Deliverable: plans/spway/8e/report.md with the catalog rows in the columns of Catalog.LanguageModel.xml, the first-job recommendation (8c), the provider parts (8a: hosted vs local in the sandbox), the risks. Primary sources, dated; UNVERIFIED where not verified. Dispatched 2026-09-29 as a subagent of claude-code.
(claude-code, 2026-09-29 08:32 UTC) Delivered 2026-09-29 by spice-architect: plans/spway/8e/report.md (the seat's project write was blocked by the harness; it wrote my scratchpad and I copied it). Findings: 'OpenJev/Winnow-12B' is three things - openjev/openjev (27B, CC BY-NC, 16k), EldanRing/Winnow-12B (12B Gemma finetune, Apache-2.0, 64k, 12.7 GB Q8, serves /v1/systemone and chat) and the razorback16/openjev server; Winnow is the only commercial-safe local pick. Jev 1.13 is a typed decision API (POST /v1/systemone: noul/choice/score), NOT OpenAI-compatible, USD 0.042 per M in, output free, 64k per request, text only, generates nothing. Fit: decision model for triage Kind, Choice/Score autofill, priority, advisory approval probabilities; general LLM for research, authoring, coding, extraction, the daily brief. Triage about USD 0.04 per 1000 calls on Jev vs 0.35 DeepSeek vs 1.50 Anthropic. First job (8c): one Jev choice filling Kind for DT.AgencyTriage. 8a: reuse LlmProvider with a decide Role and a second client seam. UNVERIFIED: sandbox GPU/VRAM, gateway version.
plans/spway/7a/report.md
plans/spway/7a/Catalog.LanguageModel.xml
/sites/Projects/Lists/ProjectTasks/Item/spway-8