Duranta Auto.
Replay the event over time.
Start with a synthetic road event. Run the sample to inspect speed, braking, movement and gaps. A supported CSV can produce a separate local record; an uploaded clip is recorded by file hash only and is not read into event facts.
A crash leaves a sequence of observations. Auto places car data, weather and road conditions on a timeline, then shows the record for review.
The worked example shows car and camera events together. A supported telematics CSV can drive a local record. Export adds an integrity hash; gaps remain visible.
01 - The input1
Onboard camera clip1
The synthetic example depicts front and cabin camera observations. The local upload path builds from telematics CSV; an uploaded clip is recorded by file hash and read for file metadata only.
clip_2026-09-04T14-02-03Z.mp4 - 12.0s - 1080p dual
Car data
Speed, braking, GPS and movement, recorded by the car itself, second by second.
tel_4471.parquet - 61 km/h > brake 82% > 0 - 1.4s
Weather and road
Weather and road conditions for the same moment, added as labeled inputs: rain 4 mm/h and a wet road surface (from the weather and road feeds in this synthetic example), 0.6 mile visibility at dusk.
wx_14-02Z.json - rain 4mm/h - grip -30% - vis 0.6mi
02 - Try it on your own files
Try it with your own file. Drop in a car data CSV (time, speed, brake, gap) or a camera clip. Everything stays in your browser. Nothing is uploaded. A telematics CSV can build the record below. An uploaded camera clip is recorded by file hash and read for file metadata, not interpreted into event facts in this demo.
03 - Run the replay1
START HERE - press Run. The synthetic example animates a replay in about 10 seconds. Exporting a record computes its local integrity hash.
04 - The replay - drawn from data, not video1
There's no video below. The synthetic frames depict car data and camera events: positions, speeds and times. Uploaded clips do not generate these frames. You can watch the crash play out, then step through it frame by frame. The original video never has to leave the car.
Swipe the scene sideways to follow the cars. →
05 - The timeline1
06 - The record1
| Incident | INC-2026-0904-4471 - rear-end collision, I-91 N mile 22.4 |
| What the example shows | synthetic example only. It shows Vehicle B following Vehicle A while Vehicle A changes lanes, with the gap and speeds from the timeline above. It does not decide fault or liability. |
| Road conditions | Stated assumption: this example assumes rain and a wet road make stopping distance about 1.4 times longer. It is a modeling assumption, not a measured value. Visibility was 0.6 miles at dusk. This is background for the synthetic example, not a finding about cause or liability. |
| Cost estimate | Example repair estimate $6,400. Estimated total range: $6,400 - $9,100 |
| Evidence files | clip hash 9f2c...a41b - car data hash 77d0...e2c9 - timeline time in synthetic example 14:02:17Z |
| Data consistency check | PASSThe synthetic braking values agree with each other. This does not check source authenticity or make an injury or fraud finding. |
| Cycle time | Example run time: 4 minutes 12 seconds. Demo figure, not a real-world benchmark. |
Data checks (synthetic example)
Illustrative checks only. No fraud, injury or prior-history evaluation.
07 - Saved records1
Each record you export is saved in your browser with a local hash-linked audit log, whether it came from the synthetic example or your own file. Reload the page and it remains in this browser. A hash can detect changes; it does not authenticate the source or signer.
08 - Testing at scale1
A synthetic fleet sends car data and camera events. We write crashes into that stream, then the system rebuilds each one and is scored against the known answer. Pick a fleet size and watch the score change. No real footage is used anywhere.
The chart percentages are illustrative synthetic-demo values, not measured outcomes. The 10,000-case button runs a local random stress test with hand-authored labels; its label match is not external validation. Test data comes with this beta product: download a sample dataset (JSON) - 50 synthetic crash records from a 100,000-record generated corpus (the full corpus and generator are not published here).
08b - A real public record next to the synthetic scenario
The replay above is synthetic. To show the same idea on real public data, this card reads one record live from FEMA OpenFEMA (NFIP Redacted Claims v3, id 6320555). Only loss date, flood zone and state are shown. No sensor or inspection sequence exists in that record.
| Status | Loading from FEMA... |
Source: fema.gov/api/open/v3/NfipClaims, id 6320555. The older v2 endpoint is scheduled to be retired October 15, 2026.
09 - How it works
Keep raw video local (design)
The proposed camera workflow would derive event data locally. 2
Compare with known crashes.
The product compares synthetic telemetry features with a synthetic reference library. 1
Hashed records, not stories.
The exported record has a SHA-256 integrity hash and a local audit entry. 2
10 - What this is and isn't
Accident Replay is in beta. It replays the data you give it and shows what's missing. A person makes the call. It does not decide fault, cause, liability or coverage.
Public figures
3 There were an estimated 5,930,496 police-reported crashes in the U.S. in 2022.
3 Weather-related crashes average 744,911 a year (2019-2023), 12% of all crashes. Rain alone averages 574,047 a year, 10% of crashes. (NHTSA FARS/CRSS data).
3 The average loss per collision claim was over $6,150 in 2021, up 18.6% from $5,185 in 2020.
3 Claims adjusters, appraisers, examiners and investigators held 365,300 jobs in the U.S. in 2024, with median pay of $36.92 an hour.
1. This is a beta product. The displayed worked example and its scored test cases are synthetic, not customer outcomes. Alpha v0.2 local-file path. Footnote numbers beside inputs and results identify their source category.
2. Supported structured local files can drive a record. Uploaded media is recorded by file hash rather than interpreted into event facts; a local unkeyed hash detects changes within its stated scope, not source or signer authenticity. No claim, fault, cause, liability or coverage decision is made.
3. Public figures on this page cite Police-Reported Crashes (FARS/CRSS); How Do Weather Events Affect Roads?; 2021-2022 Auto Insurance Database Report; Occupational Outlook Handbook.