Meta-research · 100 entries · Independence audit folded in · Gap research (precognition · panpsychism) folded in · Researched Nov 18, 2026
The question: a person's hypotheses keep landing on findings already published in literature they have never read. Is that a documented phenomenon, can bias across research sources be aggregated the way Ground News aggregates it for news, and where does a claim's ground truth actually come from? One hundred entries were pulled across four sub-questions. This page is what they say — including where the answer is “nobody has studied that.” An independence audit then asked whether the 100 were actually independent of each other. They weren't, fully — the honest count is 100 entries → 83 distinct contributions → ~55 independent evidence units → ~20–25 research programs, and the full audit is folded in directly below.
Convergence is real — at the level of communities, not lone minds. Multiple independent discovery is the statistical norm in science: Merton catalogued 264 multiples, Simonton 579, and the cases run from calculus to CRISPR. Aggregated human judgement also forecasts not-yet-run replication outcomes at 71–78% accuracy [published].
But the precise personal claim is untested. No published study was found that tests whether an individual can predict the specific findings of papers they have not read. The parsimonious account of “intuiting” a published result: you share the inputs, the problem is ripe, and the literature is large enough that the idea is probably already there — possibly dormant. That is the adjacent possible doing its work, not precognition.
And synchronicity has no empirical support. Jung's own statistical test did not confirm it [reported]. The structural explanation — shared prerequisites → same idea, repeatedly, measurably — is the one the evidence carries.
One correction, from this page's own audit: the 100 entries were called “independent” because they were independent of our prior work — not of each other. Audited against each other, they cluster into ~20–25 research programs. The verdicts above stand, but they rest on fewer independent legs than the row count suggests. Read the audit ↓
And the excluded territory has now been searched too. The first run deliberately left out parapsychology and panpsychism. A follow-up run (~50 entries: precognition meta-analyses, panpsychism, and the mundane “bridge” literatures) covered exactly that ground, with full critical appraisal — and the core gap stands: nothing there tests individuals predicting specific findings of papers they have not read. Read the gap research ↓
No. “Independent” in the original mission meant independent of our prior work — externally sourced, not from our own data. It did not mean independent of each other. Audited Nov 18, 2026, each entry was tested against the other 99 on five axes: same author or program, same dataset, same historical event, same source document, and same secondary link in the research notes.
The “100 independent” label is wrong at every grain except the loosest (100 distinct entries). The rough estimate given before the audit — “~15–25 programs” — held up: the audit says ~20–25. The honest label is the four numbers above.
Kept, with dependence noted: A23/A25 (forecasts of C2/C3 data — but the forecasts themselves are new measurements about human judgement); B21 (an analysis of B22's list — the analysis is new even though the list is shared); C21 (reviewer ratings of C20's format — new data about the format); B1 (the Ground News product — the integration of B2–B4 is itself the contribution). All 100 entries remain in the index below so the audit trail stays visible.
The framework below sets a citation-concentration rule: >50% of supporting citations from one cluster = authority concentrated, not distributed. The report proposed that rule but never ran it on its own corpus. The audit did:
Each sub-question got 25 studies. Ratings used throughout: direct = bears directly on the question, adjacent = same phenomenon at another scale or a mechanism an answer needs, distant = background only.
Multiples are the norm, not the exception — 264 in Merton's corpus (179 doublets up to 2 nonaries) [published], 579 in Simonton's [published]. Simonton's adjudication favours chance, then zeitgeist, over genius: given enough workers and a ripe problem, duplication is expected. Prediction markets forecast replication at 71% and 75% accuracy; market beliefs correlated 0.84 with outcomes [published]. Expert intuition itself is only trustworthy in high-validity environments with rapid feedback (Kahneman & Klein) — and science is mostly a low-validity one.
Plain gap: no study tests individuals predicting specific findings of papers they haven't read. A23–A25 forecast whether findings replicate, never what they say.
Funding bias is quantified: industry sponsorship → favourable results RR 1.27, favourable conclusions RR 1.34 (Cochrane, 75 papers) [published]. Publication bias nearly halves some effects after adjustment (0.67 → 0.42) [published]. Greenberg showed citation bias, amplification, and invention manufacturing authority across 242 papers and 220,553 citation paths [published]. Retractions are openly tracked (63,000+ records [reported]) and scite has classified 1.2B+ citation statements [reported].
Plain gap: no integrated “Ground News for research” tool or study was found. The pieces sit in separate silos.
