How to Build an Autonomous-Vehicle Incident Record

How to Build an Autonomous-Vehicle Incident Evidence Record

Build an autonomous-vehicle incident record as a dated evidence ledger. Separate the event, initial notice, operator filing, regulator publication, witness account, technical evidence, investigation finding, enforcement action, corrective response, litigation allegation, and unresolved fact.

Do not assign fault from the first headline or filing.

flowchart LR
    A["Incident"] --> B["Initial notice and preservation"]
    B --> C["Operator and regulator filings"]
    C --> D["Report versions and evidence"]
    D --> E["Investigation and findings"]
    E --> F["Recall, order, or corrective action"]
    F --> G["Open issues and revision history"]

Freeze the incident identity

Start with the date, local time, location, road, vehicle, operator, involved parties, reporting entity, and every available incident identifier.

Do not assume two reports describe two events. NHTSA warns that an incident may have multiple reports, versions, or reporting entities. Its current Standing General Order page provides Report ID, Report Version, Same Incident ID, and Same Vehicle ID fields to help readers reconcile records.

Preserve the original file, retrieval date, hash, and source URL. Keep later versions beside it.

Name the automation state

Record the system, hardware, software version, vehicle mode, whether automation was engaged, when it disengaged, and who performed the driving task.

The reporting rule commonly asks whether ADS or Level 2 ADAS was engaged within a defined period before the crash. That field does not prove causation.

Also record the operational design domain conditions: road type, geography, speed, traffic, weather, lighting, construction, emergency activity, connectivity, and any remote-assistance interaction.

An incident inside an authorized area can still occur outside the system's documented conditions.

Separate observations from interpretations

An observation might be that a vehicle stopped in a lane, a pedestrian was transported to a hospital, or a report listed wet pavement.

An interpretation might be that a perception failure, driver inattention, poor maintenance, unsafe policy, or another road user's conduct caused the event.

Store them in different fields. Attribute every interpretation to the person or agency that made it.

Initial reports often contain unknown or incomplete information. NHTSA says entities must report based on notice even if they have not verified or agreed with the information.

Build the source stack

The stack may include the operator report, NHTSA Standing General Order files, state autonomous-vehicle collision reports, police records, fire and emergency records, transportation-agency records, NTSB dockets, defect investigations, recall material, court filings, company statements, witness accounts, and media.

California DMV publishes collision reports submitted by permit holders. The record remains a manufacturer-submitted report unless an agency adopts its content as a finding.

NHTSA's Special Crash Investigations program explains how deeper investigations may combine scene and vehicle inspection, interviews, police reports, and medical information.

An NTSB investigation can add a factual report, analysis, probable cause, and safety recommendations. A recommendation is not a civil judgment.

Preserve evidence and custody

Relevant material can include vehicle event data, automation logs, perception and planning records, video, audio, telemetry, maps, remote-assistance records, messages, maintenance, software updates, calibration, repair history, passenger communications, roadway signals, surveillance footage, mobile-device records, and physical components.

Do not collect, publish, or transfer private data merely because it may be relevant. Evidence preservation and access require lawful authority, privacy controls, security, retention rules, and qualified counsel.

NHTSA's event data recorder page warns that EDR data have limitations and should be used with other sources.

Record who produced each item, who held it, whether it is complete, what clocks or coordinate systems it uses, whether it was transformed, and what is unavailable or redacted.

Track report versions

A later report may add injuries, correct automation classification, identify another involved party, revise the narrative, or remove a duplicate.

Do not overwrite the earlier version. Store the change date, author, fields changed, reason where known, and effect on the public account.

NHTSA's data page states that the latest report version is used in its charts and explains how duplicate reports can arise. A local ledger should preserve enough history to reconstruct the change.

Wait for findings

Use explicit states such as reported, alleged, under investigation, preliminary finding, final finding, enforcement action, recall, settled allegation, adjudicated issue, and unresolved.

NHTSA's consent order involving Cruise provides a useful example of why the timeline matters. The agency said Cruise failed to fully report a pedestrian crash, including post-crash details. The enforcement record addresses reporting conduct. It should not be paraphrased as a complete judicial determination of every actor's civil liability.

If a final investigative finding does not exist, say so.

Record corrective action without assuming effectiveness

An operator may pause service, change software, retrain staff, expand monitoring, recall vehicles, alter an ODD, revise a safety case, or change incident reporting.

Record whether the action was voluntary, ordered, proposed, completed, independently verified, or still monitored.

A corrective action establishes a response. Its effectiveness needs evidence.

Publish a bounded timeline

A useful public record shows the event, source, evidence type, state, key claim, known limitation, and next expected update. It avoids graphic personal detail that is not needed for the public purpose.

The closing section should distinguish established facts, agency findings, contested allegations, unavailable evidence, and open questions. Link to the primary records.

One incident cannot establish an industry crash rate. For comparative analysis, use [[How to Read Autonomous-Vehicle Safety Claims]]. For case-specific responsibility, use [[Who Is Liable When Automated Driving Fails]] only as an issue map.

This guide was developed with AI assistance from E025, current NHTSA, California DMV, and NTSB evidence practices, and the linked schema. Dalton Anderson remains the author. It is not an investigation, evidence-preservation instruction for a real matter, legal advice, or a fault determination. Editorial, investigative, technical, legal, privacy, source, accessibility, and founder review are required before publication. Publication is not authorized.