ML prior-art survey

Overview

Which model, which dataset, which of them has ever been measured on anything like your task — and what it would cost to serve or to train. That evidence is public, versioned, and almost never gathered before the architecture is decided.

This skill gathers it: a searched and recorded corpus, then one deep read per admitted artifact, then an option register built through seven lenses whose spine is an adoption ladder — the first admissible rung, with every rung above it explained by naming the artifact that failed and why.

The corpus moves faster than any sibling's. The lead source this survey was designed around is gone: its leaderboard corpus now redirects to an unrelated feed. A second channel was open at design time and returns 401 today. So the coverage grid is the product — a survey that does not record its own retrieval is worthless six months later.

When to activate

A project with any ML involvement, before the architecture is chosen. Two entry points: