Preregistered and blind-analysis replications of the correlation-matrix experiment have predominantly failed to reproduce the excess-correlation effect.

Grote's blind analysis and both consensus-protocol replications are unambiguous nulls, reported by the proponents themselves in the second case. The partial exceptions are honestly recorded and do not overturn the pattern that the strictest designs found nothing.

The methodologically strictest modern replications—including the proponents' own preregistered consensus protocols and an independent full blind analysis by Grote—have unambiguously failed to reproduce the matrix effect. While a few earlier or partial tests reported mixed or positive results, the overarching pattern is that when analytic flexibility is completely removed, the effect vanishes. The strongest opposing consideration is that a few in-house studies and an unreviewed online preprint still claim partial significance, preventing a completely uniform null record.

Multiple independent lines of evidence show that the strictest replications have been null. Walach et al.’s two consensus‑protocol replications, with preregistered methods and nonparametric randomization tests, both failed to find significant excesses and concluded that the matrix paradigm is not replicable (MPI‑E003). Grote’s 200‑participant blind‑analysis study, in which the entire analysis pipeline was frozen before unblinding, likewise found results squarely in the range of chance (MPI‑E004). These tests specifically target the original matrix effect with modern safeguards against bias. The main opposing consideration is that Walach 2020, a preregistration‑era replication, did report a significant surplus, and Flores’ thesis reports three of five significant studies; thus “predominantly failed” does allow non‑negligible exceptions and the record is not uniformly null.

The two strongest-designed tests in the file are null: Grote’s 200-participant blind analysis fell well inside the chance distribution, and Walach et al. 2021 reported that neither consensus-protocol replication was significant, concluding the paradigm is not replicable. Grote 2017 adds three preplanned nulls. “Predominantly” is therefore the right adverb. The strongest opposing consideration is that Walach et al. 2020 was itself a modern, simulation-evaluated positive, and Flores reported three of five thesis studies significant under permutation tests, so the record is not a clean sweep of failures.

The strictest tests in the corpus are unambiguous: Grote's 200-participant full blind analysis fell well within the chance distribution, and the proponents' own two OSF-deposited consensus-protocol replications were both null, with Grote 2017's three preplanned analyses adding to the pattern. The word "predominantly" is accurate and necessary. The strongest opposing consideration is that the record is not a clean null: Walach 2020 was itself a preregistration-era, adequately powered replication reporting significance, and Flores reports three of five studies significant under conservative permutation analyses. Those exceptions are real but weaker than they look — the 2020 effect was deflated as "fragile" by its own authors under improved analyses, and Flores's five studies share equipment, personnel, and analytic choices, so they are not independent confirmations.

Four strict outcomes point one way: Grote 2017's three preplanned analyses (combined p = 0.315), Grote 2021's full blind analysis falling well within the chance distribution at n = 200, and both consensus-protocol replications in Walach et al. 2021, whose authors concluded the paradigm "is not a replicable paradigm in PSI research." The strongest opposing consideration is that "predominantly" is not "uniformly": Walach et al. 2020 was a preregistration-era, adequately-powered replication that reported a surplus, and Flores reported three of five significant under permutation tests. Even granting those, the ordering is clear — the tighter the analytic lockdown, the weaker the result — and the second 2021 replication's low power (0.69) does not disturb the pattern set by the first (0.88) and by Grote's blind test.

Unchanged: Grote's blind analysis (MPI-E004) and both consensus-protocol replications (MPI-E003) are clean nulls under the strictest available designs, reported by proponents themselves in the latter case. No new record contests this.

MPI-E004's full blind analysis and both consensus-protocol replications in MPI-E003 are clean nulls under the strictest designs in the corpus, with MPI-E006's preplanned nulls extending the pattern. Nothing in the record contests the observation.

Grote's full blind analysis (MPI-E004) and both consensus-protocol replications (MPI-E003) are clean nulls under the strictest designs available, the latter reported by the proponents themselves. MPI-E006's preplanned nulls extend the pattern; no record contests it.

MPI-E004's blind analysis of 200 participants and both consensus-protocol replications in MPI-E003 are clean nulls under the strictest designs available, with MPI-E006's preplanned nulls extending the pattern. No record in the corpus contests the generalization.

Evidence firmly establishes this null trend under strict conditions. The most rigorous blind-analysis test (MPI-E004) and the proponents' own consensus-protocol replications (MPI-E003) both yielded explicitly null results. The strongest opposing consideration is that a large two-lab preregistered replication (MPI-E002) and some permutation-based analyses (MPI-E007) still reported partial successes, indicating the record is slightly messier than a pure uniform failure.

Grote’s 200-participant blind analysis and Walach et al.’s two consensus-protocol replications are all preregistered or preregistration-era studies that used fixed analysis plans and, in Grote’s case, locked software before unblinding; all three failed to find significant excess correlations (MPI-E003, MPI-E004). These are the methodologically strongest tests of the paradigm and directly support the statement that such replications have predominantly failed. The main counter-consideration is that not all preregistration-era work was null — Walach 2020 and parts of Flores’s thesis reported positives — but these either lacked full blind analysis or were less independently controlled, so they qualify “predominantly” without overturning it.

The two cleanest modern tests failed: Walach et al. 2021 froze a consensus protocol on OSF and obtained nulls in both replications, concluding the paradigm is not replicable; Grote 2021 wrote and locked the entire analysis before unblinding 200 participants and landed well inside the chance distribution. Grote 2017's three preplanned analyses were likewise null. "Predominantly" survives the messiness: Walach 2020 predates those tighter controls and was later called fragile by its authors, and Flores's three-of-five thesis results were not full blind-analysis, preregistered replications. The strongest opposing consideration is that the record is not a clean sweep of nulls once every modern-era positive is counted.

Among the tests that actually froze analysis in advance, the record is close to uniform: Grote's three preplanned analyses were null (p = 0.438, 0.703, 0.0949), his full blind-analysis study of 200 participants fell "well within the expected random distribution," and the proponents' own two consensus-protocol replications, deposited on OSF before collection, were both null and led their authors to call the paradigm non-replicable. The strongest opposing consideration is the scope of "predominantly": Walach 2020 was a modern, hardware-rebuilt, simulation-evaluated replication that reported a surplus, and Flores reported three of five positives — but 2020's significance broke down under the authors' own improved analyses, and the Flores studies were neither blind-analysed nor independently preregistered as described here. Low power in the second Walach 2021 replication (0.69) slightly softens the claim without changing its direction.

Grote's full blind analysis (MPI-E004) and both consensus-protocol replications (MPI-E003), the latter reported by proponents, are clean nulls under the strictest designs in the corpus; MPI-E006's preplanned nulls extend the pattern. No record contests it.

Grote's full blind analysis (MPI-E004) and both consensus-protocol replications reported by proponents (MPI-E003) are clean nulls under the strictest designs available; MPI-E006's preplanned nulls extend the pattern. No record in the corpus contests it.

The strongest-designed tests are unambiguous: Grote's 200-participant blind analysis fell "well within the expected random distribution," and the proponents' own two consensus-protocol replications were null. The partial exceptions (Walach 2020; three of Flores's five studies) predate or sidestep the strictest controls.