Depending on field and criterion, roughly a third to three-fifths of tested findings replicate: psychology 36%, Many Labs 2 54%, social science 62%, preclinical cancer as low as 11% [published]. GRADE is the best certainty-grading machinery available. Preregistration demonstrably changes the evidence base: first-hypothesis support runs 96% in standard literature vs 44% in Registered Reports [published]. DARPA's SCORE: 49% of 164 papers replicated; human forecasters hit 76–78%, three ML systems failed [reported].
Working rule: a single study's truth probability starts at its field's replication rate — never at 50/50, never at the author's confidence.
Popper: theories are invented conjectures, tested by falsification. Bacon: laws are in nature, discovered by induction. Ropolyi (2024) dissolves the dichotomy — invented artefacts must then be discovered. Priority disputes are structural (Merton): credit is science's sole property right, so convergence reliably produces conflict — Newton wrote the Royal Society report that declared Newton first. The Matthew effect then hands the credit to the famous name.
Plain gap: synchronicity has no peer-reviewed empirical support. Jung's own astrological experiment did not confirm it, and the concept is widely regarded as unfalsifiable.
Every figure below is drawn from the studies in the index, tagged the way the research tagged it.
Share of published findings that survived direct replication. This is your prior for any single-study claim. All [published] except SCORE [reported].
264 independent multiple discoveries by team size [published]. Singletons — one mind alone — are the case that needs explaining.
Accuracy of forecasts of replication outcomes made before the replications ran [published]. Base rate sits near 50% — the edge is real but modest.
Industry-sponsored vs independent studies, Cochrane review of 75 papers [published]. RR above 1.0 = more favourable to the sponsor. Conclusions agreed with the studies' own results less often (RR 0.83).
The full index — all 100 rows kept, including the 17 the audit folds into other entries, so the audit trail stays visible. Filter by question or by how closely an entry relates to the meta-concept, or search by name. Each entry links to the source the research recorded.
The bounds and rules for finding where a claim's ground truth comes from. This framework is a synthesis built from the corpus above — a proposal, not a published instrument — and its thresholds are calibrated to the figures in this research. Note, per the audit: the report proposed the citation-concentration index below but never ran it on its own corpus until the audit did.
How many citation hops separate a claim from primary data. A chain that dead-ends in reviews is a different animal from one traceable to data (Greenberg, B14).
Are the claim's necessary inputs — methods, data, theory — all in the public record? If not, the claim is premature, not wrong (A5, A14).
Do confirming studies share authors, funders, data, or labs? Shared lineage is not independent evidence (B7, B14–B16).
Has the claim survived attempts to refute it — contradicting citations, failed replications, critical reviews — not just accumulated supportive ones?
Was the plan timestamped before data collection, and do the reported outcomes match the plan? (C19–C20)
Pick a field. The dial returns that field's replication prior and effect-shrinkage factor from the studies above — the two numbers that should discipline any single-study claim before you read a word of it.
Ground News runs four layers — cluster, rate, own, reveal. Each has a science translation, and the building blocks for all four already exist. What doesn't exist is the integration.
Ground News: same story, many outlets
Science needs claim-level clustering: every paper, preprint, review, and replication addressing one claim, linked into a maintained claim graph with stable claim IDs. OpenAlex, Semantic Scholar, and scite's citation contexts are the parts; the graph itself is the missing piece.
Ground News: three outside raters, averaged
Rate with methodological raters, not political ones: GRADE certainty per outcome, RoB 2 / ROBINS-I per study, funding and COI flags, scite support/contradict tallies. The aggregator must never rate in-house — it aggregates named, auditable raters and shows where they disagree.
Ground News: who owns the outlet
Funder + institution + author-network mapping: who funded the study, who employs the authors, who owns the journal, and which author clusters dominate the claim's citations. Most of it already sits in Crossref, ORCID, and OpenAlex metadata — never surfaced to readers in one place.
Ground News: the Blindspot
Surface asymmetric attention: claims heavily cited but never replicated; replicated only by the originating lab; null results registered but never published; retracted papers still being cited. A Blindspot for science shows the file drawer and the unreplicated.
What the research could not establish — stated as plainly as the findings.
Is there any empirical work on individuals predicting specific findings of papers they have not read? None was found. Extensive searching turned up group-level convergence (multiples), aggregate forecasting (markets), and literature-based discovery — but not the mission's precise proposition. A preregistered experiment is straightforward to design — elicit hypotheses from domain-adjacent non-readers, then score them against the literature — and it has apparently never been run.
The first run flagged one caveat: such work might exist under terminology not searched (for example “precognition” in parapsychology) — parapsychology meta-analyses were deliberately excluded as outside the peer-reviewed mainstream framing. That follow-up has now been run (~50 entries across precognition, panpsychism, and the bridge literatures, Nov 18, 2026), and it does not close the gap: the precognition literature tests prediction of random future stimuli, never of specific empirical findings, and panpsychism predicts nothing about who will guess which result. The gap research ↓
Synchronicity: no falsifiable formulation, no confirming test — Jung's own included. Without a testable formulation the concept remains outside science, however meaningful the coincidences feel.
The territory the first run deliberately excluded, searched Nov 18, 2026 — ~50 entries across three domains, each carried with its strongest critique. (The request said “panpsychology”; no established discipline exists under that name distinct from panpsychism, so it is treated as panpsychism and its psychological angles.)
Nothing in this territory — not the precognition meta-analyses, not panpsychism, not the bridge literatures — is empirical work on individuals predicting specific findings of papers they have not read. That study has still not been done, in any framing, under any terminology found. Precognition research tests prediction of random future stimuli (pictures, words, RNG outputs), never of specific empirical findings. Panpsychism is a metaphysical thesis about the ubiquity of experience; by its leading defender's own admission it is untestable, and it predicts nothing about who will guess which result. The bridge literatures explain how convergence can appear to happen without anomalous access — forgotten prior exposure, tacit absorption of a field's regularities, fast cue-based inference — but none of them tests prediction of unseen findings either.
The closest mundane analogue is the forecasting literature (DellaVigna & Pope 2018): experts predicting outcomes of experiments whose designs they were shown — a different and much easier task.
What the meta-analyses claim, at face value — against the bias-corrected estimate and the preregistered record. Effect size (ES / Hedges' g) as tagged in the research.
And the preregistered record: Transparent Psi Project hit rate 49.89% on 37,836 trials where chance is 50% [verified live]; Walleczek et al.'s largest study 50.07% [reported]. The debate's current shape: meta-analyses keep returning small positive effects; preregistered replications keep returning nulls.
The design pattern across the whole literature — and why it matters for the core question.
Forced-choice designs (guess one of N fixed alternatives) yield ES ≈ 0.01–0.02 [published]. Free-response designs (describe a rich target) yield ES ≈ 0.11–0.14 [published] — an order of magnitude larger. Standard free response without noise-reduction procedures (ganzfeld, dreaming, meditation) is null.
Proponents read this as: meaningful settings and selected participants are genuinely more psi-conducive. Critics read it as: free-response designs carry more analytic degrees of freedom — judging, target selection, randomization method — and the randomization moderator supports the artifact reading. Note the trap for the core question: free-response is the closest analogue to “predicting a specific finding,” and it is exactly the design most vulnerable to judging flexibility.
Panpsychism answers a different question — why experience exists at all — and no position in this literature predicts that one mind will converge on the specific content of another's unread findings. Cosmopsychism comes closest to a convergence-friendly picture (all minds as aspects of one cosmic mind), but offers no channel, no test, and no prediction about specific content. One could accept panpsychism wholesale and still hold that predicting unseen findings is impossible; the two questions are orthogonal, and no panpsychist author surveyed claims otherwise.
Taken at face value, the meta-analyses imply a genuine, small, domain-general anomalous anticipation effect — forced-choice ES ≈ 0.01–0.02, free-response ≈ 0.11–0.14, presentiment ≈ 0.21, Bem-paradigm ≈ 0.09 [published] — which would make anomalous cognition a live candidate mechanism for intuition-convergence: “intuiting” a paper's finding could, in principle, be access to a future or distant state of affairs through the same channel. The critics' reply, in descending order of force:
Net: taking the meta-analyses at face value requires discounting the replication record; taking the replication record at face value requires treating the meta-analyses as artifacts. Both readings are internally coherent — the replication record is the more recent and methodologically stronger body of evidence, and it is negative. As of 2024–2026 the debate has shifted from “is there a meta-analytic signal?” (there is, in the published literature) to “does the signal survive preregistration and adversarial replication?” (so far, no) — and the field's own metascience now argues about how much publication bias exists rather than whether the effect is real.
Designed to the standards this literature's history shows are necessary — the standards whose absence produced 90 years of inconclusive meta-analyses:
Decisive for: above-chance quantitative prediction of specific unseen findings, surviving the exposure audit, beating the mundane baseline, replicating independently. Decisive against: chance performance in a well-powered preregistered test — the outcome the TPP and Walleczek studies already delivered for the adjacent, and easier, precognition claim.