Validation programme

Three things, kept apart because they are different claims: what the engine does on controlled cases whose answer is known in advance, what has and has not been established against real mines, and the procedure that decides the second. Every state on this page is asserted by a committed record named beside it, not inferred from the absence of evidence. HOLD means a gate is not established; it is not a failure, and it is not rolled into one. In one sentence: HaulLens has passed its controlled engineering benchmarks and is moving into blind real-mine validation.

Engineering Benchmarks loading

Synthetic terrains with a closed-form answer, each run through the product's own study path unchanged. A PASS here means the engine did what an engineer would expect on that case. It is not real-mine validation: no surface here was surveyed and no road here was built.

Loading the benchmark record…

Real-Mine Validation loading

The programme's cases against real mines. No case has reached VALIDATED. The public-terrain search for a qualifying case is closed; the record and the reason are named below.

Loading the validation registry…

Methodology Blind comparison

The blind procedure a real-mine case goes through. Nothing is re-run after the answer is seen, and the registry code refuses VALIDATED unless both the blind run and the independent comparison are PASS.

Receive the pre-construction terrain (T0), the task zones and the engineering criteria.
Independently seal the constructed reference: its hash is recorded, the file stays with the provider or a third party.
Define origin and destination zones from the facilities themselves, never from the built road.
Freeze the study inputs: hash T0, the zones, the criteria and the run configuration.
Run HaulLens once.
Hash and freeze the HaulLens output.
Unseal the reference.
Compare under pre-registered criteria: corridor separation, length ratio, grade profile, endpoint agreement, exclusion contact; tolerances from the reference's own positional uncertainty.
Record PASS, HOLD or FAIL in the registry with every hash. A FAIL is recorded as a FAIL, not tuned away.

What HaulLens can say today

It behaves as expected on controlled engineering benchmarks (above), with inputs, expected behaviour, result and evidence recorded per case.
It refuses when no compliant corridor exists and withholds figures whose evidence is missing.
Its runs are deterministic and every input and output is hashed.
Its independent reference-extraction instrument detected real road construction in public two-epoch terrain and returned no false roads on self-comparison.

What HaulLens must not say

That it is validated against a real mine. No blind real-mine run has been completed.
That any benchmark result is mine-validation evidence.
That a HaulLens corridor is an engineered alignment, or that its earthworks or economics are construction quantities.
That the reference-extraction result says anything about route generation.

Records: validation/PUBLIC_DATA_MINE_HUNT_CLOSURE.md, validation/real_mine_request/DATA_REQUEST_PACK.md, benchmarks/engineering_benchmarks/PROOF_PACK.md.