VERIFY

How verification runs and exactly what each check asserts. Source of truth: verify.mjs. Machine-readable result: verification-report.json.

Run

node verify.mjs   # exit 0 = all checks pass; exit 1 = any failure

One verification structure runs against two benchmarks:

• Synthetic (always): generateDataset({n:1200,T:12,seed:7}), 70/30 split.
• Official (auto-detected): if data/official/zonal.csv +

data/official/train-labels.csv exist, the SAME evaluation checks run on real data via out-of-fold k-fold cross-validation, writing data/official/submission.csv.

Checks and what each asserts

Structural (mode-independent):

1. dB ↔ linear power invertible — linToDb(dbToLin(x)) == x, err < 1e-6.
2. Temporal Lee filter — speckle variance cut > 40%, temporal mean within

10% (src/sar.mjs::temporalLeeFilter).

1. Agricultural-extent score monotone — cropland dynamics score above stable

cover.

1. Reproducible — same seed → identical MSE (bit-stable).
2. Performance budget — trains < 8 s (closed-form ridge).
3. Ingestion seam round-trip — long-format zonal CSV parses back into

ordered per-village stacks (src/ingest.mjs).

Evaluation (per active benchmark, identical logic in src/evaluation.mjs):

1. Skill vs baseline — model MSE > 30% below the naive column-mean baseline.
2. Pooled R² > 0.45.
3. Physical constraints — every predicted area in [0, village_area].
4. Total-area recovery — corr(predicted total, true total) > 0.6.
5. Submission schema — header exactly

ID,Rice_ha,Cotton_ha,Maize_ha,Bajra_ha,Groundnut_ha; numeric rows; count matches.

Pass/fail semantics

• status: PASS requires every check passed === true.
• proofStatus is PROOF_INCOMPLETE regardless of PASS, because the official

dataset is a disclosed seam. PASS describes the checks that ran; it does not assert real-data accuracy.

What would invalidate the synthetic claims

• Changing src/synth.mjs signatures/noise or the seed changes the numbers;

re-run verify.mjs.

• Real data will differ from synthetic; synthetic skill does not transfer

numerically (proof/LIMITATIONS.md).