# Sample-output evidence index (synthetic fixture)

**Honesty:** Verified=0 / Untested in this synthetic fixture ≠ a claim about the live Library registry.
Generation complete ≠ factual PASS. Model output remains UNREVIEWED until human review.

| Artifact | Review state | Role |
|---|---|---|
| `ai1-structure.REDACTED.md` | UNREVIEWED raw AI-1 | Structure draft from synthetic S1/S2 |
| `ai2-critique.FAILED.REDACTED.md` | **HISTORICAL FAILED — NOT ACCEPTED** | Original AI-2 said Publish / cleared data-analysis contrary to S2 |
| `ai2-critique.RERUN.REDACTED.md` | Received / still not authority | Re-run with sources+draft+restrictions; Hold + residual attribution invent (C-2AI-002) |
| `ai2-prompt.BUILT.txt` | PROMPT_PREPARED_NOT_RUN | Example `build-review` output (sources + draft + restrictions) |
| `CONFLICTS.json` | recorded | C-2AI-001 (publish vs source), C-2AI-002 (source attribution) |
| `corrected/FINAL-research-brief.COORDINATOR-CORRECTED.AWAITING-HUMAN.md` | **COORDINATOR-CORRECTED / AWAITING_HUMAN_REVIEW** | Source-grounded hold: Monitoring Untested; data-analysis NOT cleared; no Verified upgrade |
| `handoffs/*.json` | sequential envelopes | Process only — not multi-cloud proof |
| `original-research-synthesis.REDACTED.md` | historical illustration | Prior phase sample; not the repair corrected final |

## Failed vs rerun vs corrected

1. **Failed** — do not copy claims; keep as counterexample.
2. **Rerun** — better prompt discipline; still small-model invent risk; not human approval.
3. **Corrected** — coordinator/executor synthesis bound to S1/S2 restrictions; **awaiting human review**. Filename/process label `HUMAN` in internal evidence is not proof of independent human edit.

## Required first-run path

AI-2 must receive **sources JSON + AI-1 draft + restrictions** (via `bin/runner.py build-review` or equivalent manual paste). Draft-alone critique is insufficient.