DashboardThree-Axis Probes

axis 1 · empirical · operational

🎯 Orthogonal

Does this feature tell us something no existing feature already tells us? Established empirically on real bars, never from the estimator's paper. The cross-asset / cross-regime sweep lives HERE (Terry msg 429 step 3) — it is the orthogonality substrate, not the agnostic test.

Machinery

layerwhat it measureswhere
fast Spearman |ρ|monotone redundancy vs every shipped column, per cell; BAN >0.95 · WATCH 0.85–0.95 · PASSorthogonality_probe.py
h_normmass-collapse: does the quintile encoding retain any information (<0.05 = constant-like)orthogonality_probe.py
regime aggregationmin_regime_pass_rate across 7 epochs — the honest number (same-epoch slices correlate)aggregate_candidates()
W1 distribution (report-only)Wasserstein-1 to the zero ideal = mean per-cell score on [0,1]; the L1 complement of the worst-cell L∞ verdict. Never gates.campaign_summary_adapter.py
Chatterjee ξnonlinear / functional / asymmetric redundancy (ξ=1 ⇔ Y=f(X)); Lin–Han analytic nullcodec_foci_eval.py
CODEC T(a,b|rest)conditional redundancy given the rest of the panelcodec_foci_eval.py
O-informationhigher-order synergy/redundancy tripletscodec_foci_eval.py
MOWFO walk-forwardleakage-free discover-on-train / verify-on-test stability (purge 200, embargo 10)walk_forward()

Honest limits (why PASS ≠ proof yet)

Verdict contract (what the gate reads)

Two accepted evidence dialects: (1) the probe summary — PASS when min_regime_pass_rate ≥ 1.0 (holds in every regime epoch); (2) campaign/spike per-cell records via campaign_summary_adapter.py — worst-valid-cell over ‖ρ‖ + panel-R² + h_norm, stored labels re-derived from raw numbers, WATCH honest non-promotable. No evidence ⇒ PENDING, never a faked pass.

Distribution view — (worst, W1) (2026-06-13, report-only)

The verdict collapses 208 cells to the single worst (an L∞ view: "redundant somewhere?"). Beside it we now report Wasserstein-1 to the zero-redundancy ideal — the L1 complement: "redundant everywhere?". On [0,1] scores, W1 reduces exactly to the mean of the per-cell scores (Vallender 1974) — auditable, computed through scipy and identity-asserted.

Same worst, different story (real entropy-spike data):
bar_svd_entropy — worst |ρ| 0.97 / W1 0.53 → a regime-local spike.
bar_permutation_entropy — worst |ρ| 0.97 / W1 0.77 → redundant broadly.
Worst-cell alone cannot separate them; W1 does. Both stay BAN — W1 never gates, it only explains the worst.

Deep dive: AXIS-1-ORTHOGONAL.md · AXIS-1-W1-DISTRIBUTION.md