Frontier-evals sweep: Petri, Opus 4.6 interpretability audit, NLA, CoT-monitorability, Deployment Safety Hub, EVMbench

[xdotli]

Vetted via 4 parallel research subagents against the CONTRIBUTING.md bar, then spot-checked by fetching each live source myself. All additions are recent (Feb–Jul 2026) and I dropped several subagent recommendations that were either already listed, stale, or redundant with existing entries — see reasoning below.

§10 (Safety / adversarial evaluation)

• Petri (meridianlabs-ai/inspect_petri) — confirmed canonical repo (Anthropic donated the project to Meridian Labs; the older safety-research/petri is now legacy v2). Live: 1.3k★. Verbatim-verified: 38 judging dimensions, 170+ seeds, used in the Claude 4 / Sonnet 4.5 System Cards and by UK AISI.
• Claude Opus 4.6 System Card §6.5 — "Inhibiting internal representations of evaluation awareness": suppresses the model's internal eval-awareness representation via activation oracles/attribution graphs/SAE features and re-runs the alignment evals — a causal audit, not just a transcript read. Confirmed live in the system card TOC.
• Natural Language Autoencoders (Anthropic) — decodes activations to natural language; verbatim numbers confirmed live: 26% eval-awareness on SWE-bench Verified, 16% on a destructive-coding test.
• Evaluating chain-of-thought monitorability (OpenAI) — the primary methodology paper (not a specific system card), open-sourced at github.com/openai/monitorability-evals; g-mean² metric confirmed live on the GPT-5.4/5.6 system card pages.
• Deployment Safety Hub — described structurally per CONTRIBUTING's rule that live-leaderboard numbers get pinned or dropped, not cited loose.

§9 (Agent-specific evaluation)

• EVMbench — confirmed NOT already listed (checked against the existing SWE-Lancer/MLE-bench/PaperBench entries first). 117 vulnerabilities / 40 audits confirmed, independently audited by OpenZeppelin.

Declined / dropped (for your visibility, not included in this diff)

• openai/frontier-evals repo as its own §5a entry — one subagent recommended it, but SWE-Lancer/MLE-bench/PaperBench are already individually listed (that subagent was wrong that they weren't); not worth a bullet just to cross-link.
• GPT-5 "Production Benchmarks" saturation methodology — real but from the original Aug-2025 GPT-5 card, not recent enough on its own.
• Claude Opus 4.5 sabotage evals, Gemini 3 Pro harmful-manipulation, Grok 4 RMF categories — model-card sweep correctly flagged these as better cited via their primary papers (SHADE-Arena, arXiv:2603.25326) or lacking methodological novelty.