Mapped authority drift: pinned local trial

Inspection: 2026-09-23. Relates #297 and implementation #298. This records a real local CLI trial, not a native agent rollout or merge verdict.

Scope and implementation boundary

The existing adapter reads only .agent-harness/tier.json and the exact mapped root CLAUDE.md clause. Its pipeline remains local bytes, explicit mapping, requested/inspected/unknown counts, then an advisory report. No new scanner, parser framework, telemetry collector, workflow or policy authority is added. The correction is a separately reviewed prose diff, not a probe-side auto-repair.

Snapshot and executable identity

The supplied archive identifies revision 4ed6561ceabca623e556e8bca6c0574b9d555c3f. Reconstructing its Git tree from all 130 tracked files, including Git file modes, produced b391e9a5bf0df5b5258e6be95c136de76d3714c5, exactly matching GitHub’s pinned commit. This checks repository content against that retrieved identity; it is not a signature verification or a claim about the user’s installed configuration.

The executable was the drift probe at 5f6b8f8da02ededdccbac0a944e3fded848209c0, Git blob ea81677e96dd519a146ae752c82e0fd699456e7a. Its formatting-only changes preserve the original probe’s Python AST. Environment: Linux, Python 3.13.5. Each observation below used a real subprocess:

python /path/to/checker/scripts/authority_doc_drift.py /path/to/snapshot

These are caller-selected local paths, not commands extracted from repository prose. No network request, installation, hook activation or source-file write was performed by the probe. The original snapshot’s 130 tracked-file SHA-256 values were compared before and after the trial and were unchanged.

Observations

All four subprocesses exited 0 with no stderr. Every report retained merge_verdict: null, revision_verified: false and advisory authority. The baseline was repeated and its complete parsed report was identical. The concrete discrepancy key was authority-doc-v1:merge:free:gated. The probe did not establish the Git revision itself; the snapshot mapping above was checked separately by the caller.

Exact input digests

SHA-256 over the bytes supplied to the probe:

Review and reconciliation

The canonical declaration was inspected directly before proposing the correction; it already sets both push and merge to free. This PR changes neither declaration nor permissions, protection settings, review requirements, or CI. Free authority does not mean that unreviewed or failing work is safe to merge.

GitHub comparison of the archive revision with main fcea0eb3fc129d5422250133a50e1d1d47c1dba0 showed only the two #288 documentation changes, outside this mapping. Before publication, the live root guidance was retrieved again and still matched original Git blob dd4a6c3f633fe78812c16070d0af9e1e8db58ffd. The proposed corrected blob is a007ea1a71786fab81033de858ea283ffac70998; its 149-line count remains within the existing 150-line budget.

Limits and remaining acceptance

This trial covers exactly two explicitly mapped fields in one repository plus controls. It does not establish estate-wide precision, recall, runtime policy resolution, useful automation savings, or native Windows/macOS qualification. The 12 synthetic probe tests also pass, but are separate from this real-input trial. Both the implementation and this evidence/correction still need their applicable published-head CI and review before #297 is closed. No new mapping or quality gate is authorized by these observations.