Causal attribution for scenario testing
Scenarios now come from fleet logs, crash reports (human drivers and automated vehicles) and LLM generators: nobody knows them intimately any more. The Causal Map restores that knowledge — which actors a failure needs, which variables carry it, whether it was avoidable — measured in simulation.
Why now
More and more scenarios arrive undesigned: logged, reported, generated.
Pass/fail says it failed; the Causal Map measures why, by re-simulation: whether the SUT could have avoided it, and what least change avoids it.
Simple: change one thing, replay the scenario, measure the change on your KPIs. Hard: knowing which changes to try. The Causal Map computes which ones to run.
Expertise erodes when understanding is delegated. Teams that stop seeing why their scenarios fail stop being the experts.
ISO 21448 (SOTIF) asks you to identify triggering conditions behind hazardous behaviour, including in scenarios nobody on your team designed.
Method
Nothing fails yet? It first looks for a failure in the declared space, then explains what it finds.
Cases
One illustrative case per type of input: automated-vehicle crash reports, fleet logs, human-driver crash reports, LLM-generated. The same automated analysis runs on each, whatever its source.
Report SGO 31101-14925, 3 actors: the SUT moves out to pass a parked work truck; a pickup merging from the right cuts it off. Reconstructed from the report's narrative, the SUT in the reported vehicle's place.


Was it avoidable? yes: the SUT's controller does not brake before the contact or near-miss
Which actors are necessary, alone or together? the pickup; not the parked work truck
What least change avoids it? the SUT's speed, 37 → 27 km/h
Did a broken traffic rule cause it? none: the files declare no traffic rule to break
When was it lost? 0.17 s before the near-miss
Which variables carry it? {SUT speed}, {lateral speed}, {gap} under all four KPIs; {lateral speed}: how hard or close, not whether
How far can the scenario shrink and still fail? without the parked work truck: same outcome, 2 of 3 road users kept
The truck the SUT moves out to pass is not necessary: the scenario fails without it.
Reported by Avride Inc. to NHTSA under SGO 2021-01, report 31101-14925, April 2026. NHTSA, Standing General Order 2021-01 crash reports.
Waymo bf90ada8, dense urban, same direction, 7 actors: the partner starts moving into the SUT's lane. The SUT takes the part of a recorded, human-driven car; the other road users, the Waymo car included, follow the log.


Was it avoidable? yes: the SUT's controller brakes too softly
Which actors are necessary, alone or together? the partner; not the Waymo car, A1, A2, A4 and A5
What least change avoids it? when the partner starts, 1.5 s earlier
Did a broken traffic rule cause it? none: the files declare no traffic rule to break
When was it lost? 0.93 s before the contact
Which variables carry it? the partner's {speed}, {start}, {speed, start}; {speed}: how hard or close, not whether
How far can the scenario shrink and still fail? without A1, A2, the Waymo car, A4 and A5: same outcome, 2 of 7 road users kept; the partner's start fails only between 1.7 and 3.2 s: search space −36 %
5 of 7 road users removed: the partner alone carries the failure.
Case: Waymo Open Motion Dataset, scenario bf90ada8a1c8cdce, https://waymo.com/open/
Case NMVCCS 2007-03-018, 4 actors: in the report, an SUV hits a left-turning motorcycle (V2), then a pole and a pedestrian; the SUT drives the SUV's part. Reconstructed from the investigation's narrative.
“The crash occurred in a signalized intersection of two roadways. […] V1, a 2002 Ford Expedition sport utility vehicle with one occupant, was traveling southbound in lane two. V2, a Harley-Davidson motorcycle with two occupants, was traveling westbound in lane one, and attempting to make a left turn. At the intersection, the front of V1 contacted the right side of V2.”


Was it avoidable? yes: the SUT's controller brakes too late, never brakes or brakes too softly, depending on the case
Which actors are necessary, alone or together? the motorcycle; not the pedestrian and the pole
What least change avoids it? the motorcycle's gap, 5 → 20 m
Did a broken traffic rule cause it? no: no road user breaks a rule the files declare
When was it lost? 0.42 s before the near-miss
Which variables carry it? {SUT speed}, {V2 speed}, {V2 gap}, {V2 speed, V2 gap}; {SUT speed} and {V2 speed}: how hard or close, not whether
How far can the scenario shrink and still fail? without the pedestrian and the pole: same outcome, 2 of 4 road users kept; the motorcycle's gap fails only at 5 and 10 m: search space −50 %
The report's pole and pedestrian are not necessary: the motorcycle alone is.
National Highway Traffic Safety Administration (NHTSA), National Motor Vehicle Crash Causation Survey (NMVCCS), 2005–2007: US government public data. Case: NMVCCS 2007-03-018, in the NMVCCS case viewer (enter the case number).
An LLM turns the text into an OpenSCENARIO file. Light and weather are not simulated (wet: road friction).
“Take the Euro NCAP AEB car-to-car rear stationary scenario and extend it: the road descends a 4 percent grade past a crest, and the stationary target only becomes visible once the crest is passed. Two further vehicles form a stopped queue 25 m and 50 m behind the target. Night scene, wet road surface. An emergency vehicle overtakes ego in the left lane at 130 kph, 8 s after the start, and ego must move back to the right. Joint sweep over (ego initial speed: 50, 70, 90 kph), (crest-to-target distance: 40, 60, 90 m), (road friction: 0.4, 0.7), (emergency vehicle delay: 6, 8, 12 s).”


Was it avoidable? yes: the SUT's controller brakes too late or never brakes, depending on the case
Which actors are necessary, alone or together? the stationary target under two or more KPIs; the first queued car under one KPI only, the emergency vehicle not counted (the SUT's script waits on it), the second queued car under none
What least change avoids it? the SUT's speed, 50 → 70 km/h
Did a broken traffic rule cause it? no: no road user breaks a rule the files declare
When was it lost? 2.40 s before the contact
Which variables carry it? {SUT speed}, {crest distance}, {EMV delay}
How far can the scenario shrink and still fail? the second queued car's removal could not be checked: too few configurations drawn with a failure; search space 54 → 12 concrete scenarios (−78 %), every road user kept
From a requirement's text to a reduced scenario: search space −78 %.
Inside your test loop
causal_map, through a standard agent-integration protocol (MCP), from a dev environment, a chat surface or a CI step.
Consumes and emits the open scenario and road-network formats. No proprietary scenario language to adopt or migrate to.
A thin adapter over the real simulation tool rather than a reimplementation.
Locally, on-premise or in the cloud, and suited to CI/CD.
It explains failures; it does not certify safety.
Value
Every finding is measured, not modelled; every case shown ships with the scenario file that reproduces it.
Drop the actors the failure does not need, and the values that never fail: a leaner test, fewer runs every release.
Was it avoidable? If not, no ticket.
When it was lost, and the least change that avoids it, as a scenario you can replay.
The carrying variables are the coverage model, written for you in OpenSCENARIO 2.0.
Same scenario, two systems: what changes in the explanation, what holds.
AI writes the scenarios; your team still sees why each one fails.
Triggering conditions, measured: which road users matter, which variables carry the failure, over which values.
About
Tempero (Latin: to temper, to refrain) exists to mitigate the risks of high-risk AI systems. The founding team comes from automated reasoning, formal testing and constraint solving, with prior roles at Microsoft Research and autonomous-systems institutes. Automotive SDVs come first, because they rest on mature standards such as OpenSCENARIO.
See it for yourself: personal access to the Causal Map server, on request.
Request access