/calibrate

/calibrate

Feed outcome data back into the model. Updates quality weights, engagement parameters, and optionally skill/agent prompts.

Pipeline Position

• Requires: collect complete (always available once pipeline started)
• Produces: updated math module parameters, weight-history.jsonl append
• Does not advance stages — calibrate is a continuous loop, not a gate

Instructions

1. Determine target: math (default), experience, or framework.
2. Load feedback events from $PROBLEMS_ROOT/$ARGUMENTS/feedback/events/.
3. Load current weight history from feedback/weight-history.jsonl.

--target math (default)

• Load $MATH_ROOT/optimization/05-quality-score.md for current weights.
• Load ${CLAUDE_PLUGIN_ROOT}/templates/FEEDBACK_TEMPLATE.md for ladder structure.
• Process L0 events (post-match ratings) → short loop: Bayesian update on w_k.
• Process L1 events (7-day rebooking) → medium loop: engagement model update.
• Process L2 events (30-day active) → long loop: flag for human review if thresholds shift.
• Check ladder correlations. If any break → trigger ladder break protocol.
• Append to weight-history.jsonl.

--target experience

• Analyze feedback events for experience layer signals (cancellation rates, nudge response, etc.)
• Suggest updates to experience specs based on observed patterns.

--target framework

• Evaluate pipeline output quality (did specs make sense? did scaffold build?)
• Suggest updates to skill/agent prompts based on what worked vs. didn't.
• Write suggestions — don't auto-modify skills without human approval.

4. Update pipeline-state.json calibrate timestamp.