โ€บNavigation

Iteration 4 โ€” Q3: is STABLE empty, or is it a strict-test artifact? PARTIAL โ€” degenerate-as-exact; deciding test unbuilt

Read-only evaluation ยท no ClickHouse query needed this firing โ€” synthesised from committed artifacts (read-only --readonly=2 access present) ยท load1 2.60 โ‰ค 24 ยท matrix-evaluation-for-orthogonality loop ยท PR #543

Question (backlog Q3)

Is STABLE truly empty, or is the crypto 0/91โ€“0/92 regime-invariance result a strict-linear (exact) test artifact? The grounding metric is a nonparametric residual ICP (arXiv:1706.08576) read on an ฮต-tolerance / budget frontier (arXiv:2501.17354). The required trust basis is G10 shuffled-null power calibration โ€” i.e. separating "no power" (the test can't detect invariance even when present) from "no signal" (there genuinely is none).

Method (read-only, zero new compute)

Synthesised from committed artifacts; no new ClickHouse measurement, so leakage_guard.py is not invoked (R7 applies to new measurements). Sources: SOTA-RESEARCH-2026-06-23.md (ยง2.3, ยง3.1), METRIC-ADDITION-GATE.md (ยงB), STATUS-GROUNDING-AND-THRESHOLDS.md, FULL-GRID-SCALE-VERDICT.md, LICENSE-AUDIT.md.

Findings โ€” every claim with its trust basis

ClaimValue / statusTrust basis
Strict ICP exact-invariant set0/91โ€“0/92 (STABLE-empty)Type-I-controlled linear/exact ICP (ฮฑ=0.05); committed + replicated (SPIKE-2026-06-16). This is the result whose interpretation Q3 interrogates.
Is an empty EXACT set even informative?No โ€” provably degenerateTheorem (arXiv:2501.17354, Gu et al. 2025): exact invariance pursuit is NP-hard even in the linear, well-separated case โ†’ an empty exact-invariant set is the expected, computationally-degenerate outcome, not an informative "no-signal" structural verdict. So 0/91โ€“0/92 cannot be trusted as a no-signal ceiling.
Correctly-posed form of the questionฮต-tolerance frontierSame paper's two-knob relaxation: "STABLE is empty" โ†’ "STABLE is non-empty at tolerance ฮต* with budget k*". Determining ฮต* needs the relaxed test below.
Deciding test: nonlinear/relaxed ICP (F014)specified but UNBUILTarXiv:1706.08576 (Heinze-Deml/Peters/Meinshausen) โ€” a nonparametric residual test that detects invariance from nonlinear mechanisms the linear test mechanically rejects. Classified magic-number-free (MATRIX-PARAMETER-CLASSIFICATION ยงF014). FOSS = R cran/nonlinearICP (GPL-2.0+), process-isolation via Rscript (LICENSE-AUDIT). Rscript is not installed on this host โ†’ cannot run without an R install + a new harness.
G10 power calibration (no-power vs no-signal)MISSINGThe required trust basis is absent. METRIC-ADDITION-GATE ยงB is explicit: the relaxed-ICP run "must be paired with G10's shuffled-null power calibration to separate 'no power' from 'no signal'; otherwise an empty result is meaningless." No such calibration exists yet.

Verdict

PARTIAL. Q3 splits into two parts and resolves asymmetrically:

Actionable correction for the grounded state. The campaign's headline "STABLE-ORTHOGONAL empirically empty (ICP 0/91โ€“0/92)" is accurate as a strict-exact result but is over-read if taken as "no invariant structure exists." The defensible phrasing is: "no feature is exactly regime-invariant under the strict linear ICP โ€” a computationally-degenerate outcome by 2501.17354; whether a relaxed/ฮต-tolerant invariant set exists is undetermined pending the G10-paired nonlinear ICP."

Anti-inference guard (R7)

Do not read this iteration as "STABLE is actually non-empty / a stable-orthogonal set exists." Nothing here finds any invariant feature. The only positive claim is the interpretive one โ€” that an empty exact set is not informative as a no-signal verdict (a published theorem) โ€” and the only honest status of the genuine question (is there relaxed invariance?) is undecided, measurement unbuilt. No metric is promoted to a gate.

Read-only & resource-cap proof (this firing)

GuardEvidence
ClickHouse READ-ONLY (R1)Read-only access present (--readonly=2); no query was required โ€” committed-artifact synthesis (0 queries, 0 writes).
Watchdog precheck (R2)load1 = 2.60 โ‰ค 24; no compute envelope launched (the only real measurement, F014, is out-of-surface and has no runtime here โ€” Rscript absent).
Leakage discipline (R7)No new measurement โ†’ leakage_guard.py N/A. The strict 0/91โ€“0/92 inherits its committed leakage-PASS basis; the NP-hardness claim is a literature theorem, not a data result.
Write surfaces (R3)Only this dashboard folder + PR #543 body. The F014 harness was NOT authored (out-of-surface); handed to operator.

What's next

SSoT: SOTA-RESEARCH-2026-06-23.md (ยง2.3 nonlinear ICP, ยง3.1 NP-hardness), METRIC-ADDITION-GATE.md ยงB, STATUS-GROUNDING-AND-THRESHOLDS.md, FULL-GRID-SCALE-VERDICT.md, LICENSE-AUDIT.md, MATRIX-PARAMETER-CLASSIFICATION.md ยงF014 ยท citations arXiv:1706.08576, arXiv:2501.17354 (re-verified 2026-06-23) ยท audit twin 2026-05-26-forward-orthogonality-prediction/ ยท read-only evaluator ยท do NOT merge PR #543