Orthogonal features for BTCUSD — candidate set with two-sided provenance

verdict.md — orthogonal features for BTCUSD (Phase 1, post-G9)

Date: 2026-05-09 Status: Phase 1 evidence complete with G9 adversarial audit applied. 9 covered / 55 not-covered (incl. 1 source-blocked) = 64 candidates. Methodology completeness items (firecrawl, exa, crucible plugin, HTML+SVG presentation) deferred to subsequent shifts per operator direction.

Top-line summary

The deliverable's exit-criterion content requirements are met:

• Candidate list: 64 candidates with two-sided provenance (arXiv/DOI + FOSS commit SHA where applicable)
• Coverage classification: 4-source novelty check + adversarial pair on Audit Verdicts + G9 adversarial audit on the deliverable itself
• "Why not already covered": per-row evidence with file:line / audit-folder / list reference

Reproducible from this folder:

• RESEARCH-DIGEST.md — canonical 64-row table
• VERIFICATION-VERDICT.md — citation + FOSS verification with both Layer-1 (14 modern papers) and Layer-2 (21 foundational papers) explicit
• BASELINE-RULED-OUT.md — Phase-1 baseline (existing schema + ruled-out audits + settled positives + doctrine)
• DUPLICATE-ELIMINATION.md — parallel-session reconciliation (4 papers transferred to other sessions, 0 candidates dropped)
• This verdict.md — G9 audit results

Challenged and held — G9 audit (id=758/759 policy)

Two Explore agents were dispatched: one Defender (confirm deliverable claims), one Attacker (find flaws). Five attacker findings + one defender NOT-CONFIRMED claim were surfaced. All findings were independently re-verified by direct file reads before any deliverable change was applied. One attacker finding was rejected on verification.

Challenge A1 — #1 Shannon Entropy was misclassified as covered (HIGH severity)

Attacker claim: The cited evidence — shannon_entropy_signs_bar in 2026-04-27-cost-realistic-cadence-harness/CH_COLUMN_SAFETY_MANIFEST.md:134 — is from a forex audit (fxview_cache.forex_bars), not the crypto schema. Per Terry's id=209 ("opendev-py = crypto only"), forex is out-of-scope.

Independent verification:

• CH_COLUMN_SAFETY_MANIFEST.md line 1 verbatim: # ClickHouse Column Safety Manifest — fxview_cache.forex_bars (explicit forex schema)
• 2026-04-27-cost-realistic-cadence-harness/CLAUDE.md confirms forex iter-21b harness audit
• grep "shannon" opendeviationbar-py/python/opendeviationbar/clickhouse/schema.sql → NO MATCH (no shannon column in crypto schema)
• grep "shannon" crates/opendeviationbar-core/src/types.rs → NO MATCH

Disposition: ACCEPTED. Row #1 reclassified from covered to not-covered.

Why this matters: The previous adversarial pair on Audit Verdicts (run earlier in this session) had classified #1 as covered. That pair's output was wrong on this candidate. G9 caught the cross-asset scope error.

Challenge A2 — Coverage math ambiguity (MEDIUM severity)

Attacker claim: Status line "10 covered / 54 not-covered / 1 source-blocked" sums to 65, but only 64 candidates exist.

Independent verification: The math is ambiguous, not strictly wrong. #64 is BOTH not-covered AND source-blocked (the source-blocked is a sub-flag of not-covered, not a separate bucket). However, the wording reads as three buckets summing to 65.

Disposition: ACCEPTED — wording clarified. Status line now reads "9 covered + 55 not-covered = 64 total, where 1 of the not-covered is source-blocked (#64)."

Challenge A3 — Foundational verification claim incompleteness (MEDIUM severity)

Attacker claim: RESEARCH-DIGEST claims "21 of 21 foundational papers got verified DOIs" but VERIFICATION-VERDICT only shows 14 entries.

Independent verification: The 14 entries in VERIFICATION-VERDICT.md were Layer-1 (deep-research bibliography), not Layer-2 (foundational anchors). The 21 foundational paper resolutions WERE performed (OpenAlex batch run on 2026-05-09) but the resolution table was not committed to the audit folder.

Disposition: ACCEPTED — gap closed. Added a new subsection to VERIFICATION-VERDICT.md: "Layer-2 foundational paper resolution (OpenAlex API, 2026-05-09)" with the full 21-row resolution table, year-validation, and OA-URL status per row.

Challenge A4 — Year-mismatch count understated (LOW severity)

Attacker claim: RESEARCH-DIGEST claims "1 year-mismatch caught + corrected" but VERIFICATION-VERDICT defects table shows 3 year corrections.

