How this OPD was made, and how to check it

Synthetic example. The method below is what a real OPD records. Here the units, people and dates are fictional, and the data in raw/ was generated by tools/make_example_data.py in the opd-spec repo.

1. Question. How is EC-1 built, station by station, and where does yield leak?

2. Materials. 30 units of EC-1 rev B, SKU EC1-BLK, bought at retail in two lots (Sep 2026), new in box. Serial numbers redacted to EC1-U01 to EC1-U30.

3. Equipment.

   • Micro-CT: 45 µm voxels, 150 kV (raw/scan-logs/ct-run-1.yaml). Volumes are in the gated tier.
   • Digital caliper, 0.01 mm, calibrated 2026-08-02.
   • CMM, 0.001 mm.
   • Handheld XRF.
   • 10x loupe and an ESD bench.

4. Method.

   • Teardown in seven steps, with a photo per step (raw/teardown-log.csv).
   • Measurement plan: the 12 features in parts/tolerances.yaml, one reading each per unit. Operators A and B alternate units.

5. Sample size and design.

   • Why 30 units: a standard deviation within about ±25% at 95% confidence, enough to drive Monte Carlo.
   • Factors: unit, operator, instrument. Unit order randomised. Three units CT-scanned.

6. Raw data. Everything is in raw/ and hashed in SHA256SUMS:

   • measurements.csv: 360 readings
   • eol-sample.csv: 300 end-of-line records
   • XRF spots, markings, the teardown log and a review sample

7. Analysis. python analysis/spreads.py rebuilds every derived number and fails on any mismatch:

   • each feature's mean, standard deviation and n
   • the baseline RTY from the station yields
   • the BOM roll-up

   Inference rules:

   • Vendor regions come from markings and tool marks.
   • Material comes from XRF plus the moulded resin code.
   • A station is "observed" only when tool marks or witness marks show it.

8. Results. In opd.yaml. Every value carries its source and confidence; every spread carries n.

9. Deviations and limits. Nine of 30 units showed flux on the stick module, so the stick-drift lesson rests on teardown evidence. Yields are simulated, not measured, and scrap can't be seen from the box.

10. Cost to reproduce. About 40 person-hours, plus:

    • 30 units at about 35 USD
    • CT time, about 600 USD for three units
    • caliper and CMM access

11. Pre-registration. Predicted yield leaks are in prereg/predictions.yaml, hashed in prereg/predictions.sha256 before any maker data. A later backtest compares the real line against these, blind.