Testing 3 complexity-entropy (CECP) candidate features for whether each adds information orthogonal to the live production panel — one candidate per firing, read-only across 10 regime slices × 21 (symbol, threshold) cells = 210 tests on ClickHouse. Gate: ‖ρ‖≤0.85 ∧ R²≤0.85 ∧ h_norm≥0.05 (worst-valid-cell). Scope: orthogonality (Axis-1) only — these three already PASS Axis-2 (parameterless) via PR#513 reuse. Resource-capped to 2 cores @ 200% / 2 GB, spare capacity only.
3
Candidates
1
PASS
1
WATCH
1
BAN
0
Pending
How to read this dashboard
What the loop is doing: checking, one feature at a time, whether a new market-complexity measure tells us something the features we already ship don't. A PASS means it's genuinely new everywhere we tested; WATCH means it's borderline-redundant somewhere; BAN means it duplicates an existing feature.
What "grounded" means here: a PASS is earned on the worst of 210 real-data cells, not an average — a feature has to be orthogonal in every symbol, threshold, and regime.
Each candidate page explains in plain terms what the feature measures, how it was tested, what every number means, how its verdict was reached, what resources it used, and anything worth flagging to you.
Scope: orthogonality only — no profit/signal claim is made here. Promotion to a production column is a separate operator decision after the cross-candidate Finish.
Candidates: #16 statistical-complexity (MPR) · #19 cecp-velocity · #20 binary-cecp. All three ride the ratified PR#513 ordinal embedding (m=3, τ=1) — no new project-wide constant is promoted by this loop. Empirical sibling-redundancy among them is resolved by the cross-candidate matrix (all 10 slices) at Finish.