fct_driver_isolation_race

Driver isolation per driver-race: the three ratings rolled up with lap weighting (pair-lap weighting for relative), SEs that add each stint’s slope error in quadrature, lambda, shrunk value, confidence_pct and trust_label, the tier-3 identity split, and context.

Part of the Marts family.

Driver isolation per driver-race: the three ratings rolled up with lap weighting (pair-lap weighting for relative), SEs that add each stint’s slope error in quadrature, lambda, shrunk value, confidence_pct and trust_label, the tier-3 identity split, and context. Keyed with the same surrogate-key recipe as fct_driver_skill_features, which is deliberately not changed (an unvalidated rating must not enter a trained model’s contract). No minimum-race floor: lambda does that job. Positive = faster. Leakage: a function of the residual trajectory; no ML-contract mart may depend on it (T48, ml/tests/test_manifest_contract.py). WI-16a, _roadmap/_fixes/wi/WI-16-cumulative-driver-isolation.md.

Lineage

Overview

Property	Value
Layer	marts
Family	Marts
Contract enforced	No
Tags	driver_isolation
Upstream	fct_driver_isolation_lap, int_driver_isolation_stint_tyre, fct_driver_isolation_pair_lap

19 tests guard this model - 16 generic, 3 singular.

Columns

Column	Type	Nullable	Tests	Description
driver_race_id		Yes	not_null, unique	PK: dbt_utils.generate_surrogate_key(race_year, race_id, driver_id), the same recipe as fct_driver_skill_features.driver_race_id, so the two join directly.
race_year		Yes	not_null	Season.
race_id		Yes	not_null	FastF1 event id, e.g. 2021_8.
driver_id		Yes	not_null	Driver (three-letter code). In the pair mart, the focal driver whose gains the row reports.
constructor_id		Yes	not_null	Constructor (stg_laps.constructor_id).
era		Yes		Regulation era: pre<era_boundary> or post<era_boundary> (var era_boundary, 2022).
car_iso_s		Yes		Isolation car term, s, NEGATIVE = FASTER than the race’s lap-weighted car (int_constructor_car_fe_isolation). NULL when not identified.
car_term_source		Yes	accepted_values	Where car_iso_s came from: driver_id (fitted per-era), unidentified (multiple components in race), not_estimated (pyfixest singleton), or no_fit_row (no row in the fit).
n_stints		Yes	not_null	Stints with Ω laps in the race.
n_laps		Yes	not_null	Ω laps in the group.
pace_isolated_gain_s		Yes		Mean p over the race’s Ω laps, s/lap, positive = faster.
n_pure_laps		Yes	not_null	Non-extrapolated laps with pure in the group (the pure rating’s n).
n_pair_laps		Yes	not_null	Pair-laps in the group (the relative rating’s n).
n_relative_laps		Yes	not_null	Distinct laps with at least one peer in the group.
n_teammate_pair_laps		Yes	not_null	Of n_pair_laps, those against a teammate.
relative_pace_raw_gain_s		Yes		t(peer) - t(driver), s/lap: model-free, POSITIVE = the driver faster. In aggregates, the mean over pair-laps.
identity_n_pair_laps		Yes		Pair-laps with both car terms, over which the identity_* contributions are computed.
identity_relative_pace_gain_s		Yes		Mean relative_pace_gain_s over the identity_n_pair_laps rows, s/lap. = the three identity_* contributions plus the group’s mean tyre-age bias (the seed’s pricing of the peers’ age gap, at most the match tolerance; not published, 0 when every pair is on the same tyre age). W63 design 2, 2026-10-03.
identity_pace_gap_gain_s		Yes		Contribution: SUM(pace_gap) over rows with both car terms / identity_n_pair_laps, s/lap.
identity_car_advantage_gain_s		Yes		Contribution: mean car_advantage_gain_s over identity rows, s/lap; > 0 = the driver’s car faster.
identity_traffic_advantage_gain_s		Yes		Contribution: mean traffic_advantage_gain_s over identity rows, s/lap; > 0 = he lost less to dirty air.
tyre_offset_vs_field_s		Yes		Mean tyre_offset_vs_field_s over the race (context).
lift_coast_excess_share		Yes		Mean lift_coast_excess_share over the race (context).
dirty_air_lap_share		Yes		Share of the group’s laps with is_dirty_air_lap.
pure_skill_gain_s		Yes		Pure pace skill, mean over the race’s non-extrapolated laps at driver-race grain, s/lap, POSITIVE = faster / better (raw, before shrinkage). Includes pure_season_offset_gain_s.
pure_season_offset_gain_s		Yes	not_null	The declared season-level offset inside pure_skill_gain_s, s/lap (positive = the season’s pure is raised; 0 for seasons without one). W33 Option A, 2026-09-30: 2018 +0.37 s, so a 2018 pure value sits at the 2019-2024 seasons’ level. It does not change any within-season ranking or gap. Cross-season comparisons remain caveated (docs/decomposition/driver-isolation.mdx, Limitations).
pure_skill_se_s		Yes		SE of pure_skill_gain_s, s/lap (see the model header for the formula).
pure_skill_lambda		Yes		Reliability tau^2 / (tau^2 + SE^2) in [0, 1]; tau^2 per (season, rating) at this grain. 1 when SE = 0.
pure_skill_shrunk_gain_s		Yes		lambda * pure_skill_gain_s: shrunk toward 0 (the ratings are field- or peer- centred), s/lap.
pure_skill_confidence_pct		Yes		ROUND(100 * lambda * method_score); NULL until WI-16b validates the method (seed driver_isolation_method_scores).
pure_skill_trust_label		Yes	accepted_values	solid (>= 75), indicative (50-75), thin (25-50), suppress (< 25, grade F or no lambda), unvalidated (no method score yet). NULL without a raw value.
relative_pace_gain_s		Yes		Relative pace vs same-strategy peers, mean over the race’s pair-laps at driver-race grain, s/lap, POSITIVE = faster / better (raw, before shrinkage).
relative_pace_se_s		Yes		SE of relative_pace_gain_s, s/lap (see the model header for the formula).
relative_pace_lambda		Yes		Reliability tau^2 / (tau^2 + SE^2) in [0, 1]; tau^2 per (season, rating) at this grain. 1 when SE = 0.
relative_pace_shrunk_gain_s		Yes		lambda * relative_pace_gain_s: shrunk toward 0 (the ratings are field- or peer- centred), s/lap.
relative_pace_confidence_pct		Yes		ROUND(100 * lambda * method_score); NULL until WI-16b validates the method (seed driver_isolation_method_scores).
relative_pace_trust_label		Yes	accepted_values	solid (>= 75), indicative (50-75), thin (25-50), suppress (< 25, grade F or no lambda), unvalidated (no method score yet). NULL without a raw value.