›Navigation

Matrix-Evaluation for Orthogonality

A read-only evaluator + reporter for the forward-orthogonality matrix campaign. It mutates no production system: each firing answers one backlog question and attaches a trust basis to every number, then commits the finding to PR #543. Audit twin: findings/evolution/audits/2026-05-26-forward-orthogonality-prediction/.

Scope & data (operator-set): the campaign is crypto-only for now — forex is temporarily out of scope (operator directive 2026-05-30 "for now", re-confirmed 2026-06-24); it is reversible and can be reintroduced via pre-spec.md's scope-expansion path. The evaluator has read-only ClickHouse access (clickhouse-client --readonly=2). From iteration 3 onward, numbers are grounded in live read-only queries against opendeviationbar_cache.open_deviation_bars on bigblack, shown verbatim; iterations 1–2 read committed result files produced by earlier read-only runs and ran zero live queries.

‡ One number, two panel definitions (read this before the 0/91 vs 0/92 below). The crypto "STABLE-empty" result is written 0/91–0/92 because two committed artifacts count the feature panel slightly differently: full_grid_aggregate.json uses a deduplicated 91-feature panel (→ 0/91), while icp_v3_cross_regime.json and the SPIKE-2026-06-16 replication use a 92-slot panel (→ 0/92). Both return zero regime-invariant features. It is the same finding on two panel definitions, not two different runs. (The en-dash in "0/91–0/92" means "zero on both the 91- and 92-feature panels" — NOT a measured interval.)

Grounded state (already true — not re-litigated)

Orthogonal-today features persist into the next regime — crypto 95.1% (leakage gate PASS, replicated). Persistence of the conditional / regime-scoped set, NOT unconditional stability — which is empirically empty.
Who-flips ranks by today's redundancy margin: conditional AUC 0.945 crypto (matched shuffled-pairing null → ~0.5). (Forex who-flips 0.877–0.880 archived, out of scope.)
FDR-bounded set @ q=0.10 = the stamped four (aggression_ratio, bar_dispersion_entropy, bar_katz_fd, bar_petrosian_fd); PBO 0.00 crypto; deflation z 22.15 vs 1.53. CONDITIONAL / regime-scoped — not stable.
STABLE-ORTHOGONAL empirically empty (ICP 0/91–0/92, see note ‡) — but the strict exact set being empty is provably degenerate (iter 4); the relaxed question is undecided.
adv_auc ≈ 0.522 derived (matched-null). TAU = 0.85 is un-nulled policy (confirmed iter 7); 0.95 BAN ground-truth-anchored.
The live expiry monitor (G12) is now specified (iter 5) but still not built → this probe remains EVALUATING, not OPERATIONAL.
Citations re-verified 2026-06-23 (15/15 real). The 4 INCLUDE-IF matrices are literature-grounded, not data-grounded.

The 4 INCLUDE-IF matrices (literature-grounded)

#	Matrix	Bucket	Citation	Serves
1	CTM covariate-shift → WATCH	A — anytime-valid spine	arXiv:2505.04608	Q4 live expiry monitor
2	e-detectors	A — anytime-valid spine	arXiv:2203.03532	Q4 false-alarm bound
3	Nonlinear / relaxed ICP	B — STABLE-empty falsifier	arXiv:1706.08576	Q3 (is STABLE-empty a linear artifact?)
4	Group-invariance CTM	B — STABLE-empty falsifier	arXiv:2401.15461	Q3 / Q4

Plus the NP-hardness reframing (arXiv:2501.17354): recasts the 0/91–0/92 STABLE-empty result as computationally degenerate → an ε-tolerance / budget frontier (Q3).

