int_lap_thermal_proxy

Cumulative push load from tyre thermal management.

Part of the Physics family.

Cumulative push load from tyre thermal management. One row per lap. push_residual is the lap’s pace against a POINT-IN-TIME stint baseline — the expanding median of this stint’s valid laps strictly before the scored lap (macro: trailing_median) — and the two EW loads are backward sums over it. Rebuilt by 08e: the baseline was a MEDIAN over laps 1..CEIL(stint_length*0.60) per stint, which both pooled laps the scored lap had not yet run (55.92% of mart rows, mean 4.65 laps) and took its window width from stint_length_actual, the stint-life target’s own numerator. That contamination did not merely add noise — it inverted push_residual’s sign against two of the three targets.

Lineage

Overview

Property	Value
Layer	intermediate
Family	Physics
Contract enforced	No
Upstream	int_stint_geometry, stg_laps

6 tests guard this model - 4 generic, 2 singular.

Columns

Column	Type	Nullable	Tests	Description
lap_id		Yes	not_null, unique	Foreign key to stg_laps
stint_baseline_pace		Yes		Median lap_time_s over this stint’s VALID laps strictly before this one. Expanding, not trailing-N: a stint is short enough that the window stays local, and capping it would let the baseline drift with the degradation it measures. NULL until one valid prior lap exists (min_observations=1) — usually laps 1-2, since lap 1 of a stint is an out-lap and not valid. Costs 1.41pp of training-eligible coverage against the old block median. The floor is 1 by measurement, not by default: a 2026-09-10 session silently raised it to 2, and 08i (RESULT 2026-09-18) gated all four priced floors {1,2,3,5} on the v12/08m substrate and reverted to 1. Floors 3 and 5 are rejected — dominated on coverage AND signal. On the v12 training-eligible panel floor 1 leaves 2,156/119,822 rows (1.80%) thermal-NaN, +2.15pp of coverage against floor 2, and cuts the blind rate on the hardest cliff class (0_to_2) from 12.13% to 10.48%.
baseline_observations_n		Yes	not_null, expect_column_values_to_be_between	How many valid prior laps stint_baseline_pace was computed from. Ships as a feature-contract column precisely because the baseline’s NULLs are deterministic on it and on nothing else the contract carries: unlike 02a’s rebuild, whose NULLs were declarable from age_in_stint, this count rises with SC and pit disruption and is plausibly correlated with the outcome. A consumer conditions on it; it must not be imputed away.
push_residual		Yes		stint_baseline_pace - lap_time_s. Positive = faster than the stint’s own prior pace = pushing. NULL wherever the baseline is NULL: unknown, not zero.
cumulative_push_load_surface		Yes		EW sum of positive push_residual, tau~3 laps, 5-lap lookback. A NULL residual on a PRIOR lap contributes 0 (the same convention as laps before the stint began); a NULL residual on THIS lap makes the load NULL rather than 0, since GREATEST(NULL, 0) = 0 in DuckDB would fabricate “no thermal load” for a lap whose load is simply unknown.
cumulative_push_load_bulk		Yes		EW sum of positive push_residual, tau~5 laps, 8-lap lookback. Same NULL rule as cumulative_push_load_surface.