›Navigation

ERRATUM (2026-06-24, operator-directed) — finding below is unchanged; this corrects ClickHouse wording only.
This loop does have read-only ClickHouse access (--readonly=2). This page's original wording implied access was lacking; that was imprecise and has been corrected in place. It only ever meant no query was needed this firing — Q2's status was synthesized from committed artifacts (full_grid_aggregate.json) that an earlier read-only ClickHouse run produced. Access is present; it was simply not required to answer Q2. From Iteration 3 onward, numbers are grounded in live read-only queries.
Denominator note: this page writes the crypto STABLE-empty result as 0/91 (the deduplicated full_grid_aggregate.json panel) and cites the SPIKE-2026-06-16 replication as 0/92 (its 92-slot panel). Same finding, two panel definitions — both return zero invariant features.

Iteration 2 — Q2: Crypto — split FRAGILE vs CONDITIONAL per feature PARTIAL — k-of-N MISSING

Read-only evaluation · no ClickHouse query needed this firing — synthesized from committed artifacts (the loop retains read-only --readonly=2 access) · load1 3.97 ≤ 24 · matrix-evaluation-for-orthogonality loop · PR #543

Question (backlog Q2)

Per feature, can the crypto features be split into FRAGILE (invariant in ≈0 environments) vs CONDITIONAL (invariant in k of N, 0<k<N)? The grounding metric is the per-feature k-of-N invariance count over the 10 crypto environments, under Type-I-controlled ICP (α=0.05).

Method (read-only, zero new compute)

Synthesize from committed artifacts; no new ClickHouse measurement, so leakage_guard.py is not invoked (R7 applies to new measurements). Artifacts read:

full_grid_aggregate.json — the FULL-GRID-SCALE run (383 cells, leakage gate PASS), icp block.
STATUS-GROUNDING-AND-THRESHOLDS.md — the authoritative status/threshold ledger (what is and isn't measurable).
icp_v3_cross_regime.json + SPIKE-2026-06-16 reconfirmation (ICP 0/92 replicated).

What the committed ICP artifact actually contains

Field (`full_grid_aggregate.json → icp → crypto\|BTCUSDT\|250`)	Value	Meaning
`n_features`	91	crypto feature panel size
`n_invariant_ge1`	0	features in the ICP-accepted invariant set = 0
`median_set_size`	0.0	median ICP invariant-set size across cells = 0 → empty invariant set

Why this does NOT split FRAGILE from CONDITIONAL. Standard ICP (Peters–Bühlmann–Meinshausen 2016) pools all environments and returns one invariant predictor set — here, empty. That certifies STABLE = ∅ (no feature invariant across all environments), but it does not record, per feature, in how many of the 10 environments it would pass invariance. STATUS-GROUNDING-AND-THRESHOLDS.md states this verbatim: "FRAGILE vs CONDITIONAL cannot yet be split per feature — the ICP run reported only the binary 'full invariance 0/91', not the per-environment pass count (k of 10) the boundary needs. That measurement is missing."

Findings — every number with its trust basis

Bucket	Count (of 91 crypto features)	Trust basis	Status
STABLE (k = N = 10)	0	ICP invariant-set size = 0; Type-I-controlled (α=0.05); leakage gate PASS; replicated SPIKE-2026-06-16 (0/92).	GROUNDED — empty
CONDITIONAL — certified	exactly 4 certified (the stamped four: `aggression_ratio`, `bar_dispersion_entropy`, `bar_katz_fd`, `bar_petrosian_fd`) — more may be CONDITIONAL once k-of-N is measured, but only these 4 are certified today	FDR route, not ICP: BH=BY q=0.10, PBO 0.00, deflation z 22.15 vs 1.53. Regime-scoped certificate (E3).	GROUNDED — 4 certified
FRAGILE vs CONDITIONAL (remaining 87)	unresolved	Requires per-environment ICP k-of-N (k of 10) per feature. Not computed; measurement confirmed missing.	UNRESOLVED

Verdict

PARTIAL. The split is bounded but not fully apportioned from committed state: STABLE = 0/91 (empty, grounded), exactly 4 features are certified CONDITIONAL by the FDR certificate (more may be CONDITIONAL but are uncertified), and the FRAGILE↔CONDITIONAL boundary for the other 87 is not resolvable — the standard ICP run emitted only the empty invariant set, never the per-environment k-of-10 pass counts the boundary requires.

Out-of-surface fix (R3) → handed to operator. Producing the per-environment k-of-N requires a new ICP harness (run invariance environment-by-environment, record per-feature pass counts). Authoring that analysis code would write outside the campaign dashboard folder (forbidden under R3 for this read-only evaluator). So this loop reports the gap precisely and hands the measurement to the operator — it does not build it. This mirrors Q1's disposition.

Anti-inference guard (R7)

It is tempting to read n_invariant_ge1 = 0 as "all 91 are FRAGILE (k=0)." Reject that. n_invariant_ge1 is the ICP invariant-set membership count, not a per-environment pass tally — an empty pooled invariant set is fully consistent with some features passing invariance in several individual environments (0<k<N) yet failing the cross-all-environments intersection. Indeed the FDR route already certifies 4 features as CONDITIONAL, which would be impossible if every feature were truly k=0. The k-of-N counts are genuinely absent, not implicitly zero.

What's next

Q3 is the natural successor and is independent of this gap: whether the 0/91 STABLE-empty is a strict-linear artifact — nonlinear/relaxed ICP (arXiv:1706.08576) + ε-tolerance frontier (arXiv:2501.17354), calibrated by the G10 shuffled-null power test (no-power vs no-signal).
Operator action to unblock a full Q2: extend the ICP harness to emit per-environment k-of-10 pass counts per feature (the single missing measurement named in STATUS-GROUNDING §"Missing measurements").
Nothing here is promoted to a gate (R7). The "0/91" is a Type-I-controlled report, not a threshold.

Read-only & resource-cap proof

Guard	Evidence this firing
ClickHouse READ-ONLY (R1)	No ClickHouse query was needed this firing — synthesized from committed artifacts. The loop retains read-only (`--readonly=2`) access; it simply was not required for Q2.
Watchdog precheck (R2)	load1 = 3.97 ≤ 24; no compute envelope launched.
Resource cap (R2)	Reads of JSON + markdown; peak RSS negligible.
Leakage discipline (R7)	No new measurement → `leakage_guard.py` N/A. Cited numbers inherit their committed leakage-PASS bases (FULL-GRID, E3 FDR, SPIKE-2026-06-16).
Write surfaces (R3)	Only this dashboard folder + PR #543 body. New ICP harness NOT authored (out-of-surface); handed to operator.

SSoT: full_grid_aggregate.json (icp block) · STATUS-GROUNDING-AND-THRESHOLDS.md §FRAGILE/CONDITIONAL · cross-refs SPIKE-2026-06-16-RECONFIRMATION.md, FULL-GRID-SCALE-VERDICT.md, E4-FINAL-VERDICT.md · audit twin 2026-05-26-forward-orthogonality-prediction/ · read-only evaluator · do NOT merge PR #543