Does the prop scoreboard grade a line against the week that line was for? Three premises: props.settle keys actuals by player only; the arm_prediction prop ledger is never filled; model_vs_market loses lines to a raw-string name join.

Owner lane: calibration

Question

Does the prop scoreboard grade a line against the week that line was for? Three premises: props.settle keys actuals by player only; the arm_prediction prop ledger is never filled; model_vs_market loses lines to a raw-string name join.

Prediction

Frozen in prereg/2026-09-23T1118Z: M1 wk2 0 disagreements; wk1 actual differs >=85%, outcome flips 30-45%. M2 0 production callers, all 3,129 rows NULL, order B fills every wk1 entity from wk2 truth. M3 4-9% lost to raw join.

Finding and verdict

TESTED -- premises 1 and 2 CONFIRMED, premise 3 CONFIRMED AT HALF THE FILED SIZE. (1) Called on today's board, props.settle grades 1,111 week-1 lines against week-2 stats: 88.1% get the wrong actual and 47.3% (526) get the wrong over/under, so any hit rate, reliability curve or the item-1.78 bet gate (tests/test_prop_bet.py step 4) computed through it after week 2 is scoring a near-random label on every week-1 line; 33 more lines are settled for men who did not play that week and should void. (2) The prop_probability ledger has 3,129 rows, 0 graded, week NULL on all, and no production caller of arms.settle; called wk2-first it would fill 982 of 1,039 SBQ entities from the wrong week and mis-score 499. Called in week order it is right except 14 DNP rows. The ledger also stopped at week 1 (built 2026-09-13): no week-2 prop probability was ever stored. (3) The raw-string join leaves 3.2-3.5% of lines unpriced that identity resolution would price (96 lines over two weeks), not 6.5%. Consequence for this lane: no prop-probability calibration measured through either settle path after week 2 is valid; prop P(over) calibration in 2026 has not been measured at all.

Reasoning

The calibration lane cannot grade whether prop P(over) is honest until settlement uses the line's own week. This is the prerequisite RANKED.md #19 names for H-180..H-185.

Next action

Proposed fix (for a reviewer, not the change queue): settle by (player_id, event week) with the week derived from the event's kickoff, void on DNP; derive arms.settle's week per entity from its event rather than the argument, and schedule it; key model_vs_market through identity.match_name. Then the calibration question proper: reliability of stored prop P(over) on wk1-2 (~2,800 lines) by quantity, against the market's devigged P as the free baseline.