Hits

Paired [35] <-> [36] correctly as release-then-gate, not two models. Dual-use framing matches raw.

Paired [42] <-> [43] as open-weight push + distillation allegation. That's the real tension, not just a model launch.

Paired [33] <-> [34] checkout-for-fee alongside hourly contact profiling. Good catch on monetization vs surveillance in same actor.

Flagged [45] as thin, [41][39] as single-sourced. Right call — no card/benchmarks for M3.1, prospectus claim and Wiedermann-Moeller claim both need primary docs.

Misses

Cited [26]<->[27] and [28]<->[30] in Connections but omitted both from What happened. If version bumps matter, surface them; if not, drop them. Currently incoherent.

[28] <-> [30] is forced. Reusable envs + Security Cloud scans are same DevDay, not a continuation. Same for [26]<->[27] — trivial patch sequence, no insight.

Buried the lead on safety: [41] IPO liability warning and [39] 100+ org bypass alerts are top-3 agent stories, not Watch footnotes.

Dropped [32] Muse Gadgets ESP32/RPi SDKs entirely, then claimed persistent cross-device environments are the default target. That omission undercuts your own thesis.

Bullish open-weight take cites [49][42][46] but What happened never mentions [49][46]. Also ignores [47] GLM-5.3 on Bedrock running Strix pentest — which directly contradicts the bearish "builders won't get offensive caps" line.

Coverage gaps

Enterprise and sovereign beats are press-release only: [31] Desai hire, [40] $100M academy, [48] ILMUcode. No hiring memo, Anthropic curriculum, or YTL data-residency proof.

For tomorrow: query `site:sec.gov Anthropic prospectus agent liability` + `AWS Bedrock GLM-5.3 Strix` + `Meta Muse Gadgets GitHub ESP32`. Pull primaries, not rewrites.

One change for tomorrow

Enforce recall lock: every ID cited in Connections / Where this points must appear in What happened. No orphan citations.