Elusive by Law

The ladder separates cleanly, and this case is decided on its top rung. The observation rung is solid: no party in the corpus disputes that the matrix program produced early surpluses, that its strictest replications are null, or that reported effects faded and shifted across four decades (MPI-C001 through MPI-C003). The mechanism rung is also real: Generalized Quantum Theory exists as a published formalism in a serious venue, and the no-transmission axiom's decline/displacement entailment is verbatim in print from 2007 — years before the preregistered failures it now explains (MPI-C004 through MPI-C006). What this buys MPI is integrity, not evidence: the predictions were demonstrably not retrofitted to the 2017–2021 record. But the attribution rung is where the featured thesis fails to carry its weight, for three structural reasons. First, the decline prediction is shared with the null hypothesis: regression to the mean acting on selected noise predicts exactly the same trajectory, so the replication record — the program's centerpiece — is evidentially inert between the rivals. Second, the foundation is compromised from inside: the founding surpluses were counted against independence assumptions the data violate, and that critique comes from the program's own doctoral work, not from hostile skeptics (MPI-C009). Third, every mundane mechanism the conventional account requires is documented operating inside this very corpus — a coding error manufacturing a celebrated effect, a post-hoc contrast conjuring significance where three preplanned analyses were null, publication-bias sensitivity in the parent literature (MPI-C008). Steelmanning MPI: it nominated its own positive test; its proponents publish their failures under admirably candid titles; one preregistered, adequately powered replication (Walach 2020) did report a surplus; and its genuinely discriminating fine structure — strength–reproducibility reciprocity, displacement into related observables — has never been prespecified and tested, so the model's distinctive content is unexercised rather than refuted. Steelmanning the null: the best-designed experiment in the corpus found exactly nothing; the proponents' own frozen consensus protocol failed twice; and "not a replicable paradigm" is the proponents' own conclusion about their chosen benchmark. The verdict is weakly_supported rather than contradicted because MPI's insulation makes contradiction structurally elusive and residual positives (Flores's permutation-surviving results, the unreviewed RNG split) forbid a clean-null verdict — but the burden has shifted decisively onto the fine-structure tests this model has never run.

run 2026-08-23-check-kimi · 2026-08-23 · Weakly supported · an AI seat, blind to the judgment · not human reviewed · per-claim verdicts on each claim page

The core dispute in "Elusive by Law?" is whether the ubiquitous decline effect observed in correlation-matrix parapsychology experiments is the lawful signature of generalized quantum entanglement or the entirely mundane product of statistical artifact and regression to the mean. The Model of Pragmatic Information (MPI) offers an intriguing, formally coherent framework (Generalized Quantum Theory) predicting that psychophysical correlations exist but inherently resist being harnessed as reliable signals (the no-transmission axiom). Steelmanning the MPI position requires acknowledging that its proponents successfully published these theoretical predictions before the recent wave of strict replication failures. Furthermore, the correlation-matrix paradigm initially produced statistical surpluses that appeared to support the model's expectations of "jumping" and transient effects, with even some recent, adequately powered studies showing residual positive anomalies under less constrained conditions. However, the evidentiary ledger tilts heavily toward the conventional explanation. The founding positive results are severely undermined by the statistical dependence critique; when the intercorrelation of psychological variables is properly modeled, the apparent anomalies substantially diminish. More decisively, as methodological rigor has escalated to include full blind-analysis and preregistered consensus protocols, the purported effects have completely evaporated. Steelmanning the conventional account means recognizing that it does not merely offer generalized skepticism, but relies on specific, documented failure modes explicitly caught operating within this very literature: post-hoc analytic flexibility conjuring significance out of null data, publication bias broadly impacting random-event-generator studies, and software coding errors previously mistaken for anomalous phenomena. Ultimately, while MPI is not entirely unfalsifiable—its proponents concede the paradigm's lack of replicability, and its fine-structure predictions regarding displacement remain technically testable—its current deployment absorbs failure far too easily. Because regression to the mean and multiple testing on noisy data naturally produce the exact same trajectory of initial surplus followed by scattering and decline, the evidence currently lacks the diagnosticity to justify invoking a novel scientific paradigm. The conventional statistical explanation comfortably accounts for the entire corpus without requiring revisions to the known laws of physics, leaving the MPI framework weakly supported as an empirical description of reality.

run 2026-08-25-check-gemini · 2026-08-25 · Weakly supported · an AI seat, blind to the judgment · not human reviewed · per-claim verdicts on each claim page