The 6 backlog questions

Q	Question	Trust basis	Status	Iteration
1	Can features be graded on regime-invariance at all?	Type-I-controlled ICP (α=0.05)	RESOLVED-AS-SCOPED — empty result Crypto graded; result is 0/91–0/92 invariant = none (STABLE-empty), an empty finding, not a positive result. Forex resolved by descoping, not measurement.	iter 1 (superseded) · iter 3
2	Crypto: split FRAGILE vs CONDITIONAL per feature	per-feature k-of-N ICP (α=0.05)	PARTIAL. Of 91: 0 STABLE, exactly 4 certified CONDITIONAL (FDR), 87 unclassified — per-environment k-of-N is a missing measurement → operator.	iter 2
3	Is STABLE empty, or a strict-linear artifact?	G10 shuffled-null power calibration	PARTIAL. The empty exact set is provably degenerate (NP-hardness, arXiv:2501.17354). Relaxed/ε-tolerant invariance undecided: deciding nonlinear ICP (F014, arXiv:1706.08576) specified but unbuilt (no Rscript) → operator.	iter 4
4	Live "orthogonality expired" alarm with a formal false-alarm bound?	Anytime-valid / Ville bound (≤α at every t)	RESOLVED — spec delivered (build = operator) Buildable design: WATCH (2505.04608) + e-detectors (2203.03532) + group-invariance CTM (2401.15461), alarm at wealth ≥ 1/α, + E6 replay plan. Monitor UNBUILT → probe stays EVALUATING → operator builds + E6-validates.	iter 5
5	Are the flagged anomalies disqualifying? (wold_R 0.278; forex PBO 0.20)	Replication + PBO/CSCV	RESOLVED — neither disqualifies `wold_R` was already reported-not-gated → its crypto inversion changes no gate; forex PBO 0.20 is a forex watch item, parked (forex out of scope). Crypto FDR certificate untouched.	iter 6
6	Can TAU=0.85 be derived, or must it stay policy?	Ground-truth anchor (info-loss-budget)	RESOLVED — policy, not derived TAU=0.85 fails the void-band test (redundancy median 0.991, no gap) → policy. Keep the anchored `0.95 BAN`; a principled cutoff needs the info-loss-budget derivation → operator.	iter 7

Progress — evaluator pass COMPLETE (all 6 reported). Of 6 questions: 4 answered (Q1 empty result · Q4 spec delivered · Q5 disposition · Q6 honesty) and 2 partial (Q2, Q3). Nothing pending for the read-only evaluator. The only remaining work to fully close Q2 and Q3 is two operator builds, both out of the evaluator's write-surfaces (R3): (Q2) extend the ICP harness to emit per-feature k-of-10 pass counts; (Q3) install R + cran/nonlinearICP, build the relaxed-ICP (F014) harness, pair with a G10 shuffled-null power calibration, read the ε-frontier. No metric was promoted to a gate (R7); the probe stays EVALUATING (Q4 monitor specified but unbuilt).

Iterations (append-only)

Iteration 1 — Q1 (forex): originally BLOCKED (wiring defect); superseded → RESOLVED-AS-SCOPED (forex descoped).
Iteration 2 — Q2 PARTIAL: 0 STABLE; 4 certified CONDITIONAL (FDR); per-environment k-of-N for the other 87 missing.
Iteration 3 — Q1 RESOLVED-AS-SCOPED + first live read-only ClickHouse grounding (871,376 real BTCUSDT@250 bars through 2026-06-24).
Iteration 4 — Q3 PARTIAL: exact-empty provably degenerate (NP-hardness, arXiv:2501.17354); deciding nonlinear ICP unbuilt → operator.
Iteration 5 — Q4 RESOLVED (spec): live expiry monitor = WATCH + e-detectors + group-invariance CTM, anytime-valid/Ville bound, + E6 replay plan; build = operator.
Iteration 6 — Q5 RESOLVED: neither anomaly disqualifies — wold_R reported-not-gated; forex PBO 0.20 parked (out of scope).
Iteration 7 — Q6 RESOLVED: TAU=0.85 fails the void-band test → policy, not derived; keep 0.95 BAN; info-loss-budget derivation → operator.

Campaign: matrix-evaluation-for-orthogonality · PR #543 (findings/spike-2026-06-16-reconfirmation) · audit twin 2026-05-26-forward-orthogonality-prediction · read-only · resource-capped · no AI attribution · do NOT merge.