Delivered 2026-09-27 by spice-sp-fidelity-researcher: plans/spway/7a/Catalog.LanguageModel.xml + report.md (to be moved into Config/Specs by SPWAY.7b). Findings: SPICE already tracks models in two places - LlmProvider parts (what we route to) and the LLM Radar (a release-discovery funnel with its own generated fields) - so the catalog category REUSES them: ContextWindow, PriceInPerMTok, PriceOutPerMTok, ReleasedAt, License, ModelId are Existing=true; Vendor is inherited from CatalogProduct. 9 Features from the Hugging Face model-card YAML spec, 15 Extensions from the EU AI Act Annex XI/XII and the OpenAI-compatible API convention; buyer Questions: job kind (Extraction/Research/Triage/Drafting/Code), budget per M tokens, local-only or hosted, max latency, data residency, minimum context, available VRAM, substitute allowed. Seed rows named with sources: jev-1.13, OpenJev, Winnow-12B, Claude Opus 5.5, Qwen3-Coder-30B-A3B, Pollinations openai-fast (GPT-OSS 20B); smart-local is config, not a catalog row. Two engine gaps found (filed as SPWAY.7d and SPWAY.8d).
(claude-code, 2026-09-29 08:33 UTC) Correction from 8e (2026-09-29): the Substitute row named openjev/openjev, which is CC BY-NC (non-commercial); the commercial-safe local substitute is EldanRing/Winnow-12B (Apache-2.0, 65536 context, 12.7 GB Q8). Also Jev's per-request cap is 64k (32k for state + longest question), so 16,384 is a quarter, not half. 7b applies both.