Mechanism view · complements the per-scenario leaderboard, does not replace it
Mechanism Robustness
The leaderboard scores one registered intervention (range-clip) across seven scenarios. This page asks a different question of the same reference stack: across every degradation mechanism that has been run at least at a pilot scale — how much of Safety, Comfort & handling and Operation survives, and where does the response first break?
Read this before the numbers. The four axes are not on a common severity scale and their scores are not comparable in magnitude to each other — only within-axis direction, rank and shape are meaningful across mechanisms. A planned shape / geometric-noise axis has no implementation anywhere in the stack (no point-jitter or range-noise injector exists) and is omitted rather than approximated. Cell sizes start at n=3; two ladders (noted below) are excluded because they have no comparable nominal reference. Every number on this page traces to the corrected analysis package analysis/final_v30 — see Provenance.
Corrected 2026-10-11. Until this date the page showed the pre-correction package robustness_analysis_20260916. That package mislabelled 93 range-clip runs as thinning, kept a placeholder S0 label on 160 thinning runs, counted several records per run, stored peak deceleration with the wrong direction and took the nominal reference by row order. All five are fixed; the scores, the mediation table and the breaking points below changed accordingly (for example, NDT dropout no longer clears BH correction on any axis).
A second, independently-built cross-check of this page exists — computed from scratch without reading this page's code or numbers. It was computed against the pre-correction version of this page; it independently found the pooling of non-comparable nominal campaigns that the correction fixes.
Robustness score by axis0–1, higher = less damage · pooled (all runs, equal weight) vs cohort-consistent (one campaign per axis × scenario)
A dash means that metric class has no populated outcome metric in that cell (for example Safety in the urban curve, which has no other road user, so time-to-collision is undefined). The thinning campaign did not record severe braking.
Robustness by scenarioSame score, split by the seven traffic tasks
Scenario
Safe
Comfort & handling
Operation
Mechanism registerEvery degradation mechanism run at least at pilot scale, and which axis it stands in for
Axis
Mechanism
Scenarios
Severities
Runs
Metrics
RS pooled
RS cohort-consistent
Severity ladderNative physical units per mechanism — S0–S3 is ordinal within one mechanism only
Axis
Mechanism
Severity
Native level
Unit
Runs
Scenarios
Breaking pointsPer axis × scenario × mechanism × metric class · Wilson upper bound on the worst-observed failure rate · first severity level whose damage-adjusted score drops below 0.8
Axis
Scenario
Mechanism
Metric class
Wilson high
First <0.8
Absolute BP
Baseline BP
Module attribution — mediationSpearman correlation between a candidate mediator and behaviour damage, BH-corrected across this table
Mediator
Axis
Cells
Spearman ρ
p
p (BH)
No mediator clears BH correction. The largest uncorrected association is NDT dropout on the spatial axis (ρ≈0.41, p≈0.03, BH 0.31), carried by the slow lead target-removal cells; on the temporal axis the sign is reversed (ρ≈−0.43, BH 0.53). This is at most a pointer for attribution, not a causal claim; the package's Oracle-substitution runs (see Provenance) are the planned test.
—: the frozen thesis analysis package (analysis/final) rerun with the corrections, by tools/rev30/corrected/run_v30.sh in the thesis repository; every correction is listed in the package's data/correction_log_v30.txt
Rebuild
41 hashed sources → 68,722 run×metric×stage records, 40,675 after relabelling and removing duplicate records → 526 analysis-base runs, 124 metrics, 0 cross-source conflicts
Determinism
the uncorrected pipeline reproduces the frozen package byte for byte; the corrected run differs from it only through the logged corrections
Reproduces
1,119 total findings / 202 high-confidence (score ≥65), the frozen run-level catalogue, which the corrections do not touch — see the package's own headline reconciliation for two claims that do not reproduce locally (a recall-metric definition conflict, and a continuous-dose breakpoint that needs the full remote slice)
Not yet run
141 core runs + 60 Oracle-substitution runs, costed and sequenced in the package's supplement/ folder, would close every remaining cell