From raw sensor files to priced decisions

From raw sensor files to priced decisions A data-flow diagram generated by Archify. 01 / Sources 02 / Contract 03 / Features 04 / Model + policy 05 / Decide + prove Kaggle zips · 1.18M parts · 01 / Sources · 4,264 columns Kaggle zips 1.18M parts 4,264 columns Verified Parquet · rows round-trip · 02 / Contract · sha256 Verified Parquet rows round-trip sha256 Time split · first timestamp · 02 / Contract · 60 / 20 / 20 Time split first timestamp 60 / 20 / 20 baseline_v2 · route, timing, codes · 03 / Features · 445 columns baseline_v2 route, timing, codes 445 columns Sealed holdout · latest 20% · 03 / Features · hash lock Sealed holdout latest 20% hash lock LightGBM · 3-seed ensemble · 04 / Model + policy · val MCC 0.227 LightGBM 3-seed ensemble val MCC 0.227 Cost policy · $ per shift · 04 / Model + policy · band 0.008-0.258 Cost policy $ per shift band 0.008-0.258 Review console · abstain queue · 05 / Decide + prove · keyboard Review console abstain queue keyboard Confirmed labels · ship 0, scrap 1 · 05 / Decide + prove · jsonl Confirmed labels ship 0, scrap 1 jsonl Scorecard · holdout MCC 0.110 · 05 / Decide + prove · runs = 1 Scorecard holdout MCC 0.110 runs = 1 verified CSV min timestamp train + validation latest 20% feature matrix validation scores abstain band engineer decisions scored once Legend primary data policy / PII data store data flow

Contract first

  • • Every raw load verifies sha256 against the manifest
  • • Holdout sealed by hash before training begins

Dollars, not accuracy

  • • Inspect $6.50, scrap $42, field failure $1,850
  • • Policy saves ~$2,090 per shift on the holdout

Honest proof

  • • Holdout MCC 0.110 vs validation 0.227
  • • Reported beside the leaky 0.49 it refuses