int_pit_strategy_value

One row per stint.

Part of the Strategy family.

One row per stint. Evaluates pit strategy: what was the optimal pit lap vs actual? The argmin of int_pit_strategy_cost_curve’s Total_Cost(L) over every candidate lap in the horizon, replacing the “first lap where expected wear exceeds 0.5 s” threshold that stood here until 2026-08-23 and had no pit-loss term in it. That threshold returned ‘optimal’ on 8 stints of 7,129 and ‘unknown’ on 2,231; the argmin returns 715 and 266. Two horizons are reported. The verdict follows the two-stint window (this stint’s first lap through the next stint’s last), the only framing well posed for the 61% of stints in a 2-stop or longer race; the race-remainder optimum is carried alongside as a diagnostic. opportunity_cost_s is now Total_Cost(actual) - Total_Cost(optimal), so pitting early carries a cost too, which the overrun-only definition could not express. Pit-loss is per circuit from int_pit_loss_circuit’s empirical, EB-shrunk estimate (19.3–29.6 s over the 147 races that resolve one), falling back to the circuit_reference seed and then 21.0 s; pit_loss_source records which was used.

​

Lineage

​

Overview

Property	Value
Layer	intermediate
Family	Strategy
Contract enforced	No
Tags	simulation, strategy
Upstream	int_stint_geometry, int_compound_cliff_predicted, int_pit_loss_circuit, stg_pits, int_lap_air_state, int_pit_strategy_cost_curve, stg_lap_tyre_qa, int_stint_end_regime

15 tests guard this model - 13 generic, 2 singular.

​

Columns

Column	Type	Nullable	Tests	Description
stint_id		Yes	not_null, unique	PK, FK to int_stint_geometry
optimal_pit_lap		Yes		Optimal pit lap number (absolute race lap) - the lap the car enters the pit lane, i.e. one after the last lap run on the set, matching stg_pits.pit_in_lap_number. NULL when the compound has no fitted curve to minimise over.
optimal_pit_lap_in_stint		Yes		Optimal laps to run on the set before boxing, counted in VALID laps. Not a chronological lap_in_stint: SC and pit laps do not age a tyre the way a racing lap does, and the cost curve is indexed on tyre age.
actual_pit_lap		Yes		The stop that ended this stint: stg_pits.pit_in_lap_number, the latest pit-in inside the stint’s full chronological span (its first lap of any validity through end_lap_number). Non-NULL on every stint that ended in a stop, whatever flag was out (assert_pit_ended_stints_have_stop, T23). A stint that did not end in a stop (the driver’s last stint, or a retirement) is NULL unless a pit-in falls inside its span: a retirement into the pit lane, or, on 2 stints, a stop bronze did not split the stint at (2018_15 MAG lap 48; 2022_8 LEC lap 20). Until WI-13 (F31) the stop was looked for only up to the stint’s last VALID lap + 1, which missed it after a run of SC/VSC/red-flag or other invalid laps: 418 pit-ended stints (107 of 116 red-flag-ended, 194 of 653 SC-ended) came out NULL and were graded as if they never stopped. It is a window, not equality with end_lap_number: on 13 stints bronze puts the stint boundary a lap or two after the in-lap (2018_2 VET: pit-in lap 17, stint’s last lap 18).
end_lap_number		Yes	not_null	The lap the stint ended on, of any validity (int_stint_end_regime.end_lap_number): the in-lap, or up to two laps after it, for a stint that ended in a stop; the chequered flag or the retirement lap for one that did not. Not the stint’s last valid lap. The app’s pit-strategy Gantt orders stints on it and ends a bar with no stop here, where it used to fall back to the race’s last valid lap and drew those bars to the flag (WI-13).
overrun_laps		Yes		Actual laps run on the set minus optimal, in valid laps. Negative = pitted early. Compared as offsets rather than lap numbers so an invalid lap before the stop cannot skew it. NULL when either side is NULL. Median -3 valid laps per stint across 2018-2025 (n = 4,905 stints with both an actual stop and a model optimum; per-season medians -1 to -4): teams stop earlier than a degradation-only optimum, which is what track position buys.
opportunity_cost_s		Yes	expect_column_values_to_be_between	Total_Cost(actual) - Total_Cost(optimal), in seconds. Non-negative by construction: the optimum is a minimum over the same curve. Zero for a stint that never stopped, which had no pit decision to grade. Median 5.5 s, p95 83 s.
strategy_verdict		Yes	accepted_values	Strategy classification.
optimal_pit_lap_confidence		Yes	expect_column_values_to_be_between	How sharply identified the optimum is: the share of candidate laps whose total cost is more than a second worse than it. 1.0 = one lap stands out; near 0 = the curve is flat and the exact lap barely matters. 0.0 where no curve exists. Replaces the 0.0/0.1/0.8 sentinel ladder, which encoded whether cliff onset was NULL rather than anything about the estimate.
pit_lane_loss_s		Yes	not_null, expect_column_values_to_be_between	Pit-lane loss used for this stint’s undercut-threat scan, in seconds. Empirical where int_pit_loss_circuit resolves the venue, else the circuit_reference seed, else 21.0.
pit_loss_source		Yes	not_null, accepted_values	Which pit-loss estimate this row used - empirical / seed / default.
pit_loss_imputed_flag		Yes	not_null	TRUE when nothing was measured on either side - no empirical stops resolved and the seed row was itself a guess.
horizon_source		Yes	accepted_values	Which horizon the verdict’s argmin used - two_stint (4,014 stints) or the race_remainder fallback for a stint with no successor (2,642). NULL where no fitted compound curve exists to minimise over (473).
horizon_laps		Yes		Length of that horizon in valid laps.
optimal_pit_lap_race_in_stint		Yes		The race-remainder optimum, in valid laps on the set. Diagnostic: asks the harsher unconditional question “if they had stopped only once more all race, when?” rather than grading the stop inside the strategy the team actually ran.
optimal_pit_lap_race		Yes		Absolute race lap for optimal_pit_lap_race_in_stint.
opportunity_cost_race_s		Yes	expect_column_values_to_be_between	The same actual-minus-optimal difference measured on the race-remainder horizon. Systematically larger than opportunity_cost_s for a 2-stopper, because a one-stop counterfactual charges them for a stop they were never going to skip.