On the observation rung, the record is clear. Early correlation-matrix and related micro-PK studies reported excess correlations between psychological and random-physical variables (MPI-C001), but when the paradigm was re-run under stricter methodological control — preregistered protocols and, at the apex, Hartmut Grote’s full blind analysis — the effect mostly vanished (MPI-C002). Across groups, designs, and time, nominal effects weakened, failed to reappear, or shifted to different variables or analyses (MPI-C003). This decline-and-displacement pattern is uncontroversial as a description; it is also documented more widely across parapsychology (MPI-E014). On the mechanism rung, MPI is internally coherent about what it claims. Generalized Quantum Theory (MPI-C004) and the no-transmission (NT) axiom (MPI-C005) together yield, in print and well before the latest experiments, the qualitative prediction that genuine psychophysical correlations cannot form a stable, exploitable signal and must decline, reverse, or “evade” when experiments tighten. The correlation-matrix design is explicitly constructed as the model’s own test-bed: by distributing potential effects over many correlations, it is meant to allow such unexploitable correlations to appear as a surplus of significant cells while still forbidding any clear signalling channel (MPI-C006). As statements about what MPI says and how the paradigm was intended, these claims are well-supported by the theorists’ own publications. The crux lies on the attribution and falsifiability rungs. The featured hypothesis (MPI-C007) is that the messy record — early positives, decline, and displacement — is best explained by real non-local psychophysical correlations of the MPI kind. The steelman rival (MPI-C008) holds that regression to the mean, analytic flexibility, publication bias, and — critically — statistical dependence among matrix cells are jointly sufficient to explain the same pattern from null data. Here the evidence is asymmetric. On the conventional side, every mechanism invoked is independently documented in or near this corpus (MPI-E015–E017). Most importantly, Flores’s Monte Carlo and permutation work shows that the psychological variables in matrix studies are strongly intercorrelated and that dependence-aware simulations shrink the founding excess-correlation results (MPI-C009, MPI-E007). Grote’s 2017 study demonstrates how a null set of preplanned analyses can coexist with a seemingly impressive post-hoc positive (MPI-E006). A celebrated micro-PK effect in a neighboring line collapses under a discovered coding error (MPI-E017). And the largest blind-analysis test in the matrix paradigm is cleanly null (MPI-E004). These strands jointly make it provisionally credible that ordinary statistical and methodological artifacts can explain the corpus without anomalous correlations, even though no one has yet constructed a single quantitative simulation that reproduces the exact pattern across all studies. On the MPI-advocating side, there are peer-reviewed positives that resist being dismissed as trivial. Walach et al. 2020’s two-lab, preregistration-era study reported a significant matrix surplus evaluated against 10,000 simulations (MPI-E002), and Flores’s thesis found three of five in-house replications significant under permutation tests designed to respect dependence (MPI-E007). The 2023 online game replication reports a striking true-RNG/pseudo-RNG split (MPI-E008), which qualitatively points the “right” way for quantum-flavored models, though it is unreviewed and unevenly powered. These positives stop the conventional account short of “established”: they keep open a residual anomaly, especially because some were produced by methodologically sophisticated insiders aware of the statistical pitfalls. However, the strongest-designed and most insulated tests — Grote’s 200-participant blind analysis and the proponents’ own consensus-protocol replications — are null (MPI-E003, MPI-E004), and the founding series that originally motivated MPI has been weakened by the dependence critique. Given that MPI deliberately predicts decline under replication (MPI-C005), these failures neither obviously confirm nor refute it purely on pattern; but when weighed against priors and the documented artifact mechanisms, they substantially undercut MPI-C007. On this dossier alone, genuine non-local correlations remain at best weakly-supported. The falsifiability concern (MPI-C010) is real but not absolute. The proponents’ own language after failed replications — that failure is “likely part of the systematic nature of such effects” — illustrates a practice of reading both success and failure as consistent with MPI, which in practice insulates it from disconfirmation at the level of qualitative patterns. At the same time, the model is not logically unfalsifiable: it forbids any stable, exploitable psi signal and makes distinctive fine-structure predictions (e.g., stronger initial effects should decline faster; displacement should land in related observables). Those quantitative predictions have simply not been bound in advance and tested under adversarial blind conditions. Thus MPI-C010 is best rated mixed: the theory as used has been insulated, but need not be. Integrating these lines, the overall case assessment is mixed. The decline-and-displacement pattern that motivated MPI is real; the model’s internal logic and its self-nominated matrix test are clearly described; and there remain some anomalous-looking results. Yet the weight of the strongest evidence favors the conventional null-plus-artifacts account, and MPI’s distinctive content has not been sharply tested. On the present evidence, MPI remains an intriguing but weakly corroborated explanatory framework for a pattern that is, so far, more parsimoniously explained by noise, bias, and dependence, with a small residue not yet decisively allocated either way. Further progress depends on the proposed adversarial, dependence-aware, blind tests that pin down the theory’s fine-structure predictions.

run 2026-08-25-check-gpt · 2026-08-25 · Mixed · an AI seat, blind to the judgment · not human reviewed · per-claim verdicts on each claim page

