Findings

What our experiments found, one finding at a time. Each page leads with where the result stands, puts the strongest objection before the argument, marks every checkable claim with its evidence, defines each technical term where it is first used, and ends with the dated trail of corrections, the references both ways, the steps to repeat it and a Verified line. Each offers its whitepaper as a PDF and its BuildTeX source.

30 findings so far. 12 of them are honest nulls: tests that failed, claims that did not hold, and errors of our own that we found and fixed. 6 findings in all report a null result with its bound. They carry the same weight as the results that held, because knowing what does not work saves the next person the trip.

Strong signal: real data, working controls and a result that held, with a named next test.
Signal: a real effect on real data or a sharp known-answer test, with clear limits.
Honest null: a test that failed, a claim that did not hold, or our own error found and fixed.

After every link to a finding, here and anywhere on the site, its standing: what later records in our findings ledger did to it. held extended narrowed in part questioned refuted retracted A link with no mark is to a finding no later record has treated.

Every technical term is defined where a page first uses it, and again in the findings glossary.

Each mark is one finding, earliest on the left: strong signal, signal, honest null. Choose a mark to read it.

30 findings shown.

F675 Honest null prediction failed Plants, trees and earth, Method and statistics

A panel test read climate driving tree rings correctly, and our own rule still failed it narrowed in part

On 161 northern tree-ring sites a panel-wide lagged test pointed from summer temperature to ring width, the known direction, with a wide margin. Two of its locked checks missed their bars by less than 0.02, so by the rule we wrote first the method earned no standing in this round.

What fell. Two binding first-difference checks failed by less than 0.02, so the preregistered kill fired and the panel instrument earned no standing in this round.
F515 Signal held Information and signal, Computing and mathematics, Language and literature

A small language model squeezes the UDHR to 0.27 bits a character, and fresh text to 1.59

Used as a compressor, SmolLM2 encodes the English Universal Declaration of Human Rights at 0.27 bits per character, far below any classical compressor. On 30,000 characters written after its training it needs 1.59, which points at the model having seen the declaration. One fresh text cannot say how large that share is.
F508 Signal held Information and signal, Computing and mathematics, Language and literature

Shannon's 1951 letter statistics hold on old and new English; his third-order value reads high

Shannon's first- and second-order entropy figures for English letters match ours within 0.12 bits on books from 1788 to 1925 and on present-day Wikipedia. His third-order figure sits 0.23 to 0.40 bits above ours, and an n-gram reader never gets near his 1.3 bits per letter.
F362 Signal held Instruments and control, Physics, Computing and mathematics, Method and statistics

Published 1 sigma error bars hold near nominal; a 95 percent interval needs 3 to 5 sigma narrowed in part

When a particle mass or an atomic mass moves between editions of its reference table, the old 1 sigma error bar covers the new value about as often as it should (0.68 and 0.66). The tails are far heavier than a bell curve, so covering 95 percent of later values took 3 to 5 times the stated sigma in these tables. Six of our ten forecasts held; we expected error bars to cover too rarely, and counted over all values they covered 0.79.
F344 Honest null honest null Physics, Computing and mathematics, Method and statistics

No cubic pattern in public Auger cosmic-ray directions, at sensitivity that sees only large ones narrowed in part

If space were a fine cubic grid, the highest-energy cosmic rays might arrive in a pattern aligned with its axes. In the public Pierre Auger sample we find no such pattern at any of five energy cuts, and our planted tests show the search could only have seen a strong one.

What it rules out. Only strong cubic patterns are ruled out: power 0.9 at planted strength 0.2 at 8 EeV, and at 32 EeV only at strength 1.0, where the sky along the cube axes is twice the average.
F329 Signal mixed Psychology and perception, Society and economics

In held-out private die-roll reports, about 45 percent read as honest, and this method cannot tell us why narrowed in part

Payoff maximisation says everyone in a private die-roll task should claim the top number. Of 58 published studies with 27,039 subjects, 19 papers with 9,960 subjects were held out. In their tasks with three or more outcomes the reports read as about 45 percent honest, 26 percent claiming the maximum and 29 percent in between; the result file gives no interval and the reading depends on the method. Six competing theories fit those reports almost equally well, and our planted check failed, so the tie shows only that this method could not separate them.

What did not hold. The planted check failed: the method did not recover a known model at this sample size (the planted model trailed its rival by 0.0003 nats per report). By the preregistered kill condition no ranking among the six models counts. The reports do differ from a shuffled control, and the shares read from the held-out reports are a description.
F320 Signal held Method and statistics, Psychology and perception

A random split scores a stress classifier perfect on data with nothing to learn narrowed in part

On planted data where nothing carries from one person to the next, a random train-test split still gives a perfect score of 1.00, while holding out whole people gives chance. On the real WESAD recordings the classifier does hold across 15 people, at an AUC of 0.966, and how stressed people said they felt did not track how much their bodies changed.
F310 Signal mixed Physics, Computing and mathematics, Method and statistics

Titius-Bode regularity in 122 exoplanet systems sits in gap sizes; their order adds nothing narrowed in part

Exoplanet systems do look more regular than random gaps would make them, but shuffling each system's own gaps into a new order leaves that regularity in place. A planet's position is predicted better by its two neighbours than by a Titius-Bode fit.

