Feature realness and usefulness — census, gap map and instrument grounding

HALTED2026-07-07

A sweep of three codebases found the team could prove a feature was computed correctly and was not a duplicate of another feature, but had nothing checking whether it predicted anything or fired often enough to be worth a slot — so the audit then built and stress-tested twenty statistical instruments on real Bitcoin bar data, keeping nineteen and discarding one.

Why it carries this status

Lifecycle, not result. This says where the audit sits in its process — never whether what it found was good.

## ★★ CAMPAIGN-COMPLETE (2026-07-24, iter 34) All 20 candidate feature-realness instruments (#0–#19) carry a terminal verdict. **19 GROUNDED · 1 REJECTED (#12) · 0 PARKED · 0 UNVERIFIABLE.** The BOUNDED-STOP condition (HARD RULE 1) is met; the loop ends.

findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/LEDGER.md

2026-08-17. Derived from folder evidence; the adversarial challenge pass CORRECTED the first reading. The quote exists verbatim (grounding/LEDGER.md:113 and :115), but it certifies the wrong object. LEDGER.md:3 defines the loop's scope in its own words: 'The loop (`../LOOP-PROMPT.md`) reads this file's terminal-verdict count as its STOP condition (all 20 of #0–#19 decided → CAMPAIGN-COMPLETE)' — i.e. 20 measurement INSTRUMENTS, the tool-building phase. The audit itself exists to judge FEATURES: CL

Blocked on Operator — the instrument-building phase finished; the feature campaign the audit exists to run has not started. Its own plan defers it: "Only then the feature campaign: HYPOTHESIS-REGISTRY.md (F0) for the 30 promoted columns ... Separate, later PR."

The verdict

The audit's own conclusion, reproduced in full from the source below. Not a summary — this is the document, rendered. Links inside it that point at unpublished files are shown as plain text rather than as links that would 404 here.

Verdict (plain English)

Question: when we call a feature "real," what do we actually have that proves it — and does anything prove it's worth keeping?

Answer: we can prove a feature is computed correctly (bit-exact Rust/Python parity), causal (compile-time look-ahead guard), and structurally sound — orthogonal in the precise sense of measured non-redundancy against the currently shipped columns (not "unique"), plus parameterless and agnostic (three-axis gate, ξ instruments, regime-persistence campaign — genuinely rigorous statistics). We cannot prove it predicts anything (the only forward-return test is an unenforced stub with no multiple-testing correction and no effective-n deflation), we cannot point to a falsifiable economic hypothesis for any individual feature (academic pedigree exists; registered expected-sign mechanisms do not), and we have zero machinery for usefulness — nothing anywhere measures whether a feature fires often enough to matter.

The smoking gun: candidate #65, the persistence-landscape L²-norm regime detector on BTC. It passed orthogonality, was graded CONDITIONAL, and its ~5–15-events-per-year firing rate — the exact property that makes it not worth a slot — was written down as supporting evidence, because under a redundancy lens rare-event heavy tails look like quality. No gate in any of the three repos would ever have flagged it.

Root cause is architectural, not negligence: the pipeline sends the 40-feature matrix straight into GP rule search and assumes the strategy layer will discover which features were worthless. That loop is the slowest, most expensive teacher available.

Proposed fix (this audit's KICKOFF.md): a cheap-first F0–F4 gate ladder at the feature boundary — pre-registered hypotheses (F0), an events-per-year usefulness census (F1, one SQL sweep, would have retired #65 in minutes), an honest IC panel with FDR + effective-n + null calibration (F2), sign/chronology checks (F3), and lead-lag event studies for detectors (F4). Terminal status is two-axis: REAL×DENSE ships onward; REAL×SPARSE gets archived with respect; SPURIOUS goes to the boneyard; UNVERIFIABLE cannot ship by definition. Everything runs on existing data with laptop-scale compute; strategy-level machinery stays untouched as the final check.