Independent verification: The "1 year-mismatch" claim refers to the Layer-2 OpenAlex foundational batch (Di Matteo 2003 → 2004 correction). The 3 year corrections in VERIFICATION-VERDICT defects (Gemini #2, #9, #10) are Layer-1 deep-research-paper defects — a separate batch of work.

The attacker conflated Layer-1 and Layer-2.

Disposition: REJECTED on independent verification. The "1 year-mismatch" claim is technically correct for the Layer-2 batch. No change applied.

Challenge A5 — Coverage totals inconsistency (MEDIUM severity)

Disposition: Resolved by Challenge A2's wording fix.

Defender NOT-CONFIRMED #8 — Adversarial pair on Audit Verdicts findings not in committed evidence

Defender concern: Methodology Step 6 of RESEARCH-DIGEST claims an adversarial pair on Audit Verdicts produced 1 confirmed hit + 10 over-attributions, but verdict.md was a placeholder, so the per-claim findings weren't documented.

Disposition: ACCEPTED — partially. This verdict.md (which you are now reading) documents both the prior adversarial pair's outcome AND the current G9 audit. The 10 over-attributions from the prior pair are summarized below:

The prior adversarial pair on Audit Verdicts (earlier in this session) caught 11 attacker claims, of which 1 was real (originally classified as #1 Shannon Entropy covered — now overturned by G9) and 10 were over-attributions:

• #2 ApEn (research mention only, not tested in 12-trigger ORTHOEX)
• #3 SampEn (research mention only)
• #9 Spectral Entropy (Wiener-entropy conflation, mathematically distinct)
• #21 OPN (plain ordinal-patterns conflation)
• #22 Hurst R/S (DFA conflation; later moved to covered via the Hurst crate evidence)
• #45 Higuchi FD (proposed but not in 12-trigger ORTHOEX)
• #49 Takens embedding (methodology mention)
• #50 Bivariate Shannon TE (research mention in meta-labeling note)
• #52 Multivariate TE (research mention in R16)
• #57 TE BTC-equities (methodologically related to K5 OFI lag, not the same method)

These 10 over-attributions were rejected via direct file reads at the time and remain rejected after G9 review.

Net effect on the deliverable

The reclassification of #1 Shannon Entropy adds 1 candidate to the orthogonal pool — a net positive for the deliverable's substantive content.

Held — what survived G9 unchanged

• 64-candidate list scope (Gemini DR + Claude DR consolidation)
• 53 candidates with FOSS commit SHA, validated by git ls-remote (spot-checks confirmed antropy@dfbe688, OPyN@0c1de30, mfbm@f7b683c)
• 4-source novelty check methodology (Production Manifest, Ban List, Audit Verdicts, Evolutionary Ledger — all properly closed with structural reads)
• Coverage attribution for the remaining 9 covered candidates (#4, #13, #16, #18, #20, #22, #23, #28, #38)
• Layer-1 citation verification (14 modern papers + 7 defects caught and corrected)
• Layer-2 citation verification (21 foundational papers with OpenAlex DOIs, now committed to VERIFICATION-VERDICT.md)
• SWH/GitHub search outcome (2 new FOSS implementations found: #21 OPN, #41 mvBm; 10 confirmed no-public-implementation)
• DUPLICATE-ELIMINATION reconciliation against parallel sessions
• Honest scoping of orthogonality confidence (HIGH on novelty, NOT MEASURED on empirical orthogonality — Phase 2 work)

Methodology lessons recorded

1. Agent outputs require independent verification before deliverable changes. Both adversarial pairs in this session (the earlier one on Audit Verdicts, and this G9 pair) produced over-attributions: the earlier one classified #1 Shannon as covered (wrong); this G9 attacker over-claimed on year-mismatch count. Verifying agent claims against source files caught both errors before commit.

1. Cross-asset scope errors are easy to miss. The forex/crypto schema distinction is project-internal context (Terry's id=209). The original adversarial pair didn't catch the forex-side CH_COLUMN_SAFETY_MANIFEST.md provenance for #1; G9 caught it on direct re-reading. Future audits should explicitly check schema scope per coverage attribution.

1. Documentation must match committed evidence. Methodology claims in the canonical digest must reference committed-file citations, not session-conversation context. This was the gap behind Challenge A3 — the foundational paper resolutions existed in chat output but weren't in the audit folder until G9 surfaced the gap.

Outstanding work (deferred — out of scope for current closure)

These items strengthen the deliverable per Terry's broader mandates but are NOT blocking the exit criterion:

Recommended next step

Commit the G9 audit results (this verdict.md + the 3 already-applied edits in RESEARCH-DIGEST.md and VERIFICATION-VERDICT.md) as a single follow-up commit on the findings/orthogonal-features-btcusd branch.

After that, the deliverable is at a defensible stopping point for Phase 1. Operator decides next: continue with Path B items above, push branch + open PR for Terry's review, or pause.

Post-call methodology shift (2026-05-12)

After this verdict was first written, Terry's call (2026-05-12) clarified that the orthogonality test he applies is a 3-axis framework (ORTHOGONAL × PARAMETERLESS × AGNOSTIC) per his msg 428 rubric — not just the binary coverage-novelty check used in Phase 1. The original Phase 1 deliverable (PR #493) procedurally met the exit criterion's text requirements but did not engage with the deeper statistical orthogonality structure Terry expects.

In response, FRAMEWORK.md was authored as an 8-level per-candidate analysis structure addressing all 24 of Terry's standing instructions. The framework is AI-drafted and AWAITING Terry's accuracy review per his call directive ("You will send me a list of questions. I will tell you if those questions are accurate or not"). It must not be applied to candidates until Terry confirms.

Once approved, per-candidate FRAMEWORK.md application outputs will land in candidates/ and replace this verdict's "Held" + "Outstanding work" sections with concrete per-candidate KEPT/DROPPED/DEFERRED verdicts. That output is the input to Phase 2.

Phase 2 — Empirical spike (2026-05-27, COMPLETE — revised 2026-05-29)

Status: Spike folder spike_2026-05-27/ — 5 of 5 candidates COMPLETE + cross-candidate matrix run COMPLETE + strict-manifest verification COMPLETE. Final slate after cluster resolution + parameterless audit: 2 PROMOTE (Petrosian #43, Katz #44), 1 DEFERRED (SVD Entropy #10), 2 DROP (Hjorth Mobility #14, Hjorth Complexity #15). The 5×5 cross-candidate matrix run revealed a previously-undetected 3-way Hjorth-Katz sibling cluster (Mobility ↔ Complexity ρ −0.991 median; Mobility ↔ Katz +0.821 median, +0.927 max; Complexity ↔ Katz −0.819 median, −0.921 max). Under operator-direction-2026-05-27 empirical-evidence-first tiebreaker applied across the 3-way cluster, Katz wins on all three measurements (lowest max \|ρ\| 0.63, lowest max R² 0.20, lowest LOO R² 0.70 of the three). Authored under operator-authorized gate relaxation (same precedent as candidate #65 persistence-landscape pilot); the AI-drafted FRAMEWORK.md remains AWAITING Terry's accuracy review, but the Phase 2 work proceeds per operator direction in the meantime, with explicit attribution of which checks are Terry-mandated vs operator-extended.

Two preliminary verdicts from 2026-05-27 were REVISED before column promotion:

1. Hjorth sibling pair (#14 ↔ #15) preliminary verdict was REVERSED by the cross-candidate matrix — the pair turned out to be part of a 3-way cluster with Katz, so Katz wins under the same empirical-evidence-first tiebreaker; both Hjorth candidates drop, Katz promotes. See spike_2026-05-27/CROSS_CANDIDATE_FINDINGS.md for the matrix evidence.

1. SVD Entropy (#10) preliminary PROMOTE was REVISED to DEFERRED on 2026-05-29 after operator-direction-3 pushback exposed that the prior "Axis-2 parameterless by analogy to PE" justification does not survive the empirical codebase audit. The candidate has 3 formula parameters (order=3, delay=1, normalize=False) that are real per-feature knobs — PE's m=3 itself is empirically a hardcoded magic number, not a project-wide constant by any reasonable reading of Terry's _q1000 precedent (which has typed units, naming-convention suffix on dozens of features, test enforcement, and migration infrastructure). Renaming the magic numbers to a constants file would be cosmetic without applying them as a genuine family-wide convention, and we have only evaluated TWO Bandt-Pompe-family features so far (PE + SVD) — too small a sample to justify the family-wide commitment. DEFERRED keeps the candidate alive in the pipeline; revisit once ≥ 4-5 more entropy-family candidates have been evaluated so the family-convention decision can be made well-informed. The candidate's empirical orthogonality (208/208 PASS, LOO R² 0.45 independent of cluster) is documented but not acted on yet. See candidates/10-bar-svd-entropy.md for the full deferral rationale + trigger conditions for revisit.

Pre-column-promotion verification (2026-05-29) — operator's "what else to verify before promoting columns?" question triggered Check A (strict-manifest verification). Result: 53/53 manifest features were present in the spike's 100-column comparison set (no membership gap); the verdict max-|ρ| numbers hold under strict Terry-mandate re-interpretation; per-bar-set effective coverage corrected from misleading "20–55" framing to "median 20, max 55." See spike_2026-05-27/CHECK_A_RESULTS.md for the full check.

