pp. 148–149 is the one. From May 2025 to April 2026, all human-feedback vendor traffic ran without blocking bio classifiers: ~50k people, ~133M exchanges, vendor-vetted only, mostly open-ended access. An internal-use flag disabled both blocking and the logging of classifier flags, so no flags reached review for a year. The retrospective pass came back clean, 62 non-red-team transcripts flagged, no clearly concerning misuse on manual review and no customer impact. Anthropic raised its CB-1 risk assessment on the back of it. What stays is the conclusion: this raises the likelihood of similar issues they haven't found.