Nothing empirical has been run yet. This folder is the census, the gap map, and the plan. The first firing is F0 — prose pre-registration — precisely so that every number produced afterward is interpretable.

source: findings/evolution/audits/2026-07-07-feature-realness-usefulness/verdict.md

Still owed 8

Phase P4 — the feature campaign itself — has never run: no per-feature pre-registration (F0) exists and none of the 30 promoted columns has been judged by a grounded instrument. Explicitly scoped to a separate, later PR.
The root CLAUDE.md still reads 'SCAFFOLDED … No metric admitted; no feature evaluated yet' and verdict.md still reads 'Nothing empirical has been run yet' — both superseded by grounding/LEDGER.md and never updated.
One operator sub-decision remains open from amendment 17a: the forex horizon rung (default 3s-shared).
Instrument #12 (knockoffs) carries a re-test condition: a construction such as an HMM or deep-generative time-series knockoff that controls realised FDR ≤ 0.10 on the long-memory panel. Recorded as possible future work, out of scope.
Crypto cost labels still exist for BTCUSDT only; the other 15 crypto symbols are unlabeled, so no cross-sectional crypto claim is supported.
The Harvey-Liu-Zhu trial-count denominator M is frozen as a counting rule but its integer is 'stamped at first feature-panel run' — that run has not happened.
Instruments #12, #15, #16, #17 and #18 have result JSON and Python artifacts under grounding/ but no companion .md evidence file, unlike the other fifteen.
Instrument envelopes published as blind spots and routed onward (e.g. #4/#5/#7/#19 all report FNR=1 on interaction-only signal, routed to #11/#13) are recorded but not yet exercised on a real feature.

What it claims, and what backs each claim 24

Every row pairs a claim with the file it came from and the verbatim text in that file. The sources sit above the deploy root, so the quote is embedded and the path is printed as text rather than linked — a link would resolve on a laptop and 404 here.

