RP002A synthetic calibration review

Date: 2026-10-09. Run: calibration-20261009-001. Synthetic analysis stress tests only; no learned models or RP002A generator observations.

Disposition

All 90 fixed Student-t cells (1,800,000 trials) and 18 percentile-bootstrap benchmark cells completed in 39.414 seconds. The preselected twofold-inflated Bonferroni replicate-mean Student-t procedure met the declared eligibility gate at n=20,30,50. No interval factor, scenario, threshold or decision rule was chosen after outcomes. Eligibility applies only to the declared synthetic fixtures. It does not establish distribution-free or actual-loss coverage.

Coverage is the fraction of independent trials in which all seven intervals contain their true means. Upper failure bounds are Clopper-Pearson with .05/90, giving simultaneous Monte Carlo statements across the fixed grid. They describe simulation sampling error, not actual-loss uncertainty. Normal, centered Gamma(shape=2) and standardized t(df=5) effects, shared-component rho=0,.5 and within-mean fractions 0,.5,1 are all retained. The uninflated method's skew-case failures demonstrate why its normal approximation cannot be assumed safe in these stress fixtures. The fixed factor two is conservative in the grid; it is not a universally validated correction.

Bootstrap benchmarks

Each cell has only 100 independent trials and 2,000 bootstrap draws over eight synthetic episodes. These are bounded comparison fixtures, not a target-data-size calibration or a method-selection gate. Coverage values consequently have substantial Monte Carlo uncertainty.

The Student-t summaries retain actual false benefit/equivalence boundary rates, center-equivalence rates and widths; no favorable cells are discarded. Bootstrap benchmark lower-endpoint violations are conservative boundary diagnostics, not full two-sided equivalence decisions, because the opposite endpoint condition is not imposed in that benchmark. Nested variance inflation and small-n percentile undercoverage have different causes. Adding inner resampling is not a general solution to percentile tails or skew. There is no claim that these benchmark frequencies are precise estimates of nominal .95 coverage.

Provenance and reproduction

Execution source: 350090247bba8d7c8cd98e4b3ccec69e80752966. Frozen pipeline SHA256: b3bd602411fd0f7269cd6b48b25d123423f6f9f849a37bb3ae5e389b20d1fd2a. Manifest SHA256: 8f121239a0f020184f80299d54bf881851fde1808f9d6da6da612b36905db63b.

Python 3.9.6, NumPy 2.0.2, SciPy 1.13.1; fixed SHA256 seed namespace calibration-1.0. The separately bound user approval is docs/research/CALIBRATION_APPROVAL.json. All 111 output hashes are preserved, including per-trial means/SDs/decision flags and every benchmark cell. scripts/check-statistical-pipeline.py verifies source/plan/approval binding, immutable files and recomputes interval flags/rates/eligibility from stored trial diagnostics without new sampling or training. Seventy software tests pass.

All raw calibration files · Frozen pipeline protocol

Next authorized step

Proceed to the predeclared five fresh target-size training replicates per world. Only their seven contrast SDs may choose the final count by the frozen rule; means cannot tune models or margins. Actual-loss assumptions remain explicit, and no scientific Claim or Cycle obligation is resolved by calibration.