Methodology — what is Terry-mandated, what is operator-extended

Full attribution in spike_2026-05-27/PROVENANCE.md. Summary:

Candidate-by-candidate progress

Final slate (revised 2026-05-29): 2 PROMOTE (#43 Petrosian, #44 Katz), 1 DEFERRED (#10 SVD Entropy), 2 DROP (#14 Mobility, #15 Complexity).

Verdict taxonomy used in this audit

• PROMOTE: all 3 axes PASS by Terry's strict reading; ready for feature_manifest.toml landing in the column-promotion PR
• DEFERRED: empirically orthogonal but with a structural ambiguity (e.g., per-feature parameters where convention status is currently under-evidenced); kept alive in pipeline; revisit when the structural question can be resolved with better evidence
• DROP: either failed empirical orthogonality, OR passes empirically but is redundant against a stronger sibling (cluster resolution)

Spike scope (2026-05-27)

• Substrate: opendeviationbar_cache.open_deviation_bars on bigblack, read-only SELECT only, no FINAL modifier (crypto convention)
• Slices: 10 of 24 (slice01, 03, 05, 07, 09, 12, 14, 17, 20, 21) covering 8 distinct market regimes from 2021 H1 through 2026 H1 (COVID recovery → 2021 bull → 2022 top-turn → FTX collapse → 2023 recovery → 2024 ETF era → 2025 → current)
• Symbols: 6 crypto (BTCUSDT, ETHUSDT, BNBUSDT, LTCUSDT, ADAUSDT, XRPUSDT)
• Threshold combos: 21 unique (symbol, threshold) pairs after dropping LTC@100 (broken in CH — only 31 bars)
• Known empty cells: ETHUSDT@100 in slice01 and slice03 — auto-skipped (verified 2026-05-26)
• Total expected bar-set evaluations per candidate: 208 (210 − 2 known-empty)
• Comparison set: 100 columns from live CH (117 total minus 17 unambiguous non-features: keys, metadata, trade-IDs, gap flags, boolean flags). Per-bar-set effective comparison set 20–55 columns after >80% NaN drop, varying by symbol (spread/quote plugin columns are often all-NULL on alt symbols).
• Compute footprint per candidate: ~10 min wall-clock across all 10 slices, ~6.6 GB peak RAM, 1–2 CPU cores

Forward work after this batch

• Generalize the probe into a parameterized library function accepting (name, fn, kwargs) tuples — so future audits register a new candidate with a single-line entry rather than a code edit.
• Promote the generalized probe to findings/evolution/shared_data/probes/ for team-wide reuse — orthogonality testing across the team starts from this scaffold instead of re-deriving Terry's pattern from the patterns-repo audit folder.

What is NOT in scope of this Phase 2

• Edge-magnitude backtest (TAMRS-stack metric quantification per FRAMEWORK.md Level 4) — deferred to a separate downstream phase
• Forex side — explicitly out of scope per Terry's id=209 (opendeviationbar-py is crypto-only)
• Mutual-Information / non-linear redundancy check — Hjorth Mobility's combinatorial R² PASS makes this unnecessary for the first candidate; revisit per-candidate if a future candidate passes R² but seems suspicious of non-linear redundancy
• Cross-asset KS test on pooled-quintile-encoded raw candidate values — operator chose not to re-query CH; existing per-symbol PASS-rate evidence is operationally stronger for the question that matters (does the feature work as orthogonal across assets)

Challenged and held — Phase-2 G9 adversarial audit (2026-05-27)

Two Explore agents were dispatched before the Phase-2 PR was opened, mirroring the Phase-1 G9 audit pattern (id=758/759 policy). One Attacker (find flaws), one Defender (independently confirm claims). Both ran read-only against the same deliverable scope: verdict.md Phase-2 section + the 5 candidate verdict files in candidates/ + the entire spike_2026-05-27/ folder.

Defender result — 11 of 11 claims CONFIRMED

Defender summary: "All 11 claims CONFIRMED. No logical gaps identified. The deliverable's empirical claims are well-supported by cited evidence, numerical values are internally consistent across all documents, code implements the stated methodology exactly, and all operator extensions are explicitly labeled as such. The spike is defensible on both methodological and empirical grounds."

Attacker result — 7 findings, 0 severe

Total: 0 CRITICAL, 1 HIGH, 4 MEDIUM, 2 LOW. Attacker's own bottom-line: "All findings are correctable via edit without changing any empirical verdict."

Each finding was independently re-verified against the cited source before disposition was decided.

