Every generated field must show its work.

A fail-closed prototype that turns public crash records into traceable scenario knowledge—and withholds what the evidence cannot support.

Inspect a public record Download source v0.1

Built from a 30-record, location-free slice of the UK Department for Transport’s 2025 non-sensitive open data. No customer, personal, or private crash data.

01 · source DfT 2025 63324956…b80aa6

02 · row collision_index 2025010552916

03 · field road_type = 6 Single carriageway

04 · receipt accepted 66ab3488…b7b94d

30public sample records

20manually reviewed records

100%reviewed code-label checks

0 / 30fabricated candidates accepted

Public record explorer

Follow one decision back to the row.

Select a case. Every displayed attribute carries an evidence status. When one required source value is unknown, the compiler withholds the functional scenario.

dft-2025:2025010552916 priority · standard

deterministic_derived

Urban single carriageway collision; not at junction or within 20 metres; darkness - lights lit; snowing no high winds; wet or damp surface

Compiler boundary

A model may suggest. Evidence decides.

01
Declare the source
Pin the dataset URL, license, file hash, row key, and allowed columns.
02
Decode, never guess
Map only published codes. Unknown and undocumented values fail closed.
03
Gate model output
A candidate passes only when every field matches the deterministic evidence path.
04
Emit the receipt
Stable JSON, source references, refusal states, and one reproducible hash.

Truth boundary

Useful proof without automotive theatre.

Verified now

DfT code decoding on a reviewed sample
Source-level provenance for accepted fields
Unknown-value and extra-field refusal
Deterministic receipts across repeat runs
Fabricated candidate rejection

Not claimed

Automotive-domain model accuracy
Simulation or OpenSCENARIO readiness
Certification or homologation
Production, Siemens, or customer validation
AI-BOOST application, award, or revenue

Reproducible benchmark

The exact checks are downloadable.

Twelve tests cover deterministic compilation, provenance, unknown values, invalid speeds, location-field rejection, model-candidate gating, catalog matching, and receipt sensitivity.

Read metrics JSON Inspect source manifest →

compiled records 9902bbf77f768b25fcde4425d9926d828c9b64738b8394c03b7d5c0f6bc3919f Same SHA-256 on Python 3.12 and 3.14 repeat runs.

Evidence pipeline fit check · USD 75

Need a result that can survive review?

Send one safe public or redacted target. Romeo Apps will confirm fit before payment and return a source boundary, feasibility decision, and smallest validation path.

Send the safe fit brief