# E-017 — Measuring a transition does not yet prove the guard's effect

- Date: 2026-08-03
- Status: reproducible dogfooding confirmed; causal efficacy not yet demonstrated.

## Objective and trajectory

Round `R-HOOK-INTERVENTION` sought to turn hook activity into intervention evidence: identify the new finding, record its later resolution, and measure the latency between the two per harness.

The Claude and Codex hooks now record only hash identifiers for new and resolved findings. The inventory joins them by repository and identifier and exposes `detected_by`, `resolved_by`, and latency. The E2E seeded `weak-test`, detected it through both harnesses, and then fixed it through Claude; the resolution preserves the original fingerprint.

## Evidence and verdict

- `test/hooks-e2e.test.mjs`: reproducible detect → resolve sequence with Claude and Codex.
- `test/global-inventory.test.mjs`: cross-harness association and deterministic 2,000 ms latency.
- `oguard inventory report --json`: exposes transitions without paths, content, or session identifiers.
- `round-end`: `PASS` with `AC-HOOK-INTERVENTION` proven by the suite and the documentation validation.

The result proves traceability of the transition and that the guard can observe a later fix. It does not prove that the warning caused the fix, that the agent would have followed a worse trajectory without the guard, or that the latency generalizes: the current organic inventory contains only four `pass` runs and still zero observed transitions.

## Impact on the thesis

The metric is a prerequisite for the Phase 5 control/placebo/guard trial, not a substitute for the trial. Upcoming observations must label organic findings and compare resolution rate, resolution time, and false positives across conditions before claiming cost reduction or better direction.