Net effect of G9

Held — what survived G9 unchanged

• The 3-way Hjorth-Katz sibling cluster discovery (Mobility ↔ Complexity ρ −0.991; Mobility ↔ Katz +0.821 median, +0.927 max; Complexity ↔ Katz −0.819 median, −0.921 max)
• Katz as cluster winner on max \|ρ\| 0.63, max R² 0.20, LOO R² 0.696 (≈ 0.70)
• Petrosian and SVD as independent of the cluster (LOO R² ≈ 0.45 each)
• The reversal of the preliminary "Complexity PROMOTE" decision in favor of the cluster-aware "Katz PROMOTE"
• All Terry-mandated citations at commit d74f75e1
• Read-only CH access; 2-core CPU budget; 208/208 cell coverage
• The operator-extended methodology labeled-as-such throughout (combinatorial R², empirical agnostic substitute, sibling-cluster tiebreaker)

Post-merge follow-up — Check A + SVD-deferral + parameterless audit (2026-05-29)

Operator's pre-column-promotion questions surfaced three additional fixes after the Phase-2 PR (#503) merged. These are documented here for the audit trail:

A. Strict-manifest verification (Check A)

Question: do we cover all the manifest features Terry's BANLIST rule should be applied against?

Method: read-only intersection of feature_manifest.toml (53 entries) against the spike's 100-column comparison set (live CH minus 17-col EXCLUDE list), plus per-bar-set effective-coverage analysis and strict-manifest max-\|ρ\| re-computation from the existing top-5 data.

Result:

• Membership: 53/53 manifest features in the comparison set ✅ (no gap)
• Per-bar-set effective coverage (corrected): median 20 columns, max 55 — NOT the misleading "20–55 effective" framing in the original verdict files
• Strict max \|ρ\| = loose max \|ρ\| within rounding for all 3 candidates that PROMOTE/DEFERRED in the slate (Petrosian 0.701, Katz 0.635, SVD 0.743). Verdict numbers unchanged.
• 29-30 of 53 manifest features never appeared in any top-5 for the 3 PROMOTE/DEFER candidates — likely systematic NaN-filter dropouts (mostly intra_* family). Not blocking; flagged as forward work (focused re-run with relaxed NaN filter on max-coverage cells).

Full evidence: spike_2026-05-27/CHECK_A_RESULTS.md.

B. SVD Entropy verdict revision (PROMOTE → DEFERRED)

Trigger: operator-direction-3 pushback (2026-05-29) on the prior "Axis-2 PASS by analogy to PE" claim.

Codebase audit findings (empirical, read-only):

• PE's m=3 is encoded as a hardcoded magic number (literal 3 at every call site, e.g., crates/opendeviationbar-core/src/intrabar/features.rs:307); the optimized variants bake m=3 into function NAMES without a named constant linking them
• PE's τ=1 is not even an exposed parameter — the embedding implicitly uses consecutive prices
• Zero matches in production code for EMBEDDING|PE_ORDER|SVD_ORDER|BANDT_POMPE|ORDINAL_ORDER
• Terry's _q1000 precedent for a true project-wide constant has typed units (FeatureValueQ1000), naming-convention suffix on dozens of features, pre-commit test enforcement, migration infrastructure (PHASE-23-A-MIGRATION), and documented policy — none of which exist for Bandt-Pompe m=3, τ=1
• Bandt & Pompe (2002) explicitly recommend m ∈ {3, 4, 5, 6, 7}; m=3 is the minimum-useful choice, not a canonical privileged value

Disposition: REVISED preliminary PROMOTE to DEFERRED. The candidate has 3 real per-feature knobs filled in by borrowing PE's hardcoded magic numbers — this does not satisfy Terry's "uniform _q1000 = project constant" criterion under any defensible reading. Renaming the magic numbers to a constants file would be cosmetic without applying them as a genuine family-wide convention, and a sample of only TWO Bandt-Pompe-family features (PE + SVD) is too small to commit to the family-wide convention well.

Net effect: final slate 2 PROMOTE + 1 DEFERRED + 2 DROP (was 3 PROMOTE + 2 DROP).

C. Parameterless re-confirmation for the 2 PROMOTE candidates

Independent re-check confirmed Petrosian (#43) and Katz (#44) have zero per-feature parameters by direct formula inspection:

• Petrosian FD: log10(n) / (log10(n) + log10(n / (n + 0.4 * N_sign_changes))) — the 0.4 is part of the canonical Petrosian (1995) formula, same status as π in a Fourier transform. Window = LOOKBACK_COUNT project-wide constant.
• Katz FD: log10(n) / (log10(n) + log10(d/L)) — n, L, d are derived from input, not chosen. Window = LOOKBACK_COUNT project-wide constant.

Both pass Axis-2 by direct inspection with zero ambiguity. The DEFERRED status of SVD does NOT affect these two; their PROMOTE verdicts stand and they proceed to the manifest in the upcoming column-promotion PR.

