. Public paths are redacted; original command-output hashes refer to private raw text before that substitution.
For loop and goal setup, stopping and completion cases, see
the follow-up report and loop-goal-cases.json.
The report retains the rejected description candidate as well as the final
method, and distinguishes discovery from actual runtime control.
python3 tests/activation/verify.py tests/activation/results/2026-10-17-loop-final --require-complete --require-clean-files
python3 tests/activation/verify.py tests/activation/results/2026-10-17-general-v052 --require-complete --require-clean-files