Why this is in the vault
Daily AlphaSignal curation leading with Anthropic shipping a pre-built "artifacts gallery" for Claude for Legal (six ready-to-run workflows) and Liquid AI's six sub-1GB on-device models — paired with an Artificial Analysis v5 benchmark refresh (Claude Opus 5.5 leads at 58) and a BabelArena finding that agent tool-call cost varies wildly by language. The Claude for Legal item is a direct follow-on to this vault's 2026-05-14 entry on the original open-source plugin launch, now evolved into vendor/firm-contributed workflow packaging — the clearest RDCO-relevant thread in the issue.
Mapping against Ray Data Co
Concrete connection: Claude for Legal's shift from the May 2026 open-source plugin suite ([[2026-05-14-alphasignal-ruview-anthropic-legal-nous-training]]) to an in-product "artifacts gallery" — pick a workflow, load documents, review Claude's first pass — with named firms (Goodwin, Foley & Lardner) contributing their own artifacts and 60+ connectors to tools like Bloomberg Law and Docusign, is the exact pattern project_credibility_for_phdata_sales is betting on: domain credibility compounds through vertical workflow packaging on top of a general model, not through raw model capability alone. It's also a live example of the "agents in production" framing the founder uses to measure pipeline, not subs — this is Anthropic doing it at the vendor layer (firms packaging their own review playbooks as reusable Claude artifacts) rather than the model layer. Second, weaker connection: Artificial Analysis's v5 index has Claude Opus 5.5 leading at 58 against GPT-6 Astra's 57, with the open-source gap widening to 12 points — consistent with, not a change to, the standing read from the 2026-10-27 price-compression note ([[2026-10-27-alphasignal-gpt61-terra-pricing]]) that cheap-tier price wars aren't closing the frontier-coding ceiling gap the Anthropic-credibility bet rests on.
Curation section
- Anthropic ships Claude for Legal artifacts gallery (24,511 likes) — six pre-built interactive workflows inside Claude: Cite Check (flags misquotes/wrong citations), Tabular Contract Review (answers questions across a contract stack with sourced quotes), Regulatory Scan (monitors and routes new legal developments), plus Playbook Deviation, Jurisdiction Matrix, and Matter Board. 60+ connectors (Bloomberg Law, Docusign, Filevine); Goodwin and Foley & Lardner contributed their own artifacts. See mapping above.
- Liquid AI ships LFM3, six open-weight 2B-class models for on-device use (8,447 likes) — base model LFM3-1.7B beats the prior 2.6B model with 35% fewer parameters, weights fit in ~1GB, 256K-token context. Family spans text/agentic tool-use, vision/video, 18-language audio, an omni model, and two structured-output "decision" models. Fine-tunable on a single GPU; runs on Ollama/vLLM/llama.cpp. Not individually mapped — edge/on-device deployment isn't a current RDCO surface.
- Artificial Analysis drops Intelligence Index v5, 60% private test sets (8,128 likes) — direct response to benchmark contamination. New benchmarks: FrontierCode 2.0 (132 real repo tasks, binary merge/no-merge grading) and Terminal-Bench-Science (70 terminal-based science-agent tasks, all-or-nothing). Claude Opus 5.5 leads at 58, GPT-6 Astra at 57; open-source models now trail by 12 points (up from 3); GPT-6.1 Sol scores within 4 points of Opus 5.5 at ~1/8th the cost. See mapping above.
- Signals (6 items, lower signal, not individually mapped): NVIDIA open-weight speech model tops transcription/speaker-ID benchmarks; Qwen open-sources BabelArena, finding agents need up to 8x more tool calls in Telugu than English for the same output (notable cost-variance data point, not deep-fetched); new benchmark probes LLM grasp of hidden/implied meaning; a Nanjing University Hearthstone bot beats the server's top-10 ladder players; REA releases 20 open-source repos for an AI reverse-engineering studio.
⚠️ Sponsorship
Three resolved sponsor placements this issue: Fireworks.ai (native ad block promoting its "Forge" event, Nov 3 virtual session with NVIDIA/Microsoft speakers), Datadog (sponsored ebook, "7 lessons from Google's Director of AI," native-in-list), and WorkOS (native ad inside the numbered Signals list, item #2 — a one-command auth-install pitch, not an independent news item). Additionally, the newsletter masthead carries an unlabeled "In Partnership with" slot pointing to a generic alt="partner_image" logo with no resolvable sponsor name in the HTML — present but unresolved; not guessing at an identity. None of the four placements touch the Claude for Legal or Liquid AI items used in the mapping above, so the editorial content driving this note's RDCO read is unsponsored.
Related
- [[2026-05-14-alphasignal-ruview-anthropic-legal-nous-training]]
- [[2026-10-27-alphasignal-gpt61-terra-pricing]]
- [[project_credibility_for_phdata_sales]]
- [[project_phdata_cert_escalator_path]]