Learning and weights

v5.5 adds a safe, local-only learning layer on top of the ranker.

Commands:

sigmap learn --good src/auth/service.js
sigmap learn --bad src/legacy/old-api.js
sigmap learn --good src/auth/service.js --bad src/legacy/old-api.js
sigmap weights
sigmap weights --json
sigmap learn --reset

What it does

• boosts files that repeatedly help answer the task
• penalizes files that keep surfacing but mislead the answer
• stores the learned multipliers in .context/weights.json
• keeps the feedback local to the repo

Why it is safe

• no networked learning
• no cross-repo sharing
• reset is one command away
• baseline behavior is still the normal ranker when no learned weights exist

Reading the weights

sigmap weights prints the current learned multipliers in a human-readable table. --json returns the exact stored object.

Use it when you want to answer:

• Which files has the ranker learned to trust more?
• Did judge --learn actually change anything?
• Do I need to reset after a bad experiment?

Best workflow

1. Start with ask
2. Check coverage with validate if needed
3. Score the answer with judge
4. Reinforce or penalize manually with learn, or use judge --learn when you want opt-in automation