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 judged, 10 kept / 20 dropped. Two genuine Factory Candidates: Sonnet 5.5 and Haiku 5.5 generational bumps, both already auto-eval-queued across 9 products - action is just reading the scorecard. One security Factory Candidate: fake-bug-report prompt injection (agentjacking), with a concrete dry-run test against the triage skill. Six items held at Watch pending a cheap verification step (Workers Cache, LiteParse v2 canonical rep, skills-vs-MCP framework, VLM-OCR postmortem, thinking-efficiency benchmark, Replit eval-loop pattern, Vercel CLI project-update). Dropped the bulk of a 15-post LlamaIndex blog backlog as duplicate/no-novelty vendor marketing (3 LiteParse posts deduped to one canonical Watch), two Cloudflare tag-archive index pages as unverifiable, an image-only model as explicit catalog churn, and two items (file-as-context pattern, cross-model review pattern) as already-implemented Factory doctrine.
Attackers embed prompt-injection payloads inside fake bug reports/issue text; an autonomous coding agent that treats that text as trusted context can be steered into unintended actions (agentjacking).
Why this verdict: Tier 3 source but the mechanism is concrete and matches a real exposure class Neill's autonomous agents already have; the verification gate is the explicit check tier 3 requires before promotion. Gain High (prevents a real incident class) vs effort Low (one review pass), so gain exceeds effort.
New Sonnet generation (1M context, cheaper cache-write/read tiers) queued for auto-eval against the current Sonnet incumbent across 9 named products.
Why this verdict: Tier 1 provider registry data, a real generational event (not a SKU variant), matching the doctrine's efficient-frontier bar. Effort Low because eval is already automated/queued; gain potentially High across 9 products.
New Haiku generation with lower per-token pricing and a volume pricing override, queued for auto-eval as a cheap-tier rotation candidate across 9 named products.
Why this verdict: Tier 1 registry data, real generational bump (not a variant SKU), matches the efficient-frontier bar for the cheap tier. Effort Low (already queued), gain Medium-High given 9 dependent products.
LiteParse v2 is a model-free, dependency-light PDF/doc parser (Python/WASM/Rust) claiming fast deterministic text extraction without an LLM call.
Why this verdict: Tier 3 vendor blog (LlamaIndex), repeated across 4 near-identical posts this week (canonical kept, others deduped). Real potential gain if genuinely drop-in, but unverified until run on real documents, so held at Watch.
Proposes completion-token count as a proxy for chain-of-thought effort, and benchmarks a 1.5-4x token-efficiency gap between open and closed reasoning models.
Why this verdict: Tier 2/3 research post, informs future routing but no immediate decision is pending; cheap to file as Watch, not worth action yet.