Claim	Evidence
The census verdict: the repositories are strong at proving a feature is correct, causal and non-redundant, and near-silent on whether it predicts anything or matters economically. CONFIRMED 3 repos swept (opendeviationbar-py @62a93d45, opendeviationbar-patterns @d9c6be695, mql5 @e218367)	The stack is *excellent at proving a feature is correct, causal, and non-redundant* — and almost silent on whether it predicts anything, means anything economically, or fires often enough to matter.** "Orthogonal" is measured feature-vs-feature; it is never measured feature-vs-future. findings/evolution/audits/2026-07-07-feature-realness-usefulness/DEEP-RESEARCH-FINDINGS.md
The only forward-return test in the codebase is an unenforced stub with a trivial threshold, no multiple-testing correction and no autocorrelation deflation. CONFIRMED 1 stub (tier2.py) + 1 one-off screen (iter-250); threshold \|ρ\|>0.02, 1-bar horizon, 7-day window	`python/opendeviationbar/validation/tier2.py:66-102` — Spearman(feature, next-bar return) at a trivially low `\|ρ\|>0.02` threshold, no multiple-testing correction, no effective-n deflation findings/evolution/audits/2026-07-07-feature-realness-usefulness/DEEP-RESEARCH-FINDINGS.md
Nothing anywhere in the three repos measures how often a feature fires — the usefulness axis had zero standing machinery. CONFIRMED 0 event-rate gates across 3 repos	Zero machinery. No probe anywhere computes activation rate as a gate. The single confrontation with sparsity is the 2026-04-19 orthogonal-exotica trigger campaign (patterns, `verdict.md:87` — triggers killed as "too dense" or "too sparse <1K untestable"), a one-off at the trigger level. findings/evolution/audits/2026-07-07-feature-realness-usefulness/DEEP-RESEARCH-FINDINGS.md
The motivating case: candidate #65's firing rate of roughly 5–15 events a year — the property that makes it not worth a slot — was written down as supporting evidence, because no gate measures usefulness. CONFIRMED ~5–15 events/year; skew +4.16; kurtosis +22.71; max\|ρ\|=0.361; h_norm=1.0	Its ~5–15 events/year firing rate was explicitly observed (skew +4.16, kurtosis +22.71) and recorded as a virtue — "distribution-level evidence supports the hypothesis" — because heavy-tailed rarity is what a good rare-event detector looks like under a redundancy lens. findings/evolution/audits/2026-07-07-feature-realness-usefulness/DEEP-RESEARCH-FINDINGS.md
Nine specific feature-level questions were catalogued as having no standing machinery (G-1 through G-9), each anchored to a file and line. CONFIRMED 9 gaps (G-1..G-9); 52 of 113 feature-store manifest columns carry academic 'Paper:' anchors; 0 grep hits for a per-feature edge-hypothesis field	\| G-5 \| Event-rate / signal-density floor ("fires 5-6×/yr") \| Zero machinery; doc-field proposal REMAIN \| `FRAMEWORK-AMENDMENTS.md:229-236` \| findings/evolution/audits/2026-07-07-feature-realness-usefulness/DEEP-RESEARCH-FINDINGS.md
The cohort to be judged was resolved to 30 promoted ClickHouse columns — 15 crypto and 15 forex — which do not overlap at all with the 40 grammar features the earlier kickoff had named. CONFIRMED 30 columns (15 crypto + 15 forex); intersection with FEATURE_NAMES_V2 (40) = ∅	Decision (2026-07-08): the grounding target is "all features promoted by nasimubd" — resolved below to a concrete, SSoT-anchored set of 30 columns (15 crypto + 15 forex). findings/evolution/audits/2026-07-07-feature-realness-usefulness/TARGET-COHORT.md
A multi-agent literature sweep assembled a 112-method state-of-the-art corpus across eight separable axes of feature quality. MEASURED 112 deduplicated methods; 13 agents; ~1.15M tokens; 153 web tool-uses; 8 axes (A–G plus C′)	Provenance: multi-agent web sweep, 2026-07-08 — 8 research angles + completeness-critic + 4 gap-fills, 13 agents, ~1.15M tokens, 153 web tool-uses. findings/evolution/audits/2026-07-07-feature-realness-usefulness/SOTA-GROUNDING.md
An adversarial pass concluded that none of the 20 candidate metrics was admissible on its first-draft self-test battery. ASSERTED 0 of 20 admissible on first draft (n=20 instruments)	The honest current answer, from the adversarial pass below: 0 of 20 candidate metrics is admissible on its first-draft battery — every one needs the shared signal-zoo hardening bolted on first. findings/evolution/audits/2026-07-07-feature-realness-usefulness/METRIC-EVALUATION-FRAMEWORK.md
