# E91 ClawX Evidence Report

**Status: PUBLIC — PARTIAL EVIDENCE.** The signed E91 bundle is published on both hosts and was verified from a clean download. It is **not** an independent witness and **not** a ClawX API mission record.

## Public locations

- **ClawX:** https://clawx.click/evidence/e91-vertical-intelligence-2026-11-18/
- **CAIN Studio mirror:** https://cainstudio.online/proof/bundle/e91-vertical-intelligence-2026-11-18/
- Live checks: `artifacts/e91/live_publication.json` (both HTTP 200).
- Offline verifier: `verify_e91_offline.py` in the bundle, stdlib + cryptography, imports no CAIN code.
- Clean-download verification: **91/91 checks passed**, tamper test detected.
- Signer: self-generated E91 key; **NOT** an independent witness.
- Refused files: none; no source code, task secrets, hidden ground truth or private keys are published.

## What the bundle contains

- Raw held-out rows for arms A–G and the team run.
- The core analysis recomputed from those rows.
- Transfer, learning curve, retention and poisoning results.
- Causal reasoning suite, security lab, mission reports and kill-switch evidence.
- The 120-task E91 evaluation-slice report and raw rows.
- Every E91 report, the acceptance matrix, limitations and next-evolution recommendation.
- The shared paid-call ledger (tokens/cost/timing only; no prompt or output text).

## What a pass proves

The manifest signature and every file digest match; mission arithmetic is internally consistent (`verified == effects`, 0 duplicates, chain OK, 0 post-fence effects); security-lab counts match the per-scenario rows. That is **integrity and internal arithmetic**, not proof that the experiments were honestly conducted.

## What is still local

E91 missions are not written into the real ClawX API data model, tenant registry or mission service. This is a static signed evidence publication, not a ClawX API integration. That gap is recorded in the acceptance matrix (gate 19 explanation).
