ISA validation without the ride-alongs.
ISA validation (EIRA) under Delegated Regulation (EU) 2021/1958, with legal speed limits from two independent sources.
Delegated Regulation (EU) 2021/1958
The loop in one sentence.
We record a drive, establish the legal limit from two sources that never include the vehicle under test, grade the vehicle against it, and hand a person a short ranked queue with the evidence already attached.
From ride-alongs to ranked evidence.
Today: ride-alongs and weeks of desk review
- An engineer rides along and manually flags events with DewesoftX
- Every flagged frame must be re-watched at a desk, one frame at a time
- Verdicts rely on subjective individual judgment and lack reproducibility
- Ground truth requires expensive map subscriptions or is missing entirely
With Marsstein: three independent sources, recorded on board
- Three independent data legs recorded together, on-board, in real time
- Thousands of raw samples grouped into a few dozen ranked situations
- Every decision carries its written reasoning and evidence clips masked by the anonymization model
- Runs offline from an OpenStreetMap extract: no subscription, no network, no data leaving the car
Three sources, one legal limit.
Leg 1 reads posted road signs, Leg 2 resolves GPS against an offline map, Leg 3 reads the vehicle's cluster display. Legal truth is established from two independent sources that never include the vehicle under test.
- 01
Posted sign (front camera)
Detector and classifier with temporal voting across frames. Automatically rejects false positives on trucks or stickers.
- 02
Offline map (OpenStreetMap extract)
Local OpenStreetMap extract on disk. No cloud, no network requests, no subscription fees.
- 03
Instrument cluster (dash camera)
Second vision model reads the vehicle's own instrument cluster, ensuring the vehicle is graded against independent truth.
From the three cameras to the review queue.
- 01Front camera · GPS location · Cluster camera
- 02Truth fusion
- 03ISA conformity evaluation
- 04EU 2021/1958 gates
- 05Ranked review queue
How the legal limit is decided.
The car's own ISA is never an input: it is what is being tested. The legal limit comes from the posted sign or the map, and the car is graded against it. Where neither a sign nor the map settles the limit, that stretch is left out and reported as lost coverage, instead of guessing.
All agree
Sign, map and car show the same limit. Pass, nothing for a human to check.
Car wrong
A confirmed, plausible sign is the legal limit. The car shows something else, so it is flagged.
Sign still holds
A sign holds after it has passed, for a distance set by the road class: 2 km on a residential street.
Map decides
No sign in range. Outside town, the country's default limit for the road applies, and the car matches it.
Map wins
The posted sign and the map agree. The car shows something else, so it is flagged.
Synthetic drive, plates masked by our anonymization model. Sign detection is our own model's output; car and map values are illustration where the tile says so.
Delegated Regulation (EU) 2021/1958 criteria, enforced in code.
Enforced as hard software gates: if a drive misses a single criterion, it is reported as FAIL with explicit reasons.
Accuracy over distance
Accuracy weighted by distance over the whole route
Road types separately
Urban, non-urban and motorway measured separately
Road type share
A minimum share of the route enforced for each road type
Detection latency
A maximum response time after passing a sign
Route length
A minimum length for the pooled campaign route
Darkness share
A minimum share of night and darkness driving, from the sun's position
Less review work through a ranked queue.
From thousands of raw samples down to a short list for human review, ranked worst first, with masked video clips and machine recommendations attached.
A person's signed verdict is the certifiable record. Each drive produces its events and GPS track, an annotated video, evidence photos and clips per event, the conformity figures computed on board, a provenance manifest of the code, models and map extracts used, and masked copies of everything shareable.
Validated, then masked.
The drive is graded while it is recorded. Once the recording is saved, the anonymization model masks faces and license plates in it. The raw recording stays internal, the masked copy is what goes to review and delivery, and the verdict overlay stays readable.
How the anonymization model masks