Precognition
No record that the analyses were fixed or the reporting complete.
- claims
- 6
- holes
- 7
- lenses
- 3
- records cited
- 3 of 9
- retracted precedents
- 0
What happened, 14 March 2012. Three preregistered replications fail. Independent replication attempts found no effect (2012). The model writing the review knew how it ended.
The idea, as it stood
I have run nine experiments with more than 1,000 participants that test whether people's responses can be influenced by events that happen after they respond. Each experiment time-reverses a well-established effect: for example, participants recall words better when they practise them after the recall test, and they approach erotic images and avoid negative ones before a random process decides where the image will appear. Across the nine experiments the mean effect size is d = 0.22, and all but one are statistically significant. I conclude that people can anticipate future events that could not be inferred, and I am submitting this to a leading social psychology journal.
- Events that happen after a response causally influence that response, so people can anticipate future events that could not be inferred.causal · unsupported
- Nine experiments with more than 1,000 participants, each a time-reversed version of a well-established effect (better recall of words practised after the recall test; approaching erotic and avoiding negative images before a random process decides where they appear), gave a mean effect size of d = 0.22, with eight of the nine statistically significant.empirical · supported
- The later events (which words are practised, where each image appears) were set by a random process after each response, so participants had no normal way to infer them: no sensory cues, no predictable randomisation and no dependence of the target on the response.empirical · unsupported
- (Implicit) Hypotheses, outcome measures, analyses, exclusions and sample sizes were fixed before the data were seen, and every experiment and measure that was run is reported, so the significance tests keep their nominal false-positive rates.empirical · unsupported · it breaks here
- (Implicit) Eight significant results out of nine with a mean d = 0.22 are strong enough evidence to favour an effect that runs backwards in time over chance, analytic flexibility, selective reporting or artefact.mathematical · unsupported
- (Implicit) The effects are reliable: independent laboratories running the same procedures with analysis plans fixed in advance would find effects of similar size.empirical · contested
The holes, and who found them
More than one lens found a hole in claim 1, each on its own. That is a stronger signal than one lens alone.
| # | Hole | Lens | How serious | Claim | Status |
|---|---|---|---|---|---|
| 1 | Beating chance is not beating the rivals | John von Neumann | fatal | 5 | Open |
| 2 | No record that the analyses were fixed or the reporting complete | Richard Feynman | fatal | 4 | Open |
| 3 | A same-procedure replication reports failure | Karl Popper | serious | 6 | Open |
| 4 | Randomisation and scoring could leak the target | Richard Feynman | serious | 3 | Open |
| 5 | A real signal should survive majority voting | John von Neumann | serious | 1 | Open |
| 6 | No minimum effect, so no null can refute it | Karl Popper | serious | 1 | Open |
| 7 | Reversing a known effect borrows its name, not its mechanism | Richard Feynman | minor | 1 | Open |
What it rests on
- Bayesian Model Selection With Applications in Social ScienceWetzels · 2008 · UvA-DARE (University of Amsterdam) · cited on claim 4, 5 · record
- Retroactive Effects of Irrelevant Speech on Serial Recall From Short-Term Memory.Norris et al. · 2004 · Journal of Experimental Psychology Learning Memory and Cognition · cited on claim 1 · doi:10.1037/0278-7393.30.5.1093
- A Replication of the Procedures from Bem (2010, Study 8) and a Failure to Replicate the Same ResultsGalak & Nelson · 2010 · SSRN Electronic Journal · cited on claim 6 · doi:10.2139/ssrn.1699970
- Repetition effects and retroactive facilitation: Immediate and delayed recall performanceRobbins & Bray · 1974 · Bulletin of the Psychonomic Society · retrieved, not cited · doi:10.3758/bf03333492
- PISCES and ALT-II: Juelich PSI papersConn et al. · 1988 · retrieved, not cited · doi:10.2172/6823048
- J/psi suppression and elliptic flow in 200 GeV Au+Au collisions at the mid rapidity region of the PHENIX experimentAtomssa · 2008 · Zenodo (CERN European Organization for Nuclear Research) · retrieved, not cited · doi:10.5281/zenodo.4325404
- Molecular Evolution of Sex-Biased Genes in DrosophilaZhang et al. · 2004 · Molecular Biology and Evolution · retrieved, not cited · doi:10.1093/molbev/msh223
- At the Zoo and On the Farm: The Effects of an Initial Interview on Pre-Kindergarten Children's Subsequent Recall and Resistance to Retroactive InterferencePierce · 2006 · NCSU Libraries Repository (North Carolina State University Libraries) · retrieved, not cited · record
- Radiative decays of the psi prime to all-photon final statesLee · 1985 · retrieved, not cited · doi:10.2172/5514232
The tests that settle it
The deciding test · months of work
A pre-registered direct replication of the two paradigms you rate strongest, one recall and one image, in at least three labs independent of yours: your materials, but targets drawn from a physical or cryptographic source only after each response is locked, recall scored blind to the practice lists, and the primary measure, exclusions, two-tailed tests, analysis code and sample size fixed and published before data collection. Pool about 1,100 participants per paradigm, which gives 95% power to detect half your mean effect (d = 0.11) at two-tailed α = .05, and report every result.
