We wanted to know
whether it actually works.
A benchmark here is not a number. It is a dataset, a protocol, a split, a set of leakage controls, a measurement, and a statement about what has and has not been verified. Experiments that fail are recorded with the same care as the ones that succeed.
Benchmark timeline
Experiments
Measurement
Reproducibility
The experiment records above are the real per-specialist measurements, each carrying
its dataset, split, protocol, source artifact and evidence state. The one frozen,
VERIFIED benchmark number is the LEVIR-CD-256 change head (pooled IoU 0.8122) and the
verified live deployment; the grounding, fusion, change-VQA and calibration records are
measured but UNVERIFIED (two protocols or an open metric ruling). The calibration
reliability plot in section 03 is a measured result, transcribed from
artifacts/calibration_v001.json (change_vqa, Val, n=16441); the remaining
plots in section 03 are illustrative, not measured results. No system-level end-to-end
benchmark exists; the
unified benchmark adapters and the router test remain NOT RUN. A number here is only as
strong as the state attached to it.