The pre-registration froze the pass-marks before any data was looked at: BY-FDR at q=0.10, a Harvey-Liu-Zhu \|t\| ≥ 3.0 realness bar, and DSR ≥ 0.90 with PBO < 0.20 for utility. CONFIRMED q=0.10 BY-FDR; \|t\|≥3.0 HLZ; DSR≥0.90; PBO<0.20; confidence level 90%	\| FDR \| BH–Y at q = 0.10 across the full (feature × horizon × cell) panel \| findings/evolution/audits/2026-07-07-feature-realness-usefulness/PREREG.md
An amendment replaced the pre-registration's chosen bar-count horizon with the materialized 3-second cost-realism label as the thing a feature must predict. CONFIRMED horizon derived 3s; {1,5,20} bars demoted to reported-not-gated; cost hurdle = per-bar label_fwd_3s_roundtrip_cost_bps instead of a flat 500 bps	# PREREG-AMENDMENT-2026-07-17a — the return object is the materialized cost-realism label; horizon is DERIVED (3s), not a chosen bar-count findings/evolution/audits/2026-07-07-feature-realness-usefulness/PREREG-AMENDMENT-2026-07-17a-HORIZON-COST-RETURN-OBJECT.md
A companion amendment froze the three failure-mode classes the pre-registration had left as proposals: leakage, economic triviality, and regime instability. CONFIRMED 3 classes frozen (A leakage/look-ahead, B economic-triviality, C regime/rotational-instability)	Status: FREEZE (append-only companion). Freezes the three failure-mode classes that [`PREREG.md`](./PREREG.md) §6 listed as PROPOSED — NOT frozen here and explicitly deferred: "freezing them is the next pre-registration's job" (`PREREG.md:102-107`). findings/evolution/audits/2026-07-07-feature-realness-usefulness/PREREG-AMENDMENT-2026-07-17b-FAILURE-MODES-FROZEN.md
A read-only database census established the actual gateable surface: only Bitcoin carries crypto cost labels, at four thresholds and one horizon, giving 20 gateable cells in total once forex is included. MEASURED crypto labels ≈6.0M rows, BTCUSDT only, 4 thresholds (100/250/500/750 dbps), 3s only, 98.9–99.9% coverage; forex 4 symbols × {1,3,5,10}s; 16 crypto feature symbols but only 1 labeled; 20 gateable cells	→ 20 gateable cells (4 crypto BTC-thr + 16 forex symbol×rung). Enough to ground the instruments and run a first feature verdict, not enough to claim cross-symbol crypto generality findings/evolution/audits/2026-07-07-feature-realness-usefulness/DATA-UTILIZATION-PLAN.md
The four crypto cells are nested thresholds of a single asset sharing the same trades, so they are correlated rather than independent replications — the audit refuses to claim cross-sectional crypto generality. CONFIRMED 4 crypto cells, 1 underlying symbol; 4 independent forex instruments carry the cross-sectional weight	1. Crypto cost-anchored evidence is single-symbol (BTC). The 4 crypto "cells" are 4 nested thresholds of one asset, sharing the same underlying trades — they are correlated, not independent replications. findings/evolution/audits/2026-07-07-feature-realness-usefulness/DATA-UTILIZATION-PLAN.md
An operator ruling retired the planned synthetic signal-zoo entirely: every control must be a shuffle, permutation, bootstrap or subsample of real values. CONFIRMED 5 positive + 7 null synthetic archetypes retired; 4 real-data control constructors adopted	OPERATOR RULING (2026-07-17): 03c GOVERNS — real-data-only controls only. The synthetic signal-zoo is retired for this campaign. Every instrument grounds through the 03c C1′/C2′/C3′/M5′ pattern findings/evolution/audits/2026-07-07-feature-realness-usefulness/DATA-UTILIZATION-PLAN.md
First real measurement: the 3-second forward return is strongly autocorrelated, so the honest sample size is under a tenth of the nominal row count. MEASURED n = 870,161 bars (BTCUSDT@250); ρ₁ = 0.477; HAC effective-n ≈ 81,314 (9.3% of nominal); scalar AR(1) N_eff ≈ 308K, 3.8× more optimistic	- The 3s forward return is strongly autocorrelated: lag-1 ρ₁ = 0.477. Nominal N = 870,161 but the HAC (Newey-West, frozen lag=30) effective-n of the mean is ~81,314 — only 9.3% of nominal. findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/README.md
Using naive independent-sample statistics manufactures significance: t-statistics inflate 1.3×–2.3× against the autocorrelation-corrected version. MEASURED inflation 1.287×–2.339× over 4 features, n = 870,161 bars; vwap_close_deviation IC −0.315, t −294 naive vs −126 HAC	Naive-iid t manufactures significance: per-feature Spearman-IC t-stats inflate 1.3×–2.3× under naive-iid vs HAC. `vwap_close_deviation` IC = −0.315 reads t = −294 naively but t = −126 under HAC (2.3× inflation). findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/README.md
