Existing source-7 text-authoring experiment
Executed 2026-09-11 by Codex-9e8r6t /root/palette_finish. See the measured appendix in harness/release-textual-usability-audit-20260911.md.
From the repository root on this configured Windows host:
. .local/toolchain/env.ps1
& 'C:/Users/User/.cache/codex-runtimes/codex-primary-runtime/dependencies/python/python.exe' harness/experiments/textual-authoring-20260911-local/run_experiment.py
The script uses the existing server executable, a fresh ignored experiment store, and an OS-assigned loopback port. It never compiles or opens a GUI. It requires the explicitly referenced preserved source snapshot for provenance comparison. Each invocation overwrites the result files, so preserve them before rerunning. It terminates only its own server after completed synchronous requests and verifies saved source/model by reopening the store.
summary.json and http-results.json record the passing 26-request run. Their JSON string values preserve exact submitted/accepted source newlines; the separate input/stabilized/renamed text snapshots use Windows newlines through Python's text writer. Byte-preservation assertions compare the original submitted and returned strings before that snapshot conversion. provenance.json records the binary and all 120 core/server snapshot/current file hashes. review-results.json records separate inventory/hash equality, exact advisory warning counts, and artifact hashes.
The first attempt reached successful OA checks and imported the scenario, then failed because this experiment's expected JSON wrapper was incorrectly written as Shared/layer/body/Scenario. Production schema uses Oa/Shared/Scenario, as independently checked against model.rs and the existing release1 JSON. The expectation was corrected without changing either input, model semantics or server. initial-harness-failure.json and initial-http-results.json retain that harness failure. It was not a product failure; only the later complete run is counted as passing.