Every week our pipeline scrapes the model catalogs and vendor blogs, then judges each item against one question: would this actually improve something we run in production? Verdicts below. Watch = interesting but unproven claims · Adopt = earned a place · Ignore = noise · Product Input = a feature idea, not a factory change.
W29 surfaces two P0 incidents detected on review day: claude-sonnet-4.6 and claude-haiku-4.5 aliases were both re-pointed on 2026-07-19, covering 10 app roles including quincy (locked-consumer), my-day, lengua, and all three student-facing apps. Both require immediate OpenRouter snapshot verification and regression evals before the next production invocation. A third alias re-point (claude-opus-4.7, UNVERIFIED) is Factory Candidate: grep app configs first to confirm incumbency. The Supabase Data API breaking change (explicit GRANT now required for all new tables) is Factory Candidate: update the supabase skill and migration templates before the next Supabase-backed scaffolding runs. Four Watch items advance: claude-sonnet-5 and claude-opus-4.8 await eval scorecards; Cloudflare Vite Plugin 1.0 is actionable for any CF Workers + Vite app (eligibility unconfirmed); Supabase June MCP plugin promotes to Product Input once authentication is established. 22 of 30 items dropped: all four ByteBytego newsletters (Tier 3), both LlamaIndex vendor posts (Tier 3), the CF blog homepage scrape, the CF data platform case study (gain does not exceed effort), two additive-only price changes, six models with no eval or no incumbent, and four non-incumbent deprecations.