Before grading any instrument, the four real-data control constructors and the shared detection primitive were self-validated on data with known answers, and all four passed. MEASURED n = 870,304 real bars, BTCUSDT@250; duplicate ρ=1.000000; known-positive HAC-t 43.8; known-null HAC FWER 0.002 over 500 permutations; N_min = 100,000 at achievable ρ=0.0643; wall 17.6 s, peak RSS 364 MB	Date: 2026-07-22 · Operator: nasimubd · Verdict: `SEAL-PASS` — the lab is open. findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/seal_control_substrate.md
The two substrate instruments — a bit-exact causality check and a purged/embargoed cross-validation partition — both cleared, which was the campaign's own stop-or-proceed gate. MEASURED #1: 0 false positives over 221,187,473 bit comparisons, look-ahead power 1.000 (512/512); #0: naive k-fold leak t=169 over 252,512 straddling pairs → 0 straddling after purge, null FPR 0.000, 290,101 events/fold	★ SUBSTRATE GATE: PASS (2026-07-22). #1 (future-perturbation invariance) GROUNDED + #0 (CPCV purge/embargo) GROUNDED → both substrate instruments cleared → the campaign proceeds to P2 (realness/utility axis). findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/LEDGER.md
One instrument was rejected: standard model-X knockoffs could not hold their false-discovery guarantee on this data, leaving an irreducible ~1.3× overshoot and a power ceiling below the required bar. MEASURED realised FDP ≈0.13 (1.3× nominal q=0.10) at every IC; power ceiling 0.67 @IC=0.03 across 4 statistics (lasso 0.67, mlr 0.45, ridge 0.30, lcd 0.46), need ≥0.8; 4 slices across iterations 22–25	#12: REJECTED (2026-07-24, iter 27) — OPERATOR RULING (Nasim, 2026-07-24). Supersedes the iter-25 operator-ruling checkpoint. Reason: standard Gaussian model-X knockoffs do not hold on our data. findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/LEDGER.md
The bounded instrument-grounding loop reached its stop condition with all twenty instruments carrying a terminal verdict — nineteen accepted, one rejected, none parked. MEASURED 20 of 20 terminal over 34 loop iterations; 19 GROUNDED, 1 REJECTED, 0 PARKED, 0 UNVERIFIABLE	Graduation tally (FINAL): 11 DISCOVERED · 0 UNDER-EVALUATION · 19 GROUNDED (#1, #0, #2, #3, #6, #8, #9, #10, #4, #5, #7, #19, #11, #13, #14, #15, #16, #17, #18) · 1 REJECTED (#12) · 0 PARKED. ★★ 20/20 TERMINAL — CAMPAIGN COMPLETE. findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/LEDGER.md
A recurring, reusable lesson emerged across six instruments: textbook thresholds under-control error rates on this long-memory heavy-tailed data and must be replaced by permutation or high-percentile bootstrap calibration. MEASURED pattern recurs in 6 instruments (#10, #4, #5, #7, #16, #18); e.g. #10 analytic HAC FWER 0.058→0.130 (up to 2.6× α) vs permutation-calibrated 0.043	(1) PATTERN #2 conservative-calibration — analytic/nominal thresholds under-control on real long-memory heavy-tailed data → permutation / higher-percentile block-bootstrap CV (#10,#4,#5,#7,#16,#18). findings/evolution/audits/2026-07-07-feature-realness-usefulness/grounding/LEDGER.md
No feature has yet been judged — applying the accepted instruments to the 30 promoted columns was deliberately deferred to a separate, later piece of work. OPEN 0 of 30 promoted columns evaluated	### Phase P4 — The feature campaign (SEPARATE, LATER PR — not this one) findings/evolution/audits/2026-07-07-feature-realness-usefulness/EVALUATION-PLAN.md
The loop that produced the instrument verdicts ran from the laptop but dispatched every database read and heavy computation to bigblack under a hard read-only, 5-core, 5-GB, no-swap cap. CONFIRMED caps: readonly=2, CPUQuota 500%, MemoryMax 5G, MemorySwapMax 0; watchdog abort at load1 > 24	2. Hard cap 5 CPU cores / 5 GB RAM, swap banned — every compute process runs (on bigblack) inside `systemd-run --user --scope -p CPUQuota=500% -p MemoryMax=5G -p MemorySwapMax=0` with `nice -n 19 ionice -c3` and single-thread BLAS. Cap-kills are evidence; the cap is never raised. findings/evolution/audits/2026-07-07-feature-realness-usefulness/LOOP-PROMPT.md
The audit's own header status and root verdict are stale: both still describe a scaffolded campaign with nothing run, contradicting the completed instrument ledger. REFUTED verdict.md is round-1 (2026-07-07); grounding/LEDGER.md records 34 iterations through 2026-07-24	Nothing empirical has been run yet. This folder is the census, the gap map, and the plan. The first firing is F0 — prose pre-registration — precisely so that every number produced afterward is interpretable. findings/evolution/audits/2026-07-07-feature-realness-usefulness/verdict.md