What did not hold. The spacing signal is real: the systems beat gaps pooled from the whole catalogue (p 0.0005). The order and law-like verdicts are null: against each system's own gaps put in a new order none of the three forms scored above the shuffled range (p 0.85, 0.995 and 0.68).
F308 Honest null honest null Computing and mathematics, Method and statistics

Checking 213 OEIS recurrence conjectures against their b-files finds one failure, already known narrowed in part

We tested a random sample of conjectured recurrences in the On-Line Encyclopedia of Integer Sequences against the long term lists stored beside them. One of 213 failed, and the entry itself had recorded that failure in 2016, so the sweep produced no new counterexample.

What it rules out. One failure in 213 tested lines, 0.47 percent; the 95 percent Wilson interval runs to 2.6 percent, so failure rates above that are ruled out for lines of this kind.
F281 Honest null honest null Physics, Computing and mathematics

A home rerun of the GRB 090510 light-speed test, and sound waves that keep their own limit narrowed in part

Using public Fermi photons from one gamma-ray burst, we bound any energy-dependent change in light speed to under 7.6 parts in a quintillion at 30 GeV, about 24 times weaker than the published limit from the same burst. In a simulated fluid, sound obeys its own speed limit at long wavelengths and leaves it at short ones by a known law.

What it rules out. Rules out a linear energy-dependent change in light speed larger than 7.6e-18 of c at 29.9 GeV, an energy scale below 0.32 Planck energies; the published bound from the same burst is about 24 times stronger.
F190 Honest null prediction failed Biology and medicine, Society and economics

Our model of cooperation on a grid had been rewarding selfishness

We set out to reproduce a textbook result, that living near relatives lets costly cooperation survive cheaters, and our model appeared to succeed. It had a flaw: cooperators were paid back by their own contribution. With that removed, cooperation collapsed at every benefit-to-cost ratio up to 50 under our update rule, and reached a share of only 0.48 at a ratio of 400.

What fell. With goods paid to others only, local producers never reached the preregistered mean share of 0.5 (0.48 at a benefit-to-cost ratio of 400).
F183 Honest null mixed Computing and mathematics, Method and statistics

After we fixed our own bug, CMA-ES ranked first and two nature-inspired optimisers proved centre-biased

Our first benchmark said optimisers named after slime moulds, wolves, bees and ants beat the standard methods. The win came partly from a bug in our own CMA-ES code. Fixed, CMA-ES ranked first on the shifted tests, and two popular nature-inspired methods turned out to work well only when the answer sits at the centre of the search box. As groups, plain and nature-inspired methods did not differ (mean rank gap 0.09, p 0.45).

What did not hold. The group contest between plain and biologically labelled optimisers sat inside the label-permutation range (p 0.45); the method-level ranks hold.
F172 Signal mixed Psychology and perception, Method and statistics

A selection model shrinks the overstatement in published effects but cannot pick which replicate extended

Fitted to published original results from other fields, a model that accounts for the significance filter predicted replication far better than taking originals at face value. In a held-out field it did no better than the plain base rate, and a small logistic model beat it.

What did not hold. The model beat taking originals at face value (the 95 percent interval of the Brier gain starts at 0.108). It did not beat the base rate (the interval starts at -0.039) and it lost to a small logistic model. Skill was positive in half the fields against a bar of three quarters, and its Reproducibility Project rate missed by 0.16 against a tolerance of 0.05. The preregistered kill rule fired, so the claim that the model picks which effects replicate is withdrawn.
F166 Honest null prediction failed Psychology and perception, Society and economics, Method and statistics

Our combined value formula lost to plain prospect theory on 11,759 risky choice problems

We tested our own formula for how people value risky options, prospect theory plus terms for information, threat and drive, against a richer version of prospect theory on the public choices13k data. It never beat that richer model by the margin we set in advance (with feedback it led by 0.001 against a bar of 0.005), so we retired it. Fitted loss aversion came out at 1.16 when people chose from descriptions and 0.89 when they learned from feedback.

What fell. The combined formula never beat flexible prospect theory by the preregistered 0.005 in cross-validated R2 (0.634 against 0.645 without feedback, 0.597 against 0.596 with it).
F152 Honest null prediction failed Method and statistics, Information and signal

Finding missing links by shared neighbours lost to simple popularity on three corpora held

The classic way to hunt for undiscovered connections, ranking unlinked pairs of ideas by the neighbours they share, found fewer of the links that later appeared than just pairing the most common ideas. Among pairs of equal popularity it still carried some signal.

