Exp 103 (software-induced collapse) is contradicted 3-to-1 by F2, exp117, and F3 — blocking republication

[emooreatx]

Summary

Exp 103 grounds the flagship claim that software coordination alone can induce coherence collapse (ρ → 0.90 under barrier sync, → 1.0 under lockstep). Three later experiments measure the opposite on the same substrate. The corpus asserts both and reconciles neither.

This blocks the CCA v3 corrections (#11, defect C-13): a printed table is withdrawn whichever way it resolves, so the corrections version cannot ship until it is decided.

The contradiction

Experiment	Measurement	Direction
Exp 103	barrier sync ρ = 0.90; lockstep ρ = 1.0, k_eff = 1.0	correlates → collapse
F2	independent_rho 0.171 → synchronized_rho 0.0495; k_eff rises 9.48 → 15.63	decorrelates
exp117	baseline ρ = 0.394 → barrier ρ = 0.041	decorrelates
F3	lockstep_rho = 0.0036	decorrelates

All three later results are read directly from stored JSON (expF2_results.json, exp117_results.json, expF3_results.json in CIRISArray).

Note also that F2's own recorded pass criteria are rho_increased: False and response_increased: False (fragility p ≈ 0.38), yet F2 is written up as a clean finding ("sync decorrelates") rather than as a run whose stated criteria failed.

Why it matters beyond the paper

The claim propagates well past CCA v3:

• CLAUDE.md carried "Software-induced collapse (Exp 103) proves correlation structure, not nominal scale, determines resilience" until the 2026-07-31 correction pass
• It is the concrete demonstration behind "128 physical sensors collapse to k_eff = 1 through software coordination alone"
• It is the empirical anchor for the rigidity arm of the corridor — which F3 otherwise never measured (CCA v3 corrections version: Theorem 2.3 criterion (α/k not α/k_eff — verified in sympy), Table 8 defects, validation scoring — fix the Zenodo record, don't delete it #11, defect C-14: max measured ρ = 0.171 against a 0.43 edge)

If sync decorrelates, the framework loses its only demonstration of induced rigidity, and the two-sided corridor claim has no measured rigidity-side support on this substrate.

What needs deciding

1. Did Exp 103 measure a different quantity than F2/exp117/F3? (different estimator, different ρ definition, different array size or sample rate)
2. Is there a regime difference — does sync correlate at one coupling strength and decorrelate at another?
3. Or is Exp 103 simply wrong?

This requires hardware judgment; it cannot be settled from the stored artifacts. A re-run of Exp 103 under exp117's estimator, on the same array, would likely decide it.

Proposed

• Re-run Exp 103 with the estimator and array configuration used by exp117/F2
• If sync decorrelates: withdraw Exp 103, withdraw the software-induced-collapse claim, and mark the corridor's rigidity arm unmeasured on this substrate
• If Exp 103 is reproducible: identify what distinguishes it from the three contradicting runs and document the regime boundary
• Either way, correct F2's write-up to state that its own pre-registered criteria did not fire

Refs

#11 (defects C-13, C-14), CCA_PAPER/CORRECTIONS_v3.md.

[emooreatx]

RESOLVED by replication on the original hardware

Ran both experiments on this machine (RTX 4090 Laptop, idle at 64 °C, 0% util). Pre-registered before any measurement — PREREG_exp103_replication.md, sha256 7ffa2ee4386b2083727ea6ac03cd88b8bb414680017b65279b34f16fc4236516, frozen 2026-07-31T23:56:59Z, with predictions, interpretations, and a decision rule fixed in advance.

The answer is neither "Exp 103 is right" nor "F2/exp117 are right." Both are single unrepeated draws from an estimator whose run-to-run spread is larger than the effect they disagree about — and the lockstep leg is a code artifact.

Finding 1 — lockstep ρ=1.0 is an artifact, not a measurement (registered prediction P1, confirmed)

exp103_software_injection.py:127-128 computes one elapsed time for the whole batch and assigns that identical scalar to all 64 sensors:

all_timings[:, s] = t1 - t0          # one scalar -> 64 sensors

np.corrcoef is therefore being applied to 64 byte-identical rows. Replication confirms: identical_rows = True, ρ_abs = 1.0000 exactly, ρ_signed = 1.0000, and detrended ρ = 0.0000 at both n=30 and n=200. This is the same class of defect as the identity checks in #11 (C-5, C-11): a number that cannot come out any other way.

"128 physical sensors collapse to k_eff = 1 through software coordination alone" is withdrawn. The k_eff=1 is corrcoef of copied scalars.

Finding 2 — barrier correlation is real, and survives common-mode removal (P2 confirmed, P3 refuted)

I expected barrier ρ to be a shared-measurement-epoch artifact that would vanish under common-mode removal. It does not. At n=200: ρ_abs = 0.792 against a shuffle null p95 of 0.057, and 0.438 after subtracting the across-sensor mean at each sample — still far above null. Barrier sync does raise correlation. Exp 103's qualitative direction is supported.

Finding 3 — but the estimate is too unstable to support any point value (not pre-registered; post-hoc)

At F2's exact parameters (64 sensors, 50 samples, 5 trials), same code, same session:

Trial	independent	barrier
0	0.777	0.505
1	0.156	0.809
2	0.111	0.278
3	0.136	0.485
4	0.173	0.338

Barrier spans 0.278–0.809 (sd 0.18) across five identical runs; independent spans 0.111–0.777. The distributions overlap. In k_eff terms barrier ranges 1.23 to 3.46 — so whether you report "collapse" depends on which draw you got. Exp 103's 0.90 and F2's 0.0495 are both consistent with this one unchanged procedure.

Neither experiment repeated its measurement or reported an interval. That is the whole disagreement.

Finding 4 — a warm-up contaminant hits the first measurement (post-hoc)

The first measurement after idle reads ρ ≈ 0.777 regardless of mode — observed in both sessions independently (0.7777 and 0.7772). Subsequent runs settle to ~0.11–0.17. Both Exp 103 and F2 measure their independent baseline first, so both baselines are systematically inflated by GPU warm-up, and every Δρ computed against them is contaminated.

Also refuted: the estimator hypothesis (P5)

F2 uses signed mean(r); Exp 103 and exp117 use mean(|r|). This is not the explanation: correlations are overwhelmingly positive, so signed ≈ absolute in every barrier run (0.792 vs 0.792 at n=200). And expF2_causality_test.py:86-102 is byte-identical to exp103_software_injection.py:63-82 — same mechanism, opposite reported results.

Disposition

• Lockstep: withdrawn. Artifact, mechanism identified in source and confirmed empirically.
• Barrier: retained directionally, no point estimate. Any future ρ must be reported as a mean over repeated trials with an interval and a shuffle null, discarding the first run after idle.
• "Software alone can induce collapse": not supported. Its demonstration was the lockstep artifact; the barrier effect is real but cannot carry a collapse claim at this variance.
• This unblocks CCA v3 corrections version: Theorem 2.3 criterion (α/k not α/k_eff — verified in sympy), Table 8 defects, validation scoring — fix the Zenodo record, don't delete it #11. The paper's Exp 103 table is withdrawn per the pre-registered decision rule.

Methodological note for the program: no version of any of these experiments computed a null, repeated a measurement, or reported an interval. The shuffle null added here (200 draws, each sensor permuted independently) is what makes the readings interpretable at all — at n=30 the null p95 is 0.15, so a good part of what has been reported as correlation across this series may be estimator floor.

Artifacts: exp103_replication.py, exp103_replication_results.json, stability.py, stability_results.json, PREREG_exp103_replication.md.