opendeviationbar-py ยท feature-column discipline ยท living reference
Every feature column must be orthogonal ยท parameterless ยท agnostic (Terry msg id=428). Each axis is a runnable probe; one gate combines them into a single promote/skip verdict. This area is the living current-state reference โ the dated campaign spoke is the frozen record of how it was built.
Does this feature tell us something no existing feature already tells us?
Spearman |ฯ| matrix + h_norm per (symbol ร threshold ร slice) cell, regime-stratified across 7 epochs; ฮพโCODECโO-info cascade behind it.
orthogonality_probe.pyDoes this feature contain secret hand-tuned numbers?
Job 1: AST gate vs the constants registry (tunable knob โ BLOCK). Job 2: shared-knob promotion sweep across all candidates/families.
parameterless_probe.pyIs this ONE well-defined algorithm โ would two honest implementations agree?
Algorithm-invariance: no-dispatch scan + two-implementation convergence oracle + doc/code consistency. (NOT asset-agnostic โ provenance-corrected.)
agnostic_probe.pyOne promote/skip decision over the three probes. A tunable knob vetoes.
promote = parameterless โง agnostic โง orthogonal. Cost-ladder short-circuit: free static axes first. A feature-tunable knob vetoes everything.
three_axis_gate.pyWill a feature that is orthogonal today STILL be orthogonal tomorrow?
Leakage-gated walk-forward over 20 named regime slices: persistence 94-95%, flip-ranking AUC 0.88-0.95, ACI-calibrated intervals, FDR≤10% declared sets. Regime-conditional by evidence (ICP 0/92).
e123_v3_walk_forward.pyThe one consolidated substrate folder every probe reads from: the locked data scope, the registries, and the byte-identical rotation engine โ split into Crypto Shared Artifacts and Forex Shared Artifacts. Read-only ClickHouse on bigblack.
Autonomous logically-absurd (Popperian) attack/defend loop that hardens the rotational orthogonal probe until no attack can break it, then proves โ empirically, zero meta-evaluation โ that it is far more robust than the legacy orthogonality_probe.py on the 10 candidates (writing the 7 missing kernels first). Hardcaps: read-only ClickHouse, 2 CPU / 2 GB, no synthetic data, frozen bands. Runs on bigblack via ccmax-claude in tmux Rotation-Orthogonal-Probe-Operational-Campaign; offline SOTA in KB.md.
Candidate-sourcing expedition feeding the admission lab: harvest + provenance-verify + dedup + M0/M1 paper-screen 300-400 candidate matrices across ALL status boundaries (STABLE >=100 first), one-way handoff via FRONTIER-FEED batches. Paper-only, laptop-hosted, zero ClickHouse.
Standing certification lab for forward matrices: six-gate admission ladder (M0-M5), frozen pre-registration, self-test seal, append-only ledger + boneyard. Goal: FAIL/FRAGILE/CONDITIONAL/STABLE statuses empirically assignable on both substrates using admitted matrices only.
Cross-asset parameterless orthogonal-candidate discovery sweep (crypto + forex) toward a 100-candidate target โ the discovery surface feeding both candidate hubs.
Hardening the candidate registry + turning the manual 'already covered?' novelty check into a fast, automatic, evidence-backed catalogue-uniqueness pre-filter. Per-iteration loop record.
Full-grid forward-orthogonality evaluation at matrix scale โ the scale-up behind the Forward probe, with leakage-gated walk-forward across the named regime slices.
How the three probes + the one gate were built (the frozen origin record of this area).
promote = parameterless.PASS AND agnostic.PASS AND orthogonal.PASS3. Agnostic โ algorithm-invariant. Two correct implementations of the same feature should converge to the same answer. A column that silently dispatches between ApEn and PE based on window size is two columns wearing one name.
โ Terry, msg id=428 (verbatim). Full sources: PROVENANCE.md