Phase 2b — Entropy-family loop (2026-06-02)

Autonomous /loop (hybrid: laptop drives, bigblack computes) swept the 8 remaining Entropy-family candidates (210 cells each × 8 regime slices), then ran the cross-candidate matrix including the deferred SVD (#10) and the landed Petrosian/Katz (#43/#44) — the redundancy-vs-landed check. Full detail: spike_2026-06-02-entropy/ (CROSS_CANDIDATE_FINDINGS.md, CROSS_CANDIDATE_VS_LANDED.md, AGNOSTIC_AXIS3.md, PROMOTION_PLAN.md); per-candidate verdicts in candidates/ (01,02,03,04,05,09,11,12).

Result — 3 PROMOTE / 2 CONDITIONAL / 3 DROP:

• 🟢 PROMOTE: bar_dispersion_entropy (#11, max |ρ| 0.66, independent), bar_svd_entropy (#10 — deferral re-opened: independent of everything incl. landed Petrosian at ρ 0.30), bar_sample_entropy (#3, template-cluster rep over ApEn).
• 🟡 CONDITIONAL: bar_spectral_entropy (#9 — Axis-2 sf ruling), bar_shannon_entropy (#1 — Axis-1 marginal + binning).
• 🔴 DROP: bar_approximate_entropy (#2, SampEn sibling ρ 0.965), bar_weighted_permutation_entropy (#5, PE sibling ρ 0.897), bar_fisher_shannon (#12, PE sibling ρ 0.948).
• ⚠ permutation (#4) DEMOTED vs the loop's first read: borderline-redundant with the already-landed Petrosian (ρ 0.836); SVD covers its embedding space independently.

Axis status: all 3 promotes are Axis-1 ✅ (orthogonal vs existing + landed) and Axis-3 ✅ (100% per-symbol PASS, see AGNOSTIC_AXIS3.md); they clear Axis-2 only via project-wide constants (the _q1000 route) — implementation gated behind the constants PR per PROMOTION_PLAN.md. ClickHouse untouched (read-only); promotion follows PR → review → merge → deploy → schema.

Phase 2c — Fractal / multifractal / chaos slate: Axis-2 (parameterless) HELD + corpus census (2026-06-09)

The overnight /loop fractal/chaos orthogonality spike (spike_2026-06-05-fractal-chaos/, driver + compute on bigblack odb-fractal-chaos-loop; live pages hosted-only at ~/sites/2026-06-06-crypto-fractal-multifractal-chaos-10-candidate-orthogonality-evaluation/) swept 10 fractal / multifractal / chaos / complexity candidates on Axis-1 only (orthogonality, 21 (symbol,threshold) cells × 10 regime slices). That loop explicitly scoped Axis-2 and Axis-3 out. This entry records the operator Axis-2 (parameterless) ruling for the whole slate, plus a full evaluated-vs-pending census of the candidate corpus.

> Provenance note: the fractal/chaos slate lives on branch spike/fractal-chaos-followup + bigblack (odb-fractal-chaos-loop) + the hosted dashboard; it is not yet merged to main. It is cited here as external evidence pending its integration PR. Line refs below are to spike_2026-06-05-fractal-chaos/orthogonal_candidates_v2.py (registry SSoT).

A. Axis-2 (parameterless) — all 10 FAIL → HELD

Rule (gating; SSoT docs/audits/3-AXIS-DISCIPLINE-TEMPLATE.md:47-56, restated FRAMEWORK.md:446-457): a feature passes Axis-2 only if every numeric literal either derives from first principles or is a locked/canonical/project-wide constant (Petrosian's 0.4, Katz n/L/d, dispersion c=6/m=2/d=1, Terry's _q1000). Free per-feature tunable knobs FAIL. FRAMEWORK.md already pre-lists DFA/MF-DFA scale ranges and Lyapunov embedding among the Axis-2 fails (:450-457).

Finding: every one of the 10 slate candidates carries ≥1 free per-feature knob → all 10 FAIL Axis-2 — including the 6 that cleared Axis-1. So the slate's promotion is HELD (hard gate: fail any axis → env-flag-off / deprecate, never quietly retained — :30-31, :136). The precedent kill for the same ruling is lookback_permutation_entropy + lookback_hurst (PR #448, column_comments.py:271-293, active=False).

The shared input window LOOKBACK_COUNT=200 (:34) is not a per-feature knob — it is the project-wide rolling window every bar feature already uses — so it does not count against Axis-2.

Registry-metadata caveat: #21 and #17 are tagged parameterless in the _REGISTRY w200 metadata (:400, :408). That tag refers to amplitude-scale invariance (ordinal methods are rank-based), not absence of embedding knobs — both still pass explicit dx=3, taux=1, so both FAIL Axis-2 in the magic-number sense.

