AOM / Moment Exploration Stop-Go Handoff¶
Date: 2026-06-04
Source archive in the lab repo:
/home/delete/nirs4all/nirs4all-lab/_archive/studies/aom_moment_exploration_2026_06_04/README.md/home/delete/nirs4all/nirs4all-lab/_archive/studies/aom_moment_exploration_2026_06_04/campaign_timeline.md/home/delete/nirs4all/nirs4all-lab/_archive/studies/aom_moment_exploration_2026_06_04/artifact_manifest.csv/home/delete/nirs4all/nirs4all-lab/_archive/studies/aom_moment_exploration_2026_06_04/paper_map.md/home/delete/nirs4all/nirs4all-lab/_archive/studies/aom_moment_exploration_2026_06_04/paper_opportunities.csv
Versioned lab documentation mirror:
/home/delete/nirs4all/nirs4all-lab/docs/aom_moment_exploration_2026_06_04/README.md/home/delete/nirs4all/nirs4all-lab/docs/aom_moment_exploration_2026_06_04/campaign_timeline.md/home/delete/nirs4all/nirs4all-lab/docs/aom_moment_exploration_2026_06_04/artifact_manifest.csv/home/delete/nirs4all/nirs4all-lab/docs/aom_moment_exploration_2026_06_04/paper_map.md/home/delete/nirs4all/nirs4all-lab/docs/aom_moment_exploration_2026_06_04/paper_opportunities.csv
Target interpretation for nirs4all-methods:
integrate the robust AOM HPO / portfolio pieces already evidenced in the lab;
do not integrate broad GPU repeated-CV preprocessing grinding as a product method;
do not route by dataset/source/database identity;
catalog only the part that has a native public
n4m_*ABI and validation surface; keep broader Python/reference portfolio objects outside the canonical catalog until they get the same treatment.
Current Phase 7f artifacts:
docs/architecture/aom_methods_gap_analysis.mdroadmap/phase-7f-aom-robust-hpo-portfolio.mddocs/architecture/aom_robust_hpo_integration_manifest.csvdocs/architecture/aom_robust_hpo_catalog_ready_manifest.csvdocs/architecture/aom_robust_hpo_phase7f_dashboard.csvbindings/python/src/n4m/sklearn/aom_portfolio.pybindings/python/tests/test_aom_structural_policy.pybenchmarks/cross_binding/tests/test_aom_phase7f_contract.py
The current integrated product surface includes:
native
n4m_aom_robust_hpo_fit;Python
n4m.aom_robust_hpo;catalog method
aom_pop.robust_hpo;compact/wide native strict-linear preprocessing banks;
native
n4m_moments_compute,n4m_moments_subset_computeandn4m_moments_subtract;Python
n4m.momentsandn4m.moments_train_from_heldout;native
n4m_sweep_runRidge-CV mode;Python
n4m.sweep_run;CPU and CUDA-build smoke validation.
The current Python reference surface also includes:
AOMStructuralPolicyand source-free presets;AOMControlSelector;AOMTrueBankEndpointPortfolio;AOMMidPEndpointStack;AOMEndpointMarginStabilityGate.AOMFallbackBlendGate.
The lab stop-go decision is final enough to stop dataset-by-dataset
over-tuning. The first native ABI surface is now integrated. Remaining work is
to decide whether the non-native portfolio/gating objects deserve the same
product hardening, and separately to extend n4m_sweep_run with batched IKPLS,
operator descriptors and fused batched GPU preprocessing screens beyond the
current Ridge plus materialized-PLS mode if the moment substrate proves worth
industrialising.