Verify a decision

{"actor":"judge","decided_at":"2026-08-27T03:14:12.341Z","decision":"review_revise","investigation_id":"bb0f5270-464b-4b29-bfc6-0040fd189172","new_score":0,"page_slug":"step-finance-ai-agent-over-permission-exploit-january-2026","prev_score":4,"reason":"The page's factual backbone — the theft amount, the executive-device compromise, the shutdown, the Harrap quotes, the laundering route, and the recovery figure — checks out against independent tier-1/2 reporting (12 of 23 claims confirmed). However, the claim that gives this investigation its name and framing — that autonomous AI trading agents with unconstrained transfer authority amplified the breach — is disputed: no primary or forensic source corroborates it, and one of the page's own three cited sources for that claim contradicts it when read directly (claim_findings[10]). Compounding this, the page's own attack-vector section states the breach was 'entirely operational... relying on stolen signing credentials' (claim_findings[9]), which is itself confirmed by independent sources but directly conflicts with the AI-agent section elsewhere on the same page — an internal self-contradiction, not merely an unsupported assertion. A secondary, lower-stakes error was also found in the STEP token's all-time-high date. Given disputed_pct of 0.39 falls in the 30-60% band, and the disputed material concerns the page's central thesis rather than a peripheral detail, this warrants substantial revision and a public flag while the sound factual core is preserved.","score_delta":-25,"sequence_num":3,"submission_content_hash":null,"submission_id":null,"submission_kind":null,"submission_valence":null,"v":1}