What fell. Shared-neighbour ranking found fewer later links than popularity on all three corpora; among pairs of similar popularity it still separated later links (AUC 0.569 to 0.757).
F149 Strong signal held Society and economics, Plants, trees and earth, Biology and medicine, Method and statistics

Did unsupported citations about fungal networks double? It depends on the weighting

A 2023 paper reported that citations overstating the evidence for forest fungal networks roughly doubled over time. On the authors' own deposited table, their model gives about 2.2 and 2.8 times, while weighting each year by its number of citations gives about 1.8 and 1.5, and only one of the four intervals clears 1.
F147 Signal held Optics, colour and imaging, Instruments and control

Kodak's own curves put modern colour negative at 0.6 to 4.9 percent quantum efficiency narrowed in part

Reading three Kodak VISION3 data sheets as if the film were a photon counter, the best layer uses about 5 percent of the light's information and the median layer about 2 percent. In the green layer the 500T and 200T stocks match within 10 percent (4.5 and 4.9 percent) although one needs about twice the photons of the other; their blue layers differ by 11 percent and their red layers by 26.
F132 Honest null prediction failed Plants, trees and earth, Information and signal, Method and statistics

Transfer entropy could not tell which way climate and tree rings are linked questioned

Summer warmth drives tree growth, with no meaningful effect the other way, so tree rings make a clean test for tools that claim to read causal direction from data. A standard transfer entropy estimator failed that test on 135 sites, including on synthetic data where we planted the true direction ourselves.

What fell. At the real coupling strength the false direction ranked at or above the true one at 52.6 percent of planted sites, so the estimator gets no direction-reading status on data like these.
F114 Strong signal held Plants, trees and earth, Biology and medicine, Information and signal

How much summer warmth one tree ring remembers, measured in bits

A single Alpine tree-ring site keeps about a tenth of a bit per year about summer warmth on years it never saw, most of it in the density of its latewood. Averaging many sites buys far more than adding a measurement or one more tree past the usual replication threshold.
F97 Strong signal mixed Plants, trees and earth, Biology and medicine

Tree rings track the broad drought pattern in a gridded record and almost none of its local detail

Across 664 western North American tree-ring sites, the trees pass almost none of the summer drought signal at scales below about 55 km and most of it above about 460 km, measured against a gridded climate record. That grid is itself smoothed over hundreds of kilometres, so part of the missing fine detail may belong to the grid and not to the trees. Three further tests on how drought moves through space and time came back null.

What did not hold. The broad-against-fine contrast held: 0.64 against control 95th percentiles of 0.22 and 0.30. The three tests on the drought field itself, separability, travel and dispersion, came back null on the climate grid and on the trees, so six of the eleven preregistered verdicts failed.
F81 Signal held Biology and medicine, Optics, colour and imaging, Plants, trees and earth

Among colours people cannot tell apart, modelled bees and blue tits split flowers from backgrounds more often narrowed in part

On 728 measured natural surfaces, modelled bees and blue tits separate flower-against-background pairs that look identical to people more often than background-against-background pairs. No animal was tested; the observers are models. The direction held on a second surface set and under a second noise model; the size of the effect did not travel.
F76 Strong signal held Method and statistics, Information and signal, Instruments and control

How flat a spectrum is predicts forecast skill, but real series follow it loosely held

A 1940s formula says how predictable a series can be from its spectrum alone. On 1,236 real series it ranked held-out forecast skill (rho 0.41) and held as a ceiling in about 94 percent of them, yet real series tracked it less tightly than copies with the same spectrum and scrambled phases.
F63 Honest null honest null Plants, trees and earth, Biology and medicine

Across 1,175 tree-ring sites, an earlier moderate drought did not detectably toughen trees narrowed in part

We tested whether trees that lived through a moderate drought came through a later severe drought better. Across 1,175 North American tree-ring sites they did not, detectably, and a rerun with a finer drought grid gave the same null.

What it rules out. The 95 percent interval runs to +0.090 in log resilience, about a 9 percent gain, so small conditioning effects stay open.
F50 Honest null honest null Biology and medicine, Psychology and perception

Childhood trauma scores showed no meaningful link to amygdala-prefrontal coupling in 187 adults

In a preregistered test on open brain-scan data, adults' childhood trauma questionnaire scores were unrelated to resting connectivity between the amygdala and the ventromedial prefrontal cortex (partial r -0.006). The result rules out effects of 0.20 or larger and leaves smaller ones between 0.10 and 0.20 open.

What it rules out. Correlations of 0.20 or larger in size are excluded at the preregistered bound (90 percent interval -0.129 to 0.117); true effects between 0.10 and 0.20 remain open.
F47 Strong signal mixed Method and statistics, Psychology and perception

More variables raise prediction skill, yet most targets stall far below their noise ceiling

Across trees, mood, music ratings, forecasters and a similarity task, held-out skill grew as we added measured variables and always beat a shuffled baseline. In four of the five targets with a valid ceiling it stopped between 0.18 and 0.65 of the way there. We forecast that a different kind of variable would move at least three of those plateaus, and it moved one of seven.

What did not hold. Four of the six preregistered verdicts failed. One target reached 0.8 of its ceiling against a forecast of at least four, and one plateau moved when a new kind of variable arrived against a forecast of at least three. Held-out skill did rise with added variables and beat its shuffled control in every target, and the two toy-model verdicts passed.
F39 Strong signal held Music and sound, Psychology and perception

Music with long ringing tails and dark spectra is rated