Consequence (HELD — remediation path). None of the 10 enters the implementation loop as-is. To clear Axis-2 a knob must be either (a) derived from first principles, (b) locked as a justified feature-local convention constant (the bar_dispersion_entropy c/m/d precedent + the #513 integer-only SSoT pattern), or (c) ratified as a project-wide constant (the _q1000 route) — then Axis-2 is re-verdicted. This supersedes the earlier working assumption that the 6 Axis-1 survivors (or the redundancy-reduced 5, after the dh_abs_returns ↔ dh_vol_increments ρ≈0.96 cross-candidate duplication) were promotion-ready: they are Axis-1-clear but Axis-2-blocked.

B. Corpus census — evaluated vs pending (2026-06-09)

Reconciled against the canonical RESEARCH-DIGEST.md (64 rows) + the #65 persistence-landscape pilot = 65 total. 24 evaluated · 41 pending · 0 unmatched (arithmetic verified: union covers #1–#65, zero overlap).

Evaluated (24) — Axis-1 orthogonality run on real ClickHouse data:

• Entropy spike (2026-06-02): #1, #2, #3, #4, #5, #9, #11, #12 (+ #10 re-opened)
• Fractal/Hjorth spike (2026-05-27): #10, #14, #15, #43, #44
• Fractal/multifractal/chaos slate (2026-06-05): #17, #21, #24, #25, #34, #35, #36, #45, #46, #47
• Phase-1 pilot: #65
• Landed in production: #43 Petrosian + #44 Katz (PR #509), #11 Dispersion (PR #515).

Pending (41) — input set for the next evaluation loop:

Flag (#25 / #28): the slate labels its width metric bar_mfdfa_width as id #25, but digest #25 = "MF-DFA spectrum h(q)" while the spectrum width Δα is digest #28. It is bound to #25 per the explicit eval id (leaving #28 in pending). If the spike in fact measured Δα width, swap #25↔#28 when the slate merges — this does not change the 24/41 split.

Phase 2d — Axis-2 remediation: shared-knob constant promotion + next-loop plan (2026-06-09)

Follow-up to the Phase 2c HOLD: where multiple family members use the same knob at the same value, that value is a family convention (not a magic number) and can be ratified as a project-wide constant — the route that clears Axis-2 without per-feature tuning. A family-by-family scan (with an adversarial literal-same-value check) was run across the 10 failed + 41 pending candidates. Full analysis, promotable-constant list, family-usage counts, candidate loop list, and the two-stage loop plan: AXIS2-REMEDIATION-AND-LOOP-PLAN.md.

Verdict (operator decision pending):

• Reuse existing #513 ORDINAL_M=3/ORDINAL_TAU=1 — 9 ordinal/CECP members (#21,#17,#16,#18,#19,#20,#8,#4,#5); fully clears #21/#17.
• Promote 2 new structural-integer constants via the #513 pattern: MFDFA_DETREND_ORDER=1 (17 members) and TAKENS_EMBEDDING_TAU=1 (4 members).
• Harmonize-then-promote (locked-convention route, not parameterless): MF-DFA q-grid + scale schedule (one operator call unblocks the ~11-member MF-DFA block).
• Stays feature-local: higuchi kmax, gen-hurst qvals, chaos emb_dim/fit, SampEn m. Unpromotable: the TE/copula/PID slate (#50–#64) — dual-series regime, not a tunable literal.
• Loop: Stage 0 promotes the constants (so candidates are parameterless-by-construction) → Stage 1 sweeps the 23 single-series-testable pending candidates (18 are out of scope — need a second series), with the Axis-2 gate enforced up front and a full all-10-slice cross-candidate Finish.

Phase 2e — CECP loop scoping + parameterless verification (2026-06-10)

Scoping the next orthogonality loop from the ordinal/CECP family — the candidates fully covered by the existing PR #513 ordinal-embedding constant. Three go into the loop; one is dropped by definition.

A. Parameterless verification of the loop candidates — all PASS (via #513 reuse)

Verified each loop candidate's full knob set. Correction to the working assumption: these do NOT fail Axis-2 — they PASS. Their only structural knob is the Bandt–Pompe ordinal embedding (m=3, τ=1), which is the already-ratified project-wide constant from PR #513. Post-#513, a feature whose sole knob is that embedding clears Axis-2 by reuse (no new promotion) — this is precisely the deferral that blocked SVD-entropy #10 (candidates/10-bar-svd-entropy.md), now resolved for the ordinal family.

Caveat (#20): Pinto-2025 (arXiv:2504.01974) could not be fetched at write time; the standard binary-CECP binarization is sign-of-increment (up/down), which is parameter-free. If the faithful build instead uses a tunable threshold, #20 would FAIL Axis-2 and be held — to be confirmed at authoring.