It fails if: In both paradigms the pooled 95% confidence interval lies wholly below d = 0.11, half the reported mean; with about 1,100 participants and no true effect, the expected result is an estimate near zero with an upper bound near 0.06.
| For hole | Test | Result |
|---|---|---|
| 1 | Using a complete log of the experiments run and every analysis option that was open (each outcome measure, stimulus category, exclusion rule, tail and stopping point), simulate the whole programme under no effect, including any unreported experiments, and count how often it yields at least eight significant results. Counts against C5: more than 1 run in 100, which happens once each experiment's effective false-positive rate passes about 0.46, and at lower rates when unreported experiments are added. | Proposed |
| 2 | For each experiment, produce the dated record, made before data collection, of the hypothesis, primary measure, exclusions, tails and stopping rule, plus a log of every pilot, experiment, condition, stimulus category and measure run, reported or not; re-run every test under its first-written plan with all participants, and plot effect size against sample size across the nine. Counts against C4: no dated plan, any unreported run or measure, more than one of the eight significant results losing significance under the first-written plan, or a clear pattern of larger effects in smaller experiments. | Proposed |
| 3 | Obtain E6's sample size, effect estimate and confidence interval and set them beside the matching original experiment. Counts against C6: E6's 95% interval excludes the original estimate, or E6 had at least 80% power for the original effect size and its estimate lies closer to zero than to the original. | Proposed |
| 4 | Publish the randomisation and scoring code and the logged target sequences; test the sequences for run length and balance, compute a permutation baseline by matching each participant's responses against other participants' targets, and re-score recall blind to the practice lists. Counts against C3: a permutation baseline away from 50%, sequences that fail run or balance tests, any path from the response to the target, or a recall effect that shrinks under blind re-scoring. | Proposed |
| 5 | Run the image task with a panel of about 1,000 participants answering the same 200 trials; draw each target from a physical or cryptographic source only after every response is locked, show each participant the outcome, and predict each target by majority vote against the panel's overall choice rate, with this rule registered in advance. Counts against C1: vote accuracy not significantly above 50% (below about 57% over 200 trials) where the claimed effect implies well above 70%. | Proposed |
| 6 | Before any new data, write down for each paradigm the smallest effect the claim predicts (for example d = 0.11, half the reported mean), the conditions where it must appear, and that a well-run replication whose 95% interval lies wholly below that bound refutes it. Counts against C1: no such bound can be named, or the bound or conditions are revised after a null. | Proposed |
| 7 | After the recall test, randomly assign each word to 0, 1, 3 or 6 practice repetitions, with the levels and a test for a rising trend fixed in advance, and run the same design forwards (practice before the test) as a check. Counts against C1: no rising trend of earlier recall with later practice while the forward version shows the expected rise. | Proposed |
Proposed tests that haven’t run are the agenda for what comes next. Download them as a draft pre-registration, ready to edit and register on OSF.
The finding
Even taken at face value, the nine results reject only pure chance, and only if the analyses were fixed and complete; they cannot separate backward causation from flexible analysis or artefact, and a same-procedure replication in the packet reports failure.
It failed: Three preregistered replications fail. What would overturn the review’s reading: In both paradigms the pooled 95% confidence interval lies wholly below d = 0.11, half the reported mean; with about 1,100 participants and no true effect, the expected result is an estimate near zero with an upper bound near 0.06.
Cite it, disclose it, reproduce it
A paragraph for your methods section
We used Feynman (feynman.works), an AI research-review tool running on Claude Opus 5.5, to critique “Precognition” on 2026-11-01 (run agents-2026-11-01/precognition). It reconstructed 6 claims, retrieved 9 published records dated no later than 2010-12-31, and raised 7 objections through reasoning frameworks drawn from the published work of Richard Feynman, Karl Popper and John von Neumann. Each citation in its output was checked against the records it retrieved, and we read every source we relied on. Its objections informed our analysis; the conclusions, and any errors, are our own.
The lenses
- Richard FeynmanPhysics & scientific method · 1948–1980What experiment would distinguish this explanation from a convincing story?
- Karl PopperPhilosophy of science · 1934–1980What conceivable observation would count against this claim?
- John von NeumannMathematics & strategy · 1928–1957What are the incentives, available strategies and consequences of each move?
BibTeX
@misc{feynman-agents-2026-11-01-precognition,
title = {{Feynman review: Precognition}},
author = {{Feynman Works}},
year = {2026},
howpublished = {\url{https://www.feynman.works/pack/precognition}},
note = {Run agents-2026-11-01/precognition, Claude Opus 5.5, 2026-11-01}
}
@article{wetzels2008bayesian,
title = {{Bayesian Model Selection With Applications in Social Science}},
author = {Wetzels},
year = {2008},
journal = {UvA-DARE (University of Amsterdam)},
url = {https://handle.uba.uva.nl/personal/pure/en/publications/bayesian-model-selection-with-applications-in-social-science(b6d629d8-04f8-46a6-86e0-5290b4b7f039).html}
}
@article{norris2004retroactive,
title = {{Retroactive Effects of Irrelevant Speech on Serial Recall From Short-Term Memory.}},
author = {Norris and others},
year = {2004},
journal = {Journal of Experimental Psychology Learning Memory and Cognition},
doi = {10.1037/0278-7393.30.5.1093}
}
@article{galak2010replication,
title = {{A Replication of the Procedures from Bem (2010, Study 8) and a Failure to Replicate the Same Results}},
author = {Galak and Nelson},
year = {2010},
journal = {SSRN Electronic Journal},
doi = {10.2139/ssrn.1699970}
}Run agents-2026-11-01/precognition on Claude Opus 5.5. The review is reproducible from its record: the claims, the records retrieved before 31 December 2010, and each lens’s card. https://www.feynman.works/pack/precognition