FForge Outcome Delivery
SAR Multi-Crop Acreage Estimator

Executive Evidence

← Back to outcome

Executive Evidence — SAR Multi-Crop Acreage Estimator

Standard: PROOF_STANDARD = IRS_AUDITOR Certification: PROOF_INCOMPLETE (machine copy: proof/PROOF_SCORECARD.json). Why not CERTIFIED: the customer outcome is an official competition leaderboard submission; the official dataset is unavailable in this workspace, so that outcome is a DISCLOSED_SEAM. All implemented checks pass and are reproducible.

This document answers the ten required IRS_AUDITOR questions. Every number has lineage in proof/CLAIM_EVIDENCE.json and verification-report.json.

1. What exactly is being claimed?

  • A dependency-free Round-1 pipeline: SAR feature engineering → ridge

linear-unmixing model → reconciled, schema-valid per-village acreage submission for Rice, Cotton, Maize, Bajra, Groundnut, plus an auto-detecting official-data evaluation mode.

  • On a synthetic Kharif benchmark (seed-fixed): 12/12 checks pass; MSE

15434.42 ha²; 59.1% skill vs the naive column-mean baseline; pooled R² 0.5887; total cultivated-area correlation 0.897.

2. What evidence supports each claim?

  • verification-report.json (+ .md) and the raw run copy

proof/evidence/verify.log / proof/evidence/verification-report.json.

  • proof/CLAIM_EVIDENCE.json maps every claim → source, method, command,

dataset, timestamp, artifact, status.

  • proof/EXECUTION_TRACE.json records each command, exit code, and stdout

sha256; proof/ARTIFACT_MANIFEST.json + proof/CHECKSUMS.json pin every file.

3. Can an independent engineer reproduce this claim?

Yes. proof/REPRODUCE.md gives exact commands; everything is seeded and dependency-free (Node 18+). node verify.mjs reproduces the table; node tools/forge-proof-verify.mjs --outcome delivery-package/sar-crop-acreage re-checks every checksum.

4. What assumptions were made?

  • SAR backscatter mixes linearly in power space, weighted by each cover's

area within a village → acreage is a linear-unmixing problem (motivates ridge).

  • The synthetic generator's per-crop seasonal signatures, noise model (speckle

ENL, soil-moisture, measurement), and 8 land-cover composition approximate village-level zonal means. These are modelling assumptions, not measurements.

  • Village area is known and bounds each crop area (used as a physical constraint).

5. What limitations exist?

See proof/LIMITATIONS.md (authoritative). Headline: all metrics are synthetic; no official dataset; Stage A (raster→zonal) not executed; CV estimates, not a leaderboard score.

6. What seams exist? (DISCLOSED_SEAM)

  • Official competition dataset absent → no leaderboard MSE produced.
  • Raster→zonal-statistics (Stage A) documented, not executed (needs GDAL/

rasterio/SNAP/GEE).

  • The "official" verification run used a synthetic-derived fixture

(tools/synth-to-official.mjs) — proves the path executes, not real accuracy.

7. What was actually executed?

  • node verify.mjs → 12 structural+evaluation checks on synthetic data

(deterministic). Raw output: proof/evidence/verify.log.

  • Official path on a synthetic-derived fixture → 16/16 checks + submission

emission. Raw output: proof/evidence/verify-official-fixture.{json,log}, params in proof/evidence/fixture-params.json.

8. What was inferred (not directly executed)?

  • Real-world field accuracy is inferred to be unknown — not measured. The

synthetic skill (R² 0.59) bounds implementation correctness on the modelled mixing problem, and does not transfer numerically to real data.

  • Stage A correctness on real GeoTIFFs is inferred from the documented method

and the unit-checked ingestion round-trip, not from a real run.

9. What remains unverified?

  • Any metric on real official data (no dataset).
  • Calibration/speckle/zonal code paths against real rasters.
  • Deployment, security, monitoring, load, and operational behaviour (not run,

not claimed).

10. What evidence would invalidate the claim?

  • A node verify.mjs run not yielding 12/12 or different numbers (drift/

environment difference).

  • Editing src/synth.mjs or the seed (synthetic numbers are conditional on them).
  • Treating synthetic skill as real-data accuracy (explicitly not claimed).
  • On real data: out-of-fold CV MSE failing the skill/R²/constraint thresholds in

proof/VERIFY.md.

Pre-written hostile objections and responses: proof/AUDITOR_OBJECTIONS.md.