selection

Overview ¶

Package selection implements the closed-world tool-selection eval harness (docs/evals/PLAN.md §4–§6). The real ./evaluate binary is offered to a model alongside harness-side decoy tools; every tool call is graded against acceptable/forbidden outcomes fixed when the scenario was authored. The decoys are eval-harness tools, not product tools: they never leak into the MCP server or its production schema.