Skip to content

Every input of the design traced to its source, and the pilot read at K = 24

AJ: no number without a recorded origin, in code or in reporting. What that changed here:

  • N = 1,850 is now an artifact. scripts/training_size.py rebuilds the rule logged above (5th percentile, nearest rank, of the smaller half over 300 seeded splits, rounded down to 50): smallest 1,791, 5th percentile 1,856, median 1,988, N = 1,850 (training-size-openstack.json), as logged. admissible_partitions.py reads N from it rather than from a typed flag; rerun, the admissible list is identical and records the source (size_floor_source). partition_pilot.py reads the training size from the admissible list.
  • The simulation reads every input from the artifact that fixed it: K from the pilot's sizing, the planned 1,635 changes from decomposition-sensitivity-v3.json, N and the split reference from the admissible list, each source recorded in its output (inputs).
  • The run spread is calibrated, not chosen. The earlier grid (0.011, 0.015, 0.025, 0.035) had no recorded origin, and it was the run shift sigma_run, not the per-run spread its log table called it: a simulated run already varies with its partition and churn. sigma_run is now found by bisection so that K null runs on the pilot's 206 changes spread as much as four named points on the pilot's per-run spread: its 90% lower bound, its estimate (0.0140), the 90% upper bound K was sized on (0.0177), and its 99% upper bound. The stopping-rule table above is therefore indexed by sigma_run, not by spread.
  • Each chosen constant states its reason where it is defined (beta, K_MIN, K_MAX, the sizing confidence, the refusal share, the simulation's trial counts and churn), and the design spec carries a table of every number and its source.
  • A bug the tracing found. The regenerated partition-pilot-openstack.json had lost its list of run files: the report's runs key was overwritten by the cell's run count. Renamed run_files; the script now refuses any cell key that would overwrite a report key.

The pilot at the registered K (partition-pilot-openstack-k24.json). Runs 23 and 24 (jobs 208898, 208899, pinned at 205ba2f with the first 22) passed every outcome-neutral check. Over the first 24 admissible runs, H1 reads +0.0049 [-0.0088, +0.0185] at 97.5% and [-0.0073, +0.0166] at 95%, inconclusive. Sizing on all 24 would give K = 22; K stays 24, fixed from the 22 runs before these two were computed.