The featured hypothesis is not that decline occurs — that is common ground — but that decline is the lawful signature of real, unexploitable psychophysical entanglement of the kind MPI describes. Assessed as an attribution about nature rather than as a description of a published theory, that hypothesis is only weakly supported by the records in this file. The observational rungs are solid and do not settle the case. Lucadou's group did report experimental-over-control surpluses from 1986 to 2006, and the later literature describes a genuine instability: Walach's 2020 two-lab surplus was called "significant but fragile" by its own authors a year later; two consensus-protocol replications were null; Grote's 2015 signal gave way to preplanned nulls in 2017 and a fully blind null in 2021. Kennedy's field-wide review confirms that this trajectory is the usual parapsychological pattern, not a matrix-specific curiosity. None of that distinguishes a shy law from selected noise. Steelman for MPI. The no-transmission axiom and its decline, sign-reversal, and displacement corollaries were in print in 2007, before the strict failures. The correlation-matrix design was the model's self-nominated positive test, not a post-hoc alibi, and for a time it produced the predicted surplus — including a large, simulation-evaluated two-lab report in an APA journal. Flores still found three of five thesis replications significant under permutation tests meant to respect dependence, and her 2023 online split (true RNG significant, pseudo-RNG not) points the direction a quantum-flavored account wants. Generalized Quantum Theory is a real published formalism, not a metaphor. The model takes a genuine risk at the aggregate level: any stable, exploitable psi channel anywhere would refute it. Read that way, the empty box Grote opened on stage is what the theory said a well-built box would contain. Steelman for the conventional account. Counting significant cells against independence, when the psychological variables are highly intercorrelated, inflates the founding statistic; Flores's own Monte Carlo work, reported from inside the program, says the original results diminish once that dependence is respected. Grote 2017 is a live demonstration that a post-hoc contrast can extract p = 0.012 from data whose three preplanned analyses were null. Maier and Dechamps traced a celebrated neighboring micro-PK effect to a coding error. Bösch's 380-study REG meta-analysis judged publication bias the most encompassing explanation of the parent literature. The methodologically strongest test in the corpus — two hundred participants, analysis software locked before unblinding — fell well inside the chance distribution. Regression to the mean plus multiplicity produces decline and "displacement" from null data without remainder. On this reading MPI is a quantum vocabulary for the null hypothesis. The records do not make the conventional account airtight. No one has shown quantitatively that QRPs and dependence reproduce this specific published matrix record; Bierman's neighboring ganzfeld simulation covered about sixty percent, not all, of that effect; and a few structured positives survived analyses that were not naive. Those residues keep C007 from being contradicted. They do not carry it. The model's distinctive fine structure — faster decline after stronger initial effects, reappearance in related rather than random observables — has never been pinned down in advance and tested. Until it is, successes confirm entanglement and failures confirm elusiveness, and the attribution question is decided by priors about how often selected noise looks like a law. The evidence in this file, weighed without those priors doing extra work, leaves the shy-correlation thesis standing on a thinning ledge: early reports of contested statistical validity, a handful of later positives that did not survive the next tightening, and a prediction that the null hypothesis also makes.

run 2026-08-25-check-grok · 2026-08-25 · Weakly supported · an AI seat, blind to the judgment · not human reviewed · per-claim verdicts on each claim page

The case has an unusual structure: its observational spine (MPI-C002, MPI-C003) is well documented and essentially uncontested, yet almost non-diagnostic, because the featured hypothesis and its rival predict the same trajectory. Everything therefore rests on two hinges — whether the founding positives were ever real (MPI-C001 against MPI-C009), and whether MPI's distinctive content beyond the shared decline prediction has ever been put at risk (MPI-C005, MPI-C010). On the first hinge, the evidence records are thin in a way that matters: MPI-E001 rests on second-hand characterizations of proceedings and German-language reports whose raw data have never been released, while the one dependence-aware reanalysis in the corpus (MPI-E007, Flores) comes from inside the program and reports that the original significances shrink. On the second hinge, the 2007 paper does state decline, sign reversal, displacement into related observables, and strength–reproducibility reciprocity in print, before the failures — a genuine priority claim (MPI-E009). But not one of those corollaries has ever been operationalized as a prespecified metric, so the model's only tested prediction is the one it shares with the null hypothesis. Steelmanning MPI: its self-nominated benchmark did produce a peer-reviewed, two-lab, 10,000-simulation positive (MPI-E002); Flores obtained three of five positives under permutation analyses built to respect dependence; the 2023 true-RNG/pseudo-RNG split points the theoretically expected way; and proponents publish their own failures and preregistered protocols, which is not the behavior of a closed system. Steelmanning the rival: every mechanism it needs is caught operating in this literature — a coding error manufacturing a micro-PK effect, a post-hoc rescue of a null study, publication-bias sensitivity in the parent REG meta-analysis, and documented intercorrelation inflating cell counts — and the single strictest test, Grote's 200-participant blind analysis, returned exactly what the rival predicts. I do not reach "contradicted" because the discriminating experiments (dependence-aware reanalysis of the founding data; equal-power source comparison; binding fine-structure predictions) have never been run, and a fragile residue of positives survives non-blind conservative analysis. I do not reach "mixed" because the asymmetry is real: the positives cluster where controls are weakest, and the strongest designs are uniformly null. My prior against unmediated psychophysical correlation contributes here, and I flag it: without that prior, the corpus would read closer to unresolved.

run 2026-08-25-check-opus · 2026-08-25 · Weakly supported · an AI seat, blind to the judgment · not human reviewed · per-claim verdicts on each claim page