The audit folder 38 markdown files

Source of record: findings/evolution/audits/2026-07-07-feature-realness-usefulness/ — not published, so these are listed rather than linked.

File	Role
`CLAUDE.md`	Hub: status header, four-round spoke map, result summary, conventions and next steps.
`DATA-UTILIZATION-PLAN.md`	Read-only database census, the 20-gateable-cell extent, four hard robustness caveats, and the no-synthetic-data operator ruling.
`DEEP-RESEARCH-FINDINGS.md`	The census: what machinery exists across three repos, the candidate-65 case study, the 9-gap map.
`EVALUATION-PLAN.md`	The ordered build P0→P4, the 9-gap-to-instrument closure map, and the 8 open pre-registration decisions.
`KICKOFF.md`	Historical F0–F4 gate-ladder proposal, explicitly NOT ADMITTED — kept as the seed-idea record.
`LOOP-PROMPT.md`	Canonical bounded /loop prompt: laptop drives, bigblack computes, 20 instruments then STOP.
`METRIC-EVALUATION-FRAMEWORK.md`	How a metric earns the right to judge a feature: DISCOVERED→GROUNDED pipeline, 5-control battery, 20 per-instrument acceptance criteria.
`PREREG-AMENDMENT-2026-07-17a-HORIZON-COST-RETURN-OBJECT.md`	Superseding amendment: the return object becomes the materialized 3s cost-realism label; horizon derived, not chosen.
`PREREG-AMENDMENT-2026-07-17b-FAILURE-MODES-FROZEN.md`	Freezes the three deferred failure-mode classes (leakage, economic triviality, regime instability) and their guards.
`PREREG.md`	The frozen Phase-P0 pre-registration: five buckets plus the Step-0 profitability decision.
`PROVENANCE.md`	How the census was produced: three read-only subagent sweeps at pinned SHAs, trust model, spot-check protocol.
`SOTA-GROUNDING.md`	The 112-method state-of-the-art corpus organised by the eight-axis model, with primary sources.
`TARGET-COHORT.md`	Resolves the grounding target to 30 promoted columns (15 crypto + 15 forex) and corrects the kickoff's 40-feature target.
`evidence/metric-selftest-digest.md`	Per-instrument self-test protocols and adversary reports for the 20 candidate metrics.
`evidence/nasimubd-promoted-cohort.md`	Verbatim cohort sweep proving the 30-column promoted set and its authorship.
`evidence/sota-research-digest.md`	Deduplicated 112-method state-of-the-art digest organised by grounding axis.
`evidence/sweep-mql5.md`	Verbatim round-1 census sweep report for mql5/fxview-core @e218367.
`evidence/sweep-opendeviationbar-patterns.md`	Verbatim round-1 census sweep report for opendeviationbar-patterns @d9c6be695.
`evidence/sweep-opendeviationbar-py.md`	Verbatim round-1 census sweep report for opendeviationbar-py @62a93d45.
`grounding/LEDGER.md`	Append-only instrument-verdict ledger with the substrate gate, all 20 terminal verdicts and the CAMPAIGN-COMPLETE declaration.
`grounding/README.md`	Harness hub: what has run, the P1 effective-n result, and the metered next steps.
`grounding/instrument_00_cpcv_purge_embargo.md`	Evidence artifact for instrument #0, purged/embargoed combinatorial cross-validation (GROUNDED).
`grounding/instrument_01_future_perturbation_invariance.md`	Evidence artifact for instrument #1, bit-exact future-perturbation causality floor (GROUNDED).
`grounding/instrument_02_block_permutation_null.md`	Evidence artifact for instrument #2, block-permutation shuffled-label null (GROUNDED).
`grounding/instrument_03_effective_n_fdr.md`	Evidence artifact for instrument #3, effective-n deflation plus Benjamini-Yekutieli FDR (GROUNDED).
`grounding/instrument_04_rank_ic.md`	Evidence artifact for instrument #4, rank information coefficient with HAC t (GROUNDED after two checkpoints).
`grounding/instrument_05_ic_decay.md`	Evidence artifact for instrument #5, information-coefficient decay and half-life (GROUNDED).
`grounding/instrument_06_sfi_single_feature_oos.md`	Evidence artifact for instrument #6, single-feature out-of-sample importance / the return object (GROUNDED).
`grounding/instrument_07_mda_onc.md`	Evidence artifact for instrument #7, mean-decrease-accuracy with ONC clustering (GROUNDED).
`grounding/instrument_08_deflated_sharpe.md`	Evidence artifact for instrument #8, permutation-based deflated Sharpe ratio (GROUNDED after a checkpoint).
`grounding/instrument_09_pbo_cscv.md`	Evidence artifact for instrument #9, probability of backtest overfitting via CSCV (GROUNDED after a checkpoint).
`grounding/instrument_10_harvey_liu_zhu.md`	Evidence artifact for instrument #10, Harvey-Liu-Zhu multiple-testing haircut with a permutation-calibrated threshold (GROUNDED).
`grounding/instrument_11_cmi.md`	Evidence artifact for instrument #11, k-NN conditional mutual information with a local-permutation null (GROUNDED).
`grounding/instrument_13_dml_cpi.md`	Evidence artifact for instrument #13, double machine learning / conditional predictive impact (GROUNDED).
`grounding/instrument_14_spanning.md`	Evidence artifact for instrument #14, Huberman-Kandel spanning intercept (GROUNDED on a published power envelope).
`grounding/instrument_19_quantile_monotonicity.md`	Evidence artifact for instrument #19, quantile monotonicity with Page-trend and Ramsey-Wilcoxon tests (GROUNDED after a checkpoint).
`grounding/seal_control_substrate.md`	SEAL evidence artifact: self-validation of the four real-data controls and the IC+HAC detector primitive.
`verdict.md`	Round-1 plain-English bottom line; now stale relative to the grounding ledger.

Dashboard twin

findings/dashboard/realness/

Generated by findings/dashboard/build_audits.py from findings/evolution/audits/2026-07-07-feature-realness-usefulness/AUDIT_LEDGER.json — never hand-edited. Each quote was verified to occur in the file named beside it when the ledger was written.