B. #18 dropped — redundant by definition (not by measurement)

#18 "CECP coordinates (PE, C)" is excluded from the loop: it is the pair (permutation entropy #4 + statistical complexity #16) — both already in the corpus — so it introduces no new computation. This is definitional redundancy (known a priori from the formula), distinct from empirical sibling-redundancy (e.g. dh_abs ↔ dh_vol ρ≈0.96), which is only found via cross-candidate testing. Full per-candidate record: candidates/18-cecp-coordinates.md.

C. Loop scope + cross-candidate testing

Loop = #16, #19, #20 — Axis-1 orthogonality sweep on the existing probe (21 (symbol,threshold) cells × 10 slices). The Finish step runs cross-candidate testing across all 10 slices to catch empirical siblings among them and against the landed panel — and the prior fractal/chaos Finish (which covered only 5 of 10 slices) should be completed to the same standard.

Phase 2f — CECP loop results + adversarial verdict review (2026-06-10)

The 3-candidate CECP orthogonality loop ran on bigblack (spike_2026-06-10-cecp-statistical-complexity, 2-core / 2 GB / spare-only / read-only ClickHouse, 210 cells = 21 (symbol,threshold) × 10 slices, all parameterless via PR#513 reuse). Per-candidate records: candidates/16-statistical-complexity.md, candidates/19-cecp-velocity.md, candidates/20-binary-cecp.md.

Results

Adversarial review (5-agent workflow + independent compute) — what the verdicts can be trusted on

• Measurement machinery is clean (the decisive in-scope check): trailing window (no look-ahead), per-cell independent fetch (no cross-slice leakage), candidate index-aligned to the panel, Spearman pairwise-complete; all three verdicts reproduce bit-for-bit from raw JSONL. No verdict-flipping bug. The orthogonality verdicts are correctly computed.
• BAN/WATCH are genuine — stat_complexity ↔ Petrosian (0.969) and binary ↔ Katz (0.937) are real monotone ties (both reparametrise path-irregularity / sign-run structure); they survive the de-overlap control. This validates that the probe catches real redundancy.
• velocity is genuinely orthogonal — an apparent ξ≈0.65 nonlinear redundancy was an overlap artifact (199/200 rolling-window overlap); de-overlapped, it collapses. Two self-corrections recorded for honesty: (a) that ξ scare was mine and was wrong; (b) the follow-on "noise-dominated 80/20" framing was also overstated — under variance composition velocity is ~30%-of-variance structured (weak-but-real), and the IID return-shuffle null is mis-specified.
• The "informativeness gate" recommendation is out of rubric scope — signal/predictiveness is not a promotion axis; it is decided downstream (ML / cost-realism), as for Petrosian/Katz/dispersion. It is not a blocker for velocity. The cross-candidate LOO R² 0.969 worst cell is vs the un-promoted siblings → moot for the production panel.

velocity 3-axis scorecard → GO

Axis-2 caveat (honest): the 50/50 split is knob-like — a locked feature-local convention, not a first-principles or project-wide constant. Same Axis-2 category as bar_dispersion_entropy's c/m/d (PASS-with-caveat, promoted pending ratification), so it does not block GO — but ratify it, or switch the construction to Suriano 2025's canonical cross-window trajectory velocity (which removes the split → unconditional Axis-2).

bar_cecp_velocity is approved for implementation by the established 3-axis standard (Axis-2 PASS-with-caveat, as for dispersion).

Probe-hygiene findings (non-blocking, for future loops)

• The h_norm liveness gate is inert for continuous features (pd.qcut → ≈1.0 always) and has a fail-open bug (compute_h_norm returns None on <1000 non-NaN values → silently skips the gate on the ~19 smallest cells). Harmless this run (ρ-driven); make it fail-closed and stop citing h_norm as corroboration.
• The R² gate is in-sample linear and never fired (max 0.29 vs 0.85) and is blind to monotone-nonlinear redundancy — Spearman is the sole workhorse. Document as a coverage gap or replace with a rank/nonlinear measure on de-overlapped strides.
• Peak RSS hit ~93% of the 2 GB cap (1912–1915 MB) — add headroom before scaling candidate count.

Cross-references

• FRAMEWORK.md — Per-candidate analysis framework — read this BEFORE doing any per-candidate work
• candidates/ — per-candidate analysis outputs (populated after Terry approves the framework)
• RESEARCH-DIGEST.md — canonical candidate table (post-G9 reclassification)
• VERIFICATION-VERDICT.md — citation + FOSS verification (Layer-1 + Layer-2)
• LICENSE-AUDIT.md — FOSS dependency licensing audit
• BASELINE-RULED-OUT.md — Phase-1 baseline lift
• DUPLICATE-ELIMINATION.md — cross-session reconciliation
• CLAUDE.md — audit hub