Laboratory REG experiments at PEAR and elsewhere had documented small but statistically significant effects of directed human intention on random electronic devices. FieldREG studies extended this to group settings, sacred sites, ceremonies, and coherent gatherings, finding that moments of shared psychological engagement correlated with anomalous REG output. The GCP’s founding hypothesis was that world-scale events generating widespread emotional resonance would produce an analogous but globally distributed signal. Nelson framed this as testing for a “noosphere”, a layer of collective intelligence hypothesized by Teilhard de Chardin, by asking whether synchronized human attention leaves a physical trace in random data streams collected simultaneously around the world.⁴ The researcher’s preferred interpretation is that the effect reflects some aspect of consciousness interacting with physical random processes; alternative interpretations, including undetected systematic environmental correlates or statistical artifacts of the event-selection process, remain under active discussion.¹

Each GCP node uses a hardware RNG based on physical electronic noise, typically a reverse-biased semiconductor junction, producing a binary bit stream that is sampled at one trial per second and packaged into 200-bit trial blocks. The resulting data stream is transmitted via internet to the Princeton archive. The network reached a maximum of approximately 70 nodes in the original GCP 1 configuration.³ Node locations were geographically distributed across multiple continents, enabling the secondary analysis of inter-node correlations as a function of geographic distance, a feature that proved analytically important when Bancel and Nelson found that the significant cumulative result was driven by cross-network correlations rather than by individual node deviations.¹ The volunteer-hosted model introduced potential variability in local electromagnetic environments across sites, which Nelson identified as a factor requiring examination when assessing conventional physical explanations for observed deviations.⁴

For each registered event, the GCP extracts the relevant time window from the archive and computes a network-level z-score using a pre-specified algorithm. The primary algorithm computes the variance of the sum of simultaneous trial values across all active nodes, testing whether inter-node correlations exceed chance expectation. Individual event z-scores are then combined using the Stouffer method, summing z-scores and dividing by the square root of the number of events, to produce a cumulative statistic. After 236 events meeting strict criteria for network stability and correct hypothesis definition (out of 250+ replications in the first nine years), the cumulative Stouffer Z exceeded 4.5 standard deviations, corresponding to p ≈ 3 × 10⁻⁶.¹ The event-registration protocol partially addresses optional stopping, once registered, events are counted, but does not fully address the question of which events qualify for registration, a criterion that has been identified as a potential source of selection bias. The competing non-psi explanation of selection bias in event registration was examined by Bancel and Nelson, who noted that the criteria for event inclusion (network stability and correct hypothesis definition) were themselves pre-specified, and that the result remained significant when restricted to the most clearly defined events.¹ This partially addresses but does not fully eliminate the selection-bias concern.

A key secondary finding reported by Bancel and Nelson is that the significant cumulative result is structurally driven by correlations between RNG nodes across global distances, rather than by individual nodes deviating from their own expected distributions. This is analytically important because it distinguishes the GCP effect from a simple “each node is slightly biased” artifact: a systematic hardware bias in individual nodes would not produce inter-node correlations. The finding that correlations increase during registered event windows, and that this correlation structure is the primary carrier of the statistical signal, is the researcher’s preferred evidence that the effect is genuinely network-level and not reducible to local device artifacts.¹ The proposed mechanism, that coherent human attention somehow induces correlated deviations across physically separated random devices, remains contested and has no accepted physical account.

The GCP 1 main experiment comprised 500 formally registered events analyzed over a 17-year period (1998–2015). The Stouffer Z accumulated across all 500 events exceeded 7 sigma.³ An earlier published analysis covering the first 236 qualifying events (out of approximately 250 registered in the first nine years) reported a cumulative Stouffer Z > 4.5, p ≈ 3 × 10⁻⁶.¹ Nelson’s 2014 report noted that the result from more than 450 formal hypothesis tests departed substantially from expectation with odds exceeding a trillion to one against the null.² The network reached a maximum of 70 RNG nodes. No individual event effect size is reported in the pool; the cumulative result is the primary statistical claim. The experiment was not preregistered as a single study in the modern sense, the event-by-event registration protocol is the GCP’s own pre-specification mechanism, developed before contemporary preregistration infrastructure existed.

GCP 2.0 builds on the GCP 1 finding that 500 formal events over 17 years produced a cumulative Stouffer Z exceeding 7 sigma.³ The HeartMath-based network uses next-generation RNG hardware designed to track fundamental electronic behavior, with the goal of distinguishing consciousness-correlated deviations from conventional electromagnetic or hardware-drift explanations. The planned scale of 4,000 independent RNG channels would provide substantially greater statistical power per event than the original 65–70 node network, potentially enabling single-event analyses with meaningful power rather than requiring accumulation across hundreds of events. The GCP 2.0 position paper frames the project as supporting citizen scientists around the world who house a device, combining research with education on interconnectivity.³ A key open question that a well-powered GCP 2.0 analysis could address is whether the inter-node correlation structure identified in GCP 1 replicates with a larger, independently designed network, a preregistered, adversarially designed replication with pre-specified event criteria and independent analysis would be the most informative next step for settling the basic empirical claim.