CROSS-RUN ASSURANCE · DETERMINISTIC EVIDENCE EXPORT

AUTONOMY EVIDENCE COMPARISON WORKBENCH.

Compare two synthetic autonomy runs without overwriting either history. See which inputs, events, decisions, proof conditions, and assurance claims changed—and export a hash-addressed evidence bundle for independent review.

Return to the 3D/VR lab Accepted-baseline regression Open evidence API contract →

BASELINE RUN CANDIDATE RUN

Deterministic comparison loaded.

REVIEW STATEREVIEW REQUIRED

INPUT CHANGES1

OUTCOMECHANGED

SELECTED TRAJECTORYCHANGED

PROOF REGRESSIONS0

COMPARISON DIGESTdcc6fe64b104a4f3…

BASELINE

Sudden crossing object

SAFE AVOIDANCE

Scenario: sudden-crossing-object
Outcome: SAFE AVOIDANCE — ROUTE REACQUIRED
Selected: TRAJ-VERTICAL
Applied action: VERTICAL DEVIATION + SPEED REDUCTION
Proof: 14/14 pass
History: dad037e36f9a3d052b2c…

CANDIDATE

Late detection / minimum-risk transition

HOLD

Scenario: late-detection
Outcome: MINIMUM-RISK HOLD — NO COMPLIANT TRAJECTORY
Selected: NONE
Applied action: HOLD / STOP
Proof: 14/14 pass
History: 9dfff8b5f20af445295d…

MEASURED DIFFERENCE

Run metrics

NO COMPOSITE SAFETY SCORE

Candidate values are compared with the baseline without declaring one universally safer.
Metric	Baseline	Candidate	Delta
Minimum Separation	0.628	0.0	-0.628
Arrival Delay Seconds	1.05	NOT DERIVED	NOT DERIVED
Path Length Penalty	1.904	0.0	-1.904
Final Mission Progress Percent	100.0	58.0	-42.0
Total Reaction Latency Ms	450.0	450.0	0.0
Deadline Misses	0	1	1.0
Event Count	18	15	-3.0
Frame Count	131	131	0.0

CONFIGURATION DIFF

Changed inputs

Only normalized public-safe inputs may appear.
Field	Baseline	Candidate	Impacted domains
scenario_id	sudden-crossing-object	late-detection	WORLD_MODEL, PERCEPTION, PLANNING, MISSION_OUTCOME

APPEND-ORIENTED HISTORY

First event divergence

DIVERGED

INDEX9

BASELINE EVENTTRAJECTORY_SELECTED

CANDIDATE EVENTNO_FEASIBLE_TRAJECTORY

Baseline-only codes: HAZARD_CLEARED, MINIMUM_SEPARATION_RECORDED, OBJECTIVE_REACHED, ROUTE_REACQUIRED, TRAJECTORY_SELECTED
Candidate-only codes: MINIMUM_RISK_HOLD_ESTABLISHED, NO_FEASIBLE_TRAJECTORY

REPLAY CHECK

Determinism and integrity

PASS

Baseline replay: MATCH
Candidate replay: MATCH
Canonical history mutation: NONE
Persistence: NONE
Output terminal: SYNTHETIC_NULL_SINK

CHANGE-IMPACT REVIEW

Which assurance claims remain supported?

REVIEW REQUIRED

Compared Sudden crossing object with Late detection / minimum-risk transition. 1 normalized input field(s) changed. Outcome changed. Selected trajectory changed. Human or independent review is required.
Claim	State	Triggered by	Qualification
EV-CLAIM-REPRODUCIBILITY Each run reproduces the same logical history for the same version, normalized inputs, and seed.	SUPPORTED WITHIN DECLARED BOUNDS	No changed input	None stated.
EV-CLAIM-PERCEPTION-TRACKING The candidate perception and tracking record remains supported under the declared synthetic sensing conditions.	SUSPENDED PENDING REVIEW	scenario_id	A changed sensing, world-model, tracking, prediction, or uncertainty input requires renewed review even when proof invariants continue to pass.
EV-CLAIM-PLANNER-SAFETY-FILTER The selected candidate and independently filtered applied action remain reconstructable and compliant with declared normalized constraints.	SUPPORTED WITH QUALIFICATIONS	scenario_id	None stated.
EV-CLAIM-AUTHORITY-SUPERVISION The candidate action is bounded by current authority and the declared human-supervision condition.	UNCHANGED	No changed input	Connectivity or successful execution does not restore or substitute for authority.
EV-CLAIM-MISSION-OUTCOME The mission result remains the same under the candidate normalized inputs.	SUPPORTED WITH QUALIFICATIONS	scenario_id	None stated.
EV-CLAIM-PROVENANCE-REPLAY Both canonical histories remain append-oriented, hash-addressed, and independently replayable.	SUPPORTED WITHIN DECLARED BOUNDS	No changed input	None stated.
EV-CLAIM-REAL-WORLD-CERTIFICATION The comparison establishes real-world vehicle safety, operational effectiveness, legal authorization, or certification.	NOT ESTABLISHED BY SIMULATION	No changed input	Simulation evidence remains bounded by declared assumptions and does not eliminate the reality gap.

PORTABLE EVIDENCE

Export without rewriting history.

Download the comparison itself or a deterministic evidence bundle for either run. Bundles include input hashes, event history, decisions, proof checks, frame manifest, accessible equivalents, and the synthetic boundary. Full replay frames are optional.

Include all replay frames

The comparison identifies deterministic differences and affected assurance claims within the declared synthetic model. It does not determine legal responsibility, certify a vehicle, or establish real-world safety.