AAEL Phase-1 gate, made measurable (Constitution §4 rung 9 / §14 row 1): Brier + ECE over the logged (prediction → independent-verdict) pairs in the sm_records ledger, plus the per-bin reliability map (the persisted-as-reference calibration map). Reads the ledger; never fires the model. n=0 until rows accrue.

qtype selects which prediction class to score: "route" = execution/research rows (world:spine / world:web verdicts); "answer" = the proposer+grader learn-loop rows (verdict_source=grader:<model>), where the grader's pass/fail IS the actual and the proposer's stated confidence is the forecast — so the learn loop's grades show up here.