Machine learning analysis of anomalous information acquisition data

# OSF Preregistration: Replication of Target-Specific Effects in Online Remote Viewing ## Study Information ### Title Replication of "Bright" vs. "Dark" Target Effects in an Online Remote Viewing Experiment Using Siamese Neural Network Analysis ### Authors Dean Radin, Institute of Noetic Sciences ### Description This preregistration predicts that remote viewing targets identified as "bright" (high hit rate) vs. "dark" (low hit rate) in a 19-year exploratory database (2000–2018) will show significantly different performance in a 5-year confirmatory replication (2021–2025). Target brightness classifications were determined from the exploratory dataset and frozen prior to analyzing the confirmatory dataset. ### Hypotheses **Primary Hypothesis (H1):** Targets classified as "bright" in the exploratory dataset (2000–2018) will show significantly higher hit rates than targets classified as "dark" in the confirmatory dataset (2021–2025). This will be tested by correlating the semantic percentage score for each target used in the first epoch against the same targets in the second epoch. ## Design Plan ### Study Type Confirmatory replication of exploratory findings ### Blinding Analysts were blind to the confirmatory dataset's percentile scores for each target during classification of bright/dark targets. Target classifications were locked based on exploratory data before any confirmatory data analysis. ### Study Design This is a preregistered replication using archival data from an ongoing online remote viewing experiment (gotpsi.org). **Exploratory phase (2000–2018):** - 19 years of data used to identify target-specific patterns - Targets classified as "bright" (high semantic percentile scores) or "dark" (low semantic percentile scores) - Semantically biased words identified via permutation testing were not used in this analysis. - All classifications and analytical code will be frozen before confirmatory analysis **Confirmatory phase (2021–2025):** - 5 years of independent data - Same experimental protocol - Target "hit rate" classifications applied without modification - Pre-specified analysis pipeline ### Randomization Target presentation in the online experiment was randomized via pseudorandom number generator out of a pool of possible targets. --- ## Sampling Plan This is an archival database. ### Existing Data **Registration prior to analysis of the data:** This preregistration covers confirmatory analysis of the 2021–2025 dataset. The 2000–2018 exploratory dataset was analyzed previously to establish classifications and identify semantic biases. ### Data Collection Procedures Data are collected continuously via the gotpsi.org website. Participants voluntarily complete remote viewing trials online. Each trial consists of: 1. Participant concentrates on an unknown target image 2. Participant types a free-response description 3. After entering the description, an image is randomly selected and displayed ### Sample Size The confirmatory dataset includes all valid trials collected between January 1, 2021 through ~ July 2025. No trials are excluded based on outcome. N ≈ 40,000 trials. ### Sample Size Rationale Sample size is determined by the natural data collection window (5 years), not by power analysis. This ensures temporal independence from the exploratory dataset and prevents optional stopping. ### Stopping Rule Data collection is bounded by calendar dates (2021–2025), not by results. Analysis proceeds after December 31, 2025, regardless of accumulated sample size or interim patterns. --- ## Variables ### Manipulated Variables None. This is an observational study of an existing experimental paradigm. ### Measured Variables **Primary outcome:** - Semantic similarity percentile score between participant description and target image, computed using FastText embeddings and cosine similarity. **Covariates:** ### Indices Semantic similarity is computed as: 1. Participant and target image text descriptions encoded via FastText 300d embeddings 2. Cosine similarity calculated between text and image embeddings 3. Percentile rank of target similarity among actual vs. all other targets in the pool --- ## Analysis Plan ### Statistical Models **Primary analysis:** Pearson correlation between mean percentile scores per target based on user descriptions for trials in first and second epochs. ### Transformations Raw cosine similarities transformed to within-trial percentile ranks (0–1 scale) to normalize across trials with varying baseline similarities. ### Inference Criteria - Primary hypothesis: p < 0.05 (two-tailed permutation test) - ### Data Exclusion **Pre-specified exclusions:** 1. Trials with empty or null text descriptions 2. Trials with technical errors (missing information) 3. Duplicate trials from same IP address within 30 seconds 4. Other trial cleaning filters specified for the permutation analysis **Debiasing procedure:** - Words are cleaned using already fixed filtering procedures - Debiased words were NOT removed from this analysis ### Exploratory Analysis (Coverage) Any analyses not specified above will be labeled as exploratory, including: --- ## Other ### Other **Temporal separation:** A gap of approximately 2 years (2019–2020), and use of different online architectures, separates the exploratory and confirmatory datasets to ensure complete independence. **Code availability:** Analysis code (MATLAB) will be deposited on OSF upon publication. **Data availability:** Deidentified trial-level data will be made available upon publication, subject to IRB approval. ### References - Bem, D. J., Tressoldi, P., Rabeyron, T., & Duggan, M. (2015). Feeling the future: A meta-analysis of 90 experiments on the anomalous anticipation of random future events. F1000Research, 4, 1188. --- ## Registration Timeline - **Exploratory analysis completed:** [DATE] - **Target classifications frozen:** [DATE] - **Preregistration submitted:** [DATE] - **Confirmatory analysis begins:** After preregistration acceptance - **Expected completion:** Q1 2026