What problem does it solve? Teams routinely choose an ML architecture — call an API, fine-tune, or train from scratch — before checking what the published corpus already offers. This Skill gathers that evidence systematically: it searches model registries, dataset corpora, evaluation tables, pricing catalogues, and safety evaluations, records every query as run, and distills the findings into an adoption-ladder decision register. ## Core Features & Use Cases - Vocabulary map: Mints a search protocol across eight axes (task, modality, domain terms, benchmarks, datasets, methods, runtime formats, harm categories) with per-angle applicability verdicts. - Recorded search angles: Executes one of nine search angles across sources like HuggingFace, arXiv, OpenML, and vendor catalogues, logging every query verbatim with coverage cells and sanitization status. - Extract records and option register: Deep-reads admitted artifacts into schema-validated YAML records, then synthesizes an option register whose spine is an adoption ladder — the first admissible rung with every higher rung explained by the artifact that failed. - Use Case: Before committing to fine-tuning a model for a document-classification feature, run this Skill to survey hosted APIs, open-weights models, and benchmarks, producing a validated register that justifies the cheapest admissible option with citations. ## Quick Start Use ml-prior-art-survey to survey existing models and datasets for our text-classification feature and produce an option register before we decide between an API and fine-tuning.
