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.
30 items reviewed for 2026-W42. 12 promoted (1 Immediate Factory Upgrade: Supabase legacy JWT key deprecation with a hard late-2026 deadline across every Supabase-backed product; 6 Factory Candidate: board-review prompt posture, Sonnet 5.5 scorecard follow-through, Haiku 5.5 close-out verification, reasoning-effort token-proxy calibration, Supabase --use-api Docker-free deploys, Grok Build open-source harness read; 2 Product Input: entity-level PDF segmentation for floodstream/finalreview, LiteParse v2.0 drop-in parser; 3 Watch: agentjacking-via-bug-reports security pattern, filesystem-vs-RAG retrieval heuristic, and a probable triage content-mismatch anomaly worth monitoring). 18 dropped, mostly Tier 3 generic harness/agent blog posts restating doctrine already codified in CLAUDE.md/AGENTS.md (safe-outputs separation, bounded-probe testing, AGENTS.md review rigor), duplicate PDF-parsing posts from the same vendor series, one Fable-related item excluded by standing policy, and one routine model-catalog entry with no named incumbent.
Segmenting documents at the entity level before LLM extraction (rather than whole-document chunking) avoids U-shaped positional bias that otherwise drops up to 80% of table rows in multi-page PDFs.
Why this verdict: Tier 3 vendor blog but describes a specific, falsifiable failure mode directly material to named pipelines; novelty Medium; effort Medium to test, gain potentially High if confirmed — worth a bounded test, not blind adoption.
LiteParse v2.0 is an LLM-free PDF/doc parser now distributed for Python/WASM/Rust, offered as a drop-in replacement for current doc-ingestion parsing.
Why this verdict: Tier 3 vendor blog, but a verifiable open-source release naming concrete products already in this pipeline category; deferred twice by cap already (rule-of-two friction), due for a verification pass; effort Low, gain Medium if it reduces LLM-parsing cost.
Attackers file fake/malicious bug reports containing prompt-injection payloads that autonomous coding agents ingest as instructions when triaging issues, causing unintended actions ("agentjacking").
Why this verdict: Tier 3 blog, no linked primary evidence; novelty Medium (names a real attack class); materiality is real but CLAUDE.md already requires flagging suspected injection, so expected gain is Low against building new tooling — Watch pending verified evidence.
The item's 'why it matters' text (an agent QA pattern involving Fable vs GPT-5.5) does not match its title (sqlite-utils 4.0 schema migrations) — likely a triage content-attachment mismatch in the intake pipeline itself.
Why this verdict: Tier 3 and internally inconsistent — the actual sqlite-utils release has no stated relevance, and the attached rationale describes an unrelated, banned-model story; not promotable on the merits, but the mismatch is worth flagging as a possible intake defect.