Research-0114: Model overview tiny-AI documentation closeout¶
- Status: Active
- Workstream: docs-only closeout for shipped tiny-AI model docs
- Last updated: 2026-05-14
Question¶
Do the human-facing model overview and FAQ still claim that tiny-AI model weights are future work after model/tiny/registry.json already ships production and smoke ONNX artefacts?
Sources¶
docs/models/overview.md— stale "Tiny-AI models (planned)" section.docs/reference/faq.md— stale FAQ answer saying no first-milestone weights ship yet.model/tiny/registry.json— current source of truth for shipped tiny-AI ONNX entries, theirkind, and theirsmokeproduction/CI role.docs/ai/overview.md,docs/ai/inference.md, anddocs/ai/model-registry.md— canonical tiny-AI operator docs.
Findings¶
The stale text was contradicted by the registry. Current master ships production entries for VMAF-tiny regressors, codec-aware FR regressors, LPIPS-SqueezeNet, saliency students, TransNet V2, and learned filters, plus explicit smoke-only rows for CI and compatibility. The public model overview should point readers at those families without turning into a second registry, and the FAQ should show a real invocation using a shipped model ID rather than a future vmaf_tiny_vN placeholder.
Alternatives explored¶
Duplicating every registry row into docs/models/overview.md was rejected: it would drift again as soon as a model is added or promoted. The chosen shape keeps a family-level table in the overview and directs operators to model/tiny/registry.json plus per-model cards for authoritative details.
Adding a new ADR was rejected because this PR does not introduce a new model policy or runtime surface; it corrects stale public prose for artefacts that are already shipped and documented elsewhere.
Open questions¶
- None for this closeout.