int_compound_cliff_predicted

Tyre compound degradation prediction using fitted survival-analysis parameters.

Part of the Pace Baselines family.

For the conceptual foundation, read Int Compound Cliff Predicted.

Tyre compound degradation prediction using fitted survival-analysis parameters. One row per lap. Predicts degradation rate and cliff onset based on fitted compound parameters and current thermal conditions.

Lineage

Overview

Property	Value
Layer	intermediate
Family	Pace Baselines
Contract enforced	No
Upstream	int_stint_geometry, dim_compounds_season, stg_weather

7 tests guard this model - 5 generic, 2 singular.

Columns

Column	Type	Nullable	Tests	Description
lap_id		Yes	not_null, unique	Foreign key to stg_laps
compound_wear_s		Yes	not_null	The age-dependent portion of the hockey-stick curve, bounded at var(‘compound_wear_max_s_per_lap’). Exposed as its own column so the bound is assertable directly — see assert_compound_wear_bounded.sql. expected_compound_pace_s is this plus grip_peak plus the temperature offset, so the pace total legitimately exceeds the bound while this column may not. As of 08m (2026-09-16) the formula is wear_gradientage + cliff_ramp, where cliff_ramp is the de-double-counted, moment-matched saturating post-onset term defined once in macros/compound_cliff_wear.sql. It was previously wear_gradientage + 0.002age^2 + severitylaps_past_cliff. Both of the removed terms were unfounded: compound_cliff_severity is FITTED as a level shift in seconds over a ~5.5-lap window and was being charged once per lap (laps_past_cliff reaches 48), and the 0.002 quadratic was a literal constant identical across all 438 seed cells that was never fitted against anything. 08l measured them at 60.4% and 18.9% of the seed’s contribution to the ML target respectively.
compound_wear_gradient		Yes	not_null	The fitted linear wear slope (s/lap) for this lap’s circuit x compound x season cell, passed through from dim_compounds_season. Exposed by 08m alongside compound_cliff_severity because any consumer needing a post-onset RATE needs both: severity is a level shift, and de-double-counting it against this gradient is what turns it into a rate. Use the cliff_ramp_slope_s_per_lap() macro.
expected_compound_pace_s		Yes	not_null	Expected pace loss vs compound baseline, in seconds. Bounded as of Phase 8 (2026-08-25): the wear portion is capped, having previously reached 93.5 s/lap on laps that were actually run. This column is subtracted into driver_skill_residual_s, so the unbounded tail was depressing the residual on 12,575 post-onset laps and, through it, corrupting both ML targets — it hid 6,825 cliff-label crossings (99.5% of them into ‘none_in_stint’) and moved the degradation target on 9.05% of rows. 08m (2026-09-16) changed what this column MEANS, not just its precision: the severity and quadratic terms it carried were removed (see compound_wear_s). Any number in the tree measured against next_5_lap_cumulative_jump_s before that date is on the old quantity and is not comparable to one measured after it.