AI Flight Recorder for Autonomous Driving
"Every Decision, Cryptographically Recorded"
— Record every decision with verifiable cryptographic evidence
" Not asking "why" after the accident — leaving "evidence" before the accident. "
2018, Uber autonomous vehicle fatality. 2016, Tesla Autopilot fatality.
Investigators faced the same question:
"What did the AI see, what did it decide, and why did it choose that action?"
Traditional EDRs (Event Data Recorders) capture vehicle physical state.
But the AI's internal decision process goes unrecorded.
DVP is intended to fill this gap — recording the complete causal chain of AI decisions in a tamper-evident format.
Structural challenges in current autonomous driving systems
How inputs from LiDAR, camera, and radar are processed, and why decisions like "go straight," "stop," or "avoid" are made — current EDRs don't record this.
At SAE Level 3 and above, the system performs the driving task while engaged; how legal responsibility is allocated depends on the jurisdiction. After an accident, establishing whether it was "human error" or a "system defect" depends on what was recorded and on whether those records can be independently verified.
UN Regulation No. 157 requires ALKS-equipped vehicles to carry a data storage system for automated driving (DSSAD). Under the EU AI Act, vehicle AI safety components are high-risk under Article 6(1) (Annex I, Section B), and the Act's requirements are to be taken into account in type-approval legislation rather than applying directly (Articles 2(2), 107 and 109). DVP is planned as a profile for decision-level provenance records alongside these rules.
Accident Occurs
Investigation Starts
Log Retrieval
Issues Found
"Unknown if AI recognized the pedestrian"
"Cannot identify brake decision trigger"
"Timestamp integrity unverifiable"
Planned target systems and intended recording levels (no specification exists yet)
Autonomous Vehicles
SAE Level 3-5
ADAS (Advanced Driver Assistance)
Level 2+
Railway Operations AI
Autonomous Operations
Drone Autonomous Control
Autonomous Flight
Autonomous Bus/Shuttle
Level 4
Sensor Layer
Perception Layer
Planning Layer
Control Layer
Data flow and system integration
{
"event_id": "019234ab-7c8d-7def-8123-456789abcdef",
"timestamp_ns": 1734567890123456789,
"event_type": "PERCEPTION_DECISION",
"vehicle_id": "VIN_XXXXXXXXXXXX",
"provenance": {
"actor": {
"type": "AI_MODEL",
"identifier": "perception_v3.2.1",
"model_hash": "sha256:abc123..."
},
"input": {
"lidar_frame_id": "frame_12345",
"camera_frame_ids": ["cam_front_12345", "cam_left_12345"],
"input_hash": "sha256:def456..."
},
"context": {
"speed_kmh": 45.2,
"weather": "RAIN_LIGHT",
"visibility_m": 120,
"active_mode": "AUTONOMOUS_L4"
},
"action": {
"decision": "EMERGENCY_BRAKE",
"confidence": 0.94,
"trigger": "PEDESTRIAN_DETECTED",
"predicted_collision_ms": 1200
}
},
"prev_hash": "sha256:789xyz...",
"signature": "ed25519:..."
}
Design targets against international regulations (planned)
| Regulation | Jurisdiction | Requirements | DVP Support |
|---|---|---|---|
| UNECE WP.29 R157 | UN 1958 Agreement parties (e.g. EU, Japan) | ALKS data recording obligations | Design target (planned) |
| EU AI Act Art. 6(1), Annex I Section B | EU | Vehicle AI safety components are high-risk (Art. 6(1)); the Act's requirements do not apply directly (Art. 2(2)) and are to be taken into account in type-approval acts (Arts. 107, 109) | Design target (planned) |
| ISO 26262 | International | Functional safety, traceability requirements | Complementary |
| ISO 21448:2022 (SOTIF) | International | Safety of the Intended Functionality | Evidence Support |
| NHTSA AV Policy | USA | Autonomous vehicle data recording guidance | Design target (planned) |
| Japan Road Transport Vehicle Act (2019 amendment, in force April 2020) | Japan | Level 3+ data recording requirements | Planned Support |
Design targets against the Data Storage System for Automated Driving (DSSAD) requirements of UN Regulation No. 157. DVP is planned; no specification document exists yet.
Recording of ALKS occurrences (R157 para. 8.2)
Recorded events linked in a hash chain
Timestamp for each occurrence (R157 para. 8.3; accuracy ±1.0 s)
UUID v7 + PTP synchronization
Protection against manipulation, e.g. data erasure (R157 para. 8.5)
Cryptographic hash chain makes alteration detectable (tamper-evident)
Data availability (R157 para. 8.4)
Standard format export capability
Role separation between existing EDR and DVP
Event Data Recorder
→ Records "what the vehicle did"
AI Decision Recording
→ Records "why the AI decided that"
Physical State + AI Decisions
How DVP is intended to support accident investigation and liability determination
| Phase | EDR Only | With DVP Integration |
|---|---|---|
| Accident Occurs | Records collision speed and deceleration | Same + AI recognition/decision history |
| Investigation Starts | "Brake was activated" | "Pedestrian detected 2.3 seconds prior, misclassified as bicycle with 0.67 confidence" |
| Cause Identification | "Unknown if system or human error" | "Recognition model v3.2 low-light classification accuracy issue" |
| Improvement Measures | Speculation-based | Specific model improvement points identified |
| Litigation Response | Insufficient evidence | Cryptographically verifiable evidence trail |
Accident Occurs
DVP Log Retrieval
Cryptographic Verification
Liability Clarification
Design targets — DVP is planned and no specification document exists
DVP development and standardization timeline — no dates set; none of these milestones has been reached, and DVP remains Planned with no specification document
Initial draft specification for public review and feedback
Begin proof-of-concept testing with automotive industry partners
Submit to UNECE Working Party 29 GRVA (Autonomous Driving)
Stable specification release with reference implementations
International standardization activities with ISO and SAE
DVP's position in the framework hierarchy
Join the development of DVP and shape the future of autonomous vehicle safety
"Aircraft have physical black boxes. Autonomous vehicles need AI black boxes too."
— VeritasChain Standards Organization
"The question is not whether autonomous vehicles will have accidents.
The question is whether we can verify what happened when they do."
This work is licensed under CC BY 4.0 International