# Agent Action Preflight / Red Team v0.1 — public provenance

**Study date:** 15 September 2026. **Type:** internal, synthetic laboratory study. **Status:** public summary with selected aggregate source files; not a client evaluation.

The published files are read-only copies from the study package. Their SHA-256 digests are:

| File | SHA-256 |
|---|---|
| `agent-action-lab-v01-report.md` | `0D23B2061C4EA1C5B42CB811C989CE121B4A3600B2454FD534A80622C986D024` |
| `agent-action-lab-v01-fixture-summary.json` | `E1434E192D2E332EB268906956BA74C62F9015245069CAC6DD00EE9E1186301D` |
| `agent-action-lab-v01-verification.json` | `7C11B94F4B92FE26916292BB6B771F708B68F7BBBCC62C6CB63A93A4D71EFE31` |

The source report is in Russian. The English Evidence page summarizes its material method, observations, and limitations. The full raw event and replay package is not published here, so this website alone does not support independent replay of every result.

The 93 scenarios were author-designed. The 72 dangerous attempts per mode are a selected fixture cohort, not a sample of natural model errors. The full deterministic comparator uses the same contract logic as ordinary code. No real payment, customer CRM, or public publication was exercised. A separate Gemini batch produced no dangerous proposals; a comparative dangerous-action rate could not be measured in that batch.