Quality gates¶
Development is gated by quantitative, adversarially reviewed checkpoints.
All reports and reviewer verdicts are version-controlled under
gates/reports/ as public evidence.
| Gate | What it proves | Status |
|---|---|---|
| G1 | Name availability across GitHub/PyPI/HF | PASS (2026-08-04) |
| G2 | Schema covers 61/61 parameters from a 23-paper survey; round-trip; unit lint | PASS + independent review |
| G3 | Dual-implementation physics cross-validation (DA < 0.1%, WMS < 1%, 1000 pts each) | PASS + independent review |
| G4 | Instrument noise statistics inside literature envelopes (18-paper anchors) | PASS + independent review |
| G5 | Independent cold-start: install -> data -> train -> score from docs alone | see gates/reports/ |
Phase 3 extensions¶
Phase 3 engine additions (TIPS polynomial, multi-species superposition, 3f/4f demodulation) are covered by the existing G3 dual-implementation framework:
- TIPS has its own cross-validation: independent reference implementation
(
tests/reference_impl/ref_tips.py) with separately derived coefficients, verified to < 0.5% relative error against the main implementation. - Multi-species superposition:
test_multi_species_absorbance_superposition(tests/test_absorption.py) confirms Beer-Lambert additivity numerically (< 0.01% relative);test_generate_record_with_interferent(tests/test_generator.py) exercises the full generator pipeline with a CH4 target species plus an H2O interferent end to end. - 3f/4f demodulation reuses the existing
simulate_wms()harmonic machinery and the same independent reference implementation used for 1f/2f (ref_wms.py's Fourier-quadrature harmonic coefficients are generic in harmonic order, not special-cased to 1f/2f). Dedicated tests check 3f/4f output shape and physical plausibility; the G3-WMS random-point numerical cross-validation sweep itself currently still samples only 1f/2f, so full dual-implementation coverage of 3f/4f is open follow-up work rather than a completed gate pass.
No new gates were added: TIPS and multi-species superposition extend the existing G3 test suite directly, and 3f/4f demodulation shares G3's machinery and reference implementation without requiring new threshold scripts.
Design principles: reviewers run in fresh sessions with read-only access; gate thresholds cannot be changed in the PR that passes them; every claim in the anchor tables is traceable to a cited paper.
Known honest limitations: G4 validates statistical similarity to published systems, not point-wise truth; official CH4 line data must come from HITRAN via hapi (the built-in demo line list is approximate and so labeled).