[Ayushxsingh100]

P8b — scoring, metrics and the dev report

Branch feature/p8b, created from origin/main at 2db255b.

• This PR ends P8. No gate, mini-gate or milestone is claimed; the external review decides (C-28).
• The only gate wording here is the dictated P8a status line (verdict on PR P8a: the experiment harness (conditions, TerraPreserve loader, trials, records) #30), with that verdict's limits 1–7.
• Do not merge before the verdict.

Decisions

• Decided (authors, 13 Nov 2026), "accepted as built after external review": C-200 to C-204 and C-207.
• Proposed (P8b):
  • C-212, metric definitions metrics-draft-1;
  • C-213, the identity truth table;
  • C-214, descriptive statistics only;
  • C-215, my readings while building.

What exists

• experiments/scoring: the only code that reads TerraPreserve label fields. It checks the hash pins, refuses the test split, derives the truth of each transition, and holds the identity truth table written from D4_OPERATORS.md.
• experiments/metrics: the versioned extractor. It produces MetricRecords (Doc 13 Appendix B) and TrialOutcomes (Doc 13 §31).
• experiments/report: the ExperimentReport (Doc 13 §33) and its Markdown rendering. Command: python -m experiments.report extract|select|compare.
• Blinding contract:
  • (a) no module outside experiments/scoring names a label field;
  • (b) no execution module imports scoring, metrics or report;
  • both are shown failing on a mutant.

Dev report

• docs/P8_DEV_PILOT.md is generated from the stored records of run 2 (ec2b4fa8…); a test checks that it equals the rendering of docs/P8_DEV_PILOT.json.
• Under C: no unsafe promotion, 0 of 98 regressions missed, no truth change outside the verified set.
• These are dev numbers of a draft protocol, not findings.

Reproducibility

• Extractor: a second run wrote byte-identical files.
• Trial re-run: 20 scenarios × 4 conditions (80 trials) were re-run as a new experiment; 0 decision differences and 0 invariant-result differences against run 2.

Checks (Python 3.12.10)

• dev.py test: 4461 tests OK, 6 skipped.
• pytest: exit 0.
• dev.py lint: clean.
• Mutation checks: 12 of 12 metric mutants and 2 of 2 blinding mutants caught.
• No new dependency.

Review verdict: (link to be added)

🤖 Generated with Claude Code