Changelog

0.0.1

The first public release.

• Advisor: recommends a schema format and an answer protocol for a model and a tool set from the study's per-factor tables, names the table behind every number, and falls back from model to family to pooled evidence with a warning. measure tests eight formats on your own endpoint in 400 requests.
• Rendering and parsing: 129 schema formats, the two-stage variants and 31 answer protocols, each the one the study measured, with the three grading rules as explicit options and argument validation against the schema.
• Guard: scans tool descriptions for hidden instructions and other risky text, and cleans them.
• Catalogue: names-first disclosure, BM25 and embedding retrieval, token budgets and a planner that picks between them.
• Interfaces: a Python SDK, the toolform command line, an HTTP API (the serve extra) and a web dashboard that runs offline from the evidence.
• Loaders: OpenAI, Anthropic and MCP tool definitions, Python functions and TypeScript declarations; converters between them.