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 in, 8 survived the gates (0 Immediate Upgrade, 1 Product Input, 3 Factory Candidate, 4 Watch), 22 dropped. The 8 langchain/latent-space/llamaindex conceptual blog posts are all Tier 3 vendor content restating patterns (agent loops, skills, RAG retrieval choice) this Factory already runs in production — dropped as no-novelty. Of 17 model-watch entries, only Fable 5.1 and DeepSeek V4 Flash are genuinely new-generation models (Watch, pending their auto-eval scorecards); everything else is a batch/preview SKU of an already-superseded generation, and claude-opus-4.8 is flatly dropped per the standing Opus-deprecation policy. Three Vercel changelog items (Konsistent, dry-run deploys, Service Bindings) are Tier 1 primary evidence — Konsistent and dry-run deploys promote to Factory Candidate as low-effort, cross-portfolio reliability/consistency gates; Service Bindings stays Watch pending a named product with the pain point. Simon Willison's shot-scraper video promotes to Factory Candidate as a low-effort self-documentation artifact for UI-facing verification. The Three Wise Men PEP 636 report is Product Input for tax-scanner-api, but its code samples are explicitly 'assumed' rather than read from the real repo, so it's held on a verification first_step rather than scheduled as a refactor. Mistral OCR 4 is Watch — plausible ingestion upgrade, but no incumbent OCR is named and pricing/capability claims are unverified vendor copy.
Proposes replacing nested if/elif rule-dispatch and LLM-JSON-response branching in tax-scanner-api with a decorator-based rule registry (O(1) dispatch, no silent fall-through) and Pydantic-validated `match` statements over tier-1/tier-2 Claude responses.
Why this verdict: Tier 1 in form (direct code-level proposal) but content is explicitly speculative/unverified against real source; novelty Medium (billing-correctness fix, not style); expected gain Medium-High if the silent-fallthrough risk is real (this is a billing engine), effort Medium (real registry refactor) — gain plausibly exceeds effort, but held on verification of the assumed code before any refactor is scheduled.
Vendor-hosted OCR API returns bounding boxes, block classification, and confidence scores per page at a stated $2/1K pages (batch), upgrading raw-text OCR to structured, confidence-scored extraction.
Why this verdict: Tier 3 vendor announcement with unverified pricing/capability claims; novelty Medium (bounding-box+confidence is a real delta over plain-text OCR); expected gain Medium if it beats the incumbent, effort Medium (new vendor integration, no incumbent OCR confirmed) — held at Watch pending the smoke test rather than promoted blind.
New Anthropic model family member priced above Opus 5 ($10/$50 per M vs Opus 5's $5/$25), flagged by model-watch as a lineage-based upgrade candidate for nine products.
Why this verdict: Tier 3 model-watch listing; novelty High (genuinely new model, not a variant); pricing is premium-above-Opus so it is unlikely to clear the efficient-frontier bar without a compelling scorecard — expected gain unclear, effort Low (eval already queued) — Watch until evidence exists.