Module learning_foundry 

User-directed, evidence-first background learning.

A learning track is intentionally separate from a generic scheduled task: it has source boundaries and budgets, produces an inert candidate, and never grants the research run authority to change live skills.

Functions

compute_next_learning_run

evaluate_learning_candidate

Run paired, no-tool trials for a Learning Foundry candidate. Both arms use the same model and scenario; only the candidate arm receives the proposed skill. The harness has no filesystem, network, credentials, or tool schema.

promote_learning_candidate

Materialize a reviewed external-learning candidate through the same version registry used by other agent-created skills. Existing skill names are rejected: improvements to an active skill must use comparative reflection and shadow evidence instead of bypassing its baseline.

spawn_learning_foundry