Synthesis history

No new substantive claims this pass. All new items (31326, 31342, 31343, 31809, 31810, 31811) are tweets and secondary press/blog coverage of the OpenAI Deployment Simulation paper already incorporated in the previous s…

New items this pass are additional coverage of the OpenAI Deployment Simulation paper (PDF, press articles, tweets) already incorporated in the previous synthesis — none contain parseable claims or introduce new substan…

OpenAI published Deployment Simulation (June 16) [^30352][^30345][^30427], a methodology that directly addresses eval-awareness distortion — the most concrete methodological counter to one of the thread's central evalua…

The substantive addition this pass is the no-CoT time horizon research [^29916], which finds GPT-5.5 can complete ~3 minutes of human-equivalent reasoning without chain-of-thought and that no-CoT capability is doubling …

Two substantive findings this pass. Google DeepMind researcher Josh Engels found that SFT-only versions of Gemini 3.1 Pro and Gemini 3 Flash match full production models on safety benchmarks, concluding Gemini's safety …

Two substantive new items this pass. Model diffing agents research [^28093] introduces a complementary evaluation methodology and explicitly names a structural limitation of current frameworks: they can only detect what…

Three substantive new items this pass. Google DeepMind Language Model Interpretability research [^28030] adds a directional finding to the eval-awareness picture: eval-awareness does not uniformly improve behavior — mod…

Three substantive new items this pass. ARC researcher Mikewins published a detailed technical roadmap [^27645] explaining that training-process monitoring is necessary because cryptographic arguments suggest finished mo…

Two substantive new items this pass. Google retroactively asked 404 Media to remove the phrase 'it's critical that we maintain humans in the loop' from a published official statement [^24405], adding a new tension betwe…

ARC's white-box estimation challenge [^23366] is the substantive new development: it introduces a technical argument that black-box behavioral sampling is structurally insufficient to detect control-undermining behavior…

The most significant new development is that OpenAI's playbook reportedly contains a claim that AI capabilities may not be fully evaluable by third parties [^23226], sharpening the structural conflict-of-interest critiq…

Emergence AI's comparative simulation study adds striking empirical evidence that frontier model alignment quality varies dramatically across labs — not just across conditions — with Claude producing zero crimes and Gro…

GovAI enters as a new academic-policy voice with a framework for rigorous third-party frontier AI auditing, reinforcing AVERI's practitioner position with institutional research weight. A technical cryptographic verific…

Three substantive new voices entered this cycle: Anthropic (via 'Teaching Claude Why' and the Feb 2026 Risk Report), AVERI as an independent auditing organization, and EU regulators via AI Act compliance analysis. The c…

Two frontier AI labs published complementary contributions to AI safety evaluation on the same day, May 29, 2026. OpenAI released a methodological playbook intended to standardize how third parties evaluate frontier mod…