Can Consciousness Nudge Randomness?

Holmberg, U. (2026). Can consciousness nudge randomness? Journal of Scientific Exploration, 40(1), 75–100. https://doi.org/10.31275/20263735

AI Assessment

A single-investigator, two-year naturalistic RNG field study reporting a very small but highly significant deviation from randomness during a pre-identified “stressful” morning window. The paper reports every figure this audit could check against its own text, and its design carries genuine strengths: a collection window fixed in advance, all pre-defined time windows reported whether significant or not, and a conservative Bonferroni correction. Those sit alongside real interpretive limits the paper is largely candid about, including the absence of formal preregistration, an “attention” value that is derived from the model rather than measured, a distance effect shown only by simulation, and no external replication. This audit describes what the paper reports and how it was run; it takes no position on whether consciousness influences randomness.

Provenance

DOI. 10.31275/20263735 · Gold open access (Creative Commons CC-BY-NC 4.0).

Study type. Single-author, two-year continuous random number generator (RNG) field study in a domestic setting, paired with a theoretical model (the Cognitive Entropy Shift Model, CESM) and a retrospective reanalysis of three earlier studies.

Author. Ulf Holmberg, independent researcher, Stockholm, Sweden (ORCID 0000-0001-5872-0076). Submitted 16 May 2025, accepted 4 June 2025, published 26 March 2026.

Funding. No funding statement appears in the article.

Data availability. The paper reports full descriptive statistics in Table 1 and additional methodological detail in appendices; it names no public data repository for the raw second-by-second dataset.

Source basis. Every figure below was confirmed against the primary article’s own stored full text (the born-digital JSE PDF, 26 pages; DOI cross-checked against Crossref). Two caveats apply to that text layer: the displayed equations render as extraction glyph codes (for example “/g86”) and Figures 1 to 5 do not survive in it, so any equation or figure claim must be checked against the published PDF; and minor extraction artifacts appear in the article’s text (the word “notable” is repeatedly split as “no table,” and endnote 16 reads “Thes values… There precise”). None of these affect the statistics audited here, which are confirmed from the numeric text.

What the paper reports

Over a two-year window (March 2022 to March 2024), a TrueRNG v3 hardware generator logged one value per second in a Stockholm apartment shared by one adult and two young children. The author reports that during a morning interval identified in advance as emotionally intense (school departure), the normalized RNG output deviated from chance by a very small amount, roughly 0.5% to 0.7% of the mean. The strongest window, 07:55 to 08:10, gave a Welch’s t of −4.347 (p < 0.001), while a matched control period immediately after, 08:10 to 08:25, showed no significant deviation.¹ The paper frames this within CESM, treats a “consciousness as informational constraint” hypothesis as the explanation, and applies the model retrospectively to three earlier studies.

The reported effect is very small (about 0.005 standard-deviation units); by the paper’s own power analysis roughly 710,000 observations suffice to detect it, and the participants were “largely unaware of the RNG’s presence,” so no cognitive or emotional state was measured at the time.

How it was run

• Design. A naturalistic, uncontrolled field study: one hardware RNG ran continuously for two years while the analyst compared pre-defined time windows against the full dataset.
• Apparatus. A TrueRNG v3 (avalanche-effect semiconductor source) connected to a Raspberry Pi 400, logging one value per second; internal firmware applied XOR whitening (combining roughly 20 raw bits per output bit).
• Placement. The device sat about 10 meters from the area identified a priori as the likely locus of elevated emotional intensity; temperature and power were monitored periodically.
• Participants. Two young children and one adult in the household, whose emotional states were hypothesized to influence the output; they were largely unaware of the device.
• Windows. A morning window (07:30 to 08:15) was designated in advance and split into 15-minute segments, plus a targeted 07:50 to 08:10 interval and a control period (08:10 to 08:25). The full two-year series was also divided into 32 non-overlapping 45-minute intervals across the day.
• Analysis. Output was normalized against each subset’s rolling monthly mean and variance; a two-sided Welch’s t-test compared each window to the full dataset, with a conservative Bonferroni correction for the overlapping windows.
• Sample. 47,731,465 values were generated; 8,789,615 were excluded (participants known to be away), leaving 38,941,850 valid observations. (The paper labels this sample about 15.2 months in Table 1 but 15.0 months in Appendix D; see Transparency.)

Results, as reported

Values are reproduced verbatim from the article’s Table 1 and Results section. The article’s prose reports the control-window p as 0.585 in one sentence and 0.592 in Table 1, a minor internal inconsistency; the table value is used here. The “A” column is not a measured quantity but an attention level back-computed from each window’s deviation using the CESM equation.

Eleven-dimension audit

Pre-registration

The study was not formally preregistered in a public repository. The author states the collection period and the morning target window were “identified in advance,” and explicitly lists “the absence of preregistration” among the criticisms the design was built to address, citing the standard methodological critique of the field.⁵ A priori designation of the primary window is real, but without a timestamped public registration it cannot be independently confirmed that the analysis windows and normalization choices were fixed before the data were seen. One detail sharpens the point: the paper’s stated a priori target is the “07:50 to 08:10” window, yet the strongest reported effect (t = −4.347) is credited to a 5-minute-shifted 07:55 to 08:10 window that has no row in Table 1’s pre-defined set, so the interval carrying the headline result is not the one the hypothesis pre-designated.

Randomization

There is no participant randomization or allocation; the “randomness” here is the RNG output itself. The source is a hardware avalanche-noise generator with firmware XOR whitening, and the null baseline is the full dataset’s own normalized distribution. The whitening step works against the hypothesis (it masks short-term structure), which the author argues makes any surviving deviation more notable rather than less.

Sensory leakage

This is not a percipient-guessing paradigm, so classical sensory leakage does not apply. The analogous confound is mundane physical coupling, temperature, power fluctuation, electromagnetic interference, or movement near the device, driving the deviations. The author addresses this by argument: environmental noise should peak during 07:30 to 07:45 when household movement was highest, yet the effect peaked later, at 07:55 to 08:10. The control is reasoned rather than instrumented per interval, and the paper concedes that future work “should incorporate more detailed environmental logging.”

Blinding

Participants were largely unaware of the device, which removes expectancy on their side. The analyst, however, was not blind: the same investigator designated the hypothesized windows, performed the normalization, and ran the tests, in his own home. No independent or blinded analysis is reported.

Optional stopping

The author states the collection period (March 2022 to March 2024) was fixed in advance and that no decision to stop or extend was made on the basis of interim results, even though notable effects appeared within the first year. As reported, this removes the optional-stopping concern, though the claim rests on the author’s account rather than a registered protocol.

Outcome measure

The primary outcome is a Welch’s t-test on the normalized RNG mean for each window versus the full dataset. It is a well-defined, pre-stated statistic. A distinct measure, the CESM attention value A, is derived from the same deviation via the model, so it cannot serve as independent corroboration of the effect; treating A = 8.60 as evidence for the mechanism would be circular, and the paper presents it as an estimate rather than a measurement.

Effect size

The effect is very small, a 0.5% to 0.7% shift in the normalized mean, about 0.005 SD units. Significance is driven substantially by the enormous sample: the author’s own power analysis shows roughly 7.1 × 10⁵ observations suffice to detect such a shift at 95% power, well below the 3.9 × 10⁷ analyzed. With tens of millions of observations, a vanishingly small departure from randomness reaches p < 0.001, so the effect size, not the p-value, is the load-bearing quantity, and the author acknowledges the effect falls “below the thresholds needed for real-time detection.”

Multiple comparisons

Six overlapping morning windows plus a 32-interval 24-hour scan were tested. The author applies a conservative Bonferroni correction and reports the overlap explicitly, noting the comparisons are “not entirely independent.” The peak window (07:55 to 08:10) survives correction, and one window (07:45 to 08:00) that was significant at 5% does not survive it, which is reported transparently. Bonferroni is a defensible choice here; the residual concern is the number of researcher-defined windows in a single dataset.

Internal replication

There is one continuous dataset from one household and one device, so internal replication is limited. The 32-interval 24-hour scan functions as a within-dataset specificity check, only the morning window was significant, which is a meaningful internal control but not an independent repetition of the effect.

External replication

No external replication is presented. The retrospective application of CESM to Dunne, Nelson and Jahn (1988), Nelson (2024), and Leskowitz (2011) is model-fitting to previously published results, not replication: the model estimates attention and intention values from those studies’ reported deviations.² The paper itself calls for “future preregistered studies” to replicate and extend the predictions.

Transparency

Transparency cuts both ways. On the positive side, the paper is gold open access, reports all pre-defined windows in Table 1 (including the non-significant ones), shows the full 24-hour scan in a figure, includes a power analysis, and devotes a whole section to standard statistical criticisms. Against that, the raw second-by-second dataset is not linked to a public repository, and the manuscript carries several internal inconsistencies a citing reader must reconcile:

• Dates. The main text dates the experiment March 2022 to March 2024 (stating it ended 19 March 2024), while Appendix D dates the same two-year study “from June 2021 to June 2023.” The duration matches; the calendar does not.
• Sample duration. The filtered sample is labeled about 15.2 months in the Table 1 caption but “450.6 days or 15.0 months” in Appendix D; the unfiltered total is given as “552.4 days, or 18.4 months” there, which a reader should reconcile with the “two-year” framing used elsewhere.
• Window label. The main text defines the targeted interval as “07:50 to 08:10,” but the Results and Table 1 report it as 07:55 to 08:10 (this audit uses the Table 1 value).
• Control p. The control window p is 0.585 in the Results prose and 0.592 in Table 1.

None of these changes the headline result, but a single investigator collecting and analyzing data in his own home, without a public preregistration, is exactly the setting in which such discrepancies matter most.

The adversarial record

• Lineage. The study sits squarely in the RNG micro-psychokinesis and field-consciousness tradition of Schmidt, the Princeton Engineering Anomalies Research (PEAR) program (Jahn and Dunne), and the Global Consciousness Project (Nelson). It explicitly reanalyzes Dunne, Nelson and Jahn (1988), Nelson (2024), and Leskowitz (2011).³⁴
• Contested literature. This research area is heavily critiqued, and the paper engages that critique directly. Holmberg situates his result against skeptics (Bösch et al. 2006, Alcock 2003, Scargle 2002) who argue such RNG deviations can reflect statistical noise, inconsistent methods, or selective reporting, and against the analytic-flexibility critique of psi research (Wagenmakers et al., 2011) that bears directly on a single-dataset, researcher-defined-window design. The RNG-psychokinesis tradition it builds on has itself been the subject of a large meta-analysis.⁶
• Reading against the study. The honest weaknesses are structural, not arithmetic: a sole investigator collecting and analyzing data in his own home; participants who were unaware of the device, so the causal story (their emotional states nudging the RNG) is inferred rather than designed or measured; an “attention” value produced by the model it is meant to support; a distance-decay result demonstrated by scaling a parameter in the equation rather than by physically moving the device; an effect size so small it is significant only because n is enormous; and no independent replication. The paper compares its 0.5% to 0.7% deviation to the small effects in Bem et al. (2015),⁷ a meta-analysis that is itself contested.