open)¶

This was a 22-idea ideation backlog from a 4-agent FE sweep. It is now ~86% resolved: every shipped item carries an inline module reference and every rejected item carries a # bench-attempt-rejected note at its call site, so the full rationale lives in the source and in git history. This file is trimmed to the residual — the three ideas that are neither shipped nor rejected — plus a short ledger of what landed, so nobody re-treads resolved ground.

Guiding lesson (still applies): across the rejected ideas the binding constraint was always the admission / selection machinery, not the richness of feature construction. The highest-confidence wins let a genuinely-good feature pass a gate it currently fails. Anything that lowers a bar must keep the order-2 maxT permutation-null floor as the outer guard and must debias both sides of any ratio consistently.

Open items (NOT started)¶

#5 — Permutation-null-calibrated prevalence bar. Replace the fixed engineered-MI-prevalence threshold with a per-run permutation-null-calibrated bar so the admission cutoff adapts to the data's null MI level. No module yet.
#10 — CMI complementarity Apriori lattice. Bottom-up Apriori-style lattice over candidate interactions gated by conditional-MI complementarity. Distinct from the shipped greedy-CMI path (_mi_greedy_cmi_fe.py, _orthogonal_cmim_fe.py); the lattice enumeration is not built.
#19 — KSG continuous-MI gate tie-breaker. Use the KSG estimator (_ksg.py, already present as a general estimator) as a tie-breaker inside the FE gate's 0.90–0.97 retained-MI band. The general estimator exists; the gate wiring does not.

Shipped this session (default-on or opt-in)¶

Order-2 Westfall-Young maxT permutation-null floor (_permutation_null.py), per-gate rejection ledger (_fe_rejection_ledger.py), occupied-K / Miller-Madow bias correction (_feature_engineering_pairs/), interaction-information ranking + routing (_interaction_information.py, opt-in), hinge / change-point basis (_hinge_basis_fe.py, opt-in), conditional-dispersion features (_extra_fe_families_dispersion.py), Haar wavelet basis (_wavelet_basis_fe.py, _wavelet_basis_fe_batched.py), cross-fold recipe stability voting (_fe_stability_vote.py), successive-halving rung schedule (_fe_rung_schedule.py), robust heavy-tail 1-D prewarp (hermite_fe/), gradient-interaction seeder (_gradient_interaction_seeder.py, opt-in), sufficient-summary early-stop (_fe_sufficient_summary.py).

Benchmarked + rejected (rationale in-code)¶

Trimmed-abs-floor, surrogate-GBM co-occurrence seeder (_surrogate_interaction_seeder.py), order-3 maxT floor (blocked on the seeder), RFF interaction pre-screen, isotonic reshaping (subsumed by spline), confidence-weighted MI under imbalance (_orthogonal_univariate_fe/), cross-fit warm-start prior. Each carries a bench-attempt-rejected note at its site so the negative result isn't re-run.