Inference pilots are the mechanic for bounded local model experiments, promotion loops, benchmark evidence, and adopted local-worker paths.
Use this package when changing llama.cpp, Qwen, OVMS, LangGraph, local trial, benchmark, model card, or winner-promotion surfaces.
- local inference runtime wrappers
- public-safe trial and benchmark contracts
- model profile and card routing
- runtime winner-promotion posture
- bounded local-worker deployment support
- resource-gated Tree of Sophia OCR, structure, translation, semantic, retrieval, graph, and golden-kernel transfer experiments
Model files, hardware behavior, upstream inference engines, and owner-authored
evaluation truth remain outside this repository. aoa-evals owns portable eval
truth when a claim becomes evaluation doctrine.
Machine-fit records, model cards, compose tuning overlays, public-safe trial packets, benchmark results, and operator-selected profiles.
Pilot commands, benchmark indexes, bounded worker routes, promotion candidates, and runtime evidence refs.
- a model is generally best from one local run
- a pilot is production-ready without the promoted live check
- benchmark evidence replaces eval-owner truth
- a source card proves live endpoint availability
Run the commands in AGENTS.md.
Use machine-fit for host capability and governed-execution when a pilot becomes a reviewable local-worker path.
Current source surfaces stay in package-local parts/ routes, root
scripts/ wrappers, compose/tuning/,
mechanics/machine-fit/parts/inference-tuning/docs/model-cards/, package
benchmark surfaces under mechanics/inference-pilots/parts/local-trials/, the
Tree of Sophia A/B/C suite under
mechanics/inference-pilots/parts/tos-foundation-lab/, and package tests under
mechanics/inference-pilots/parts/. Archived pilot
surfaces now stay under legacy/ with quiet root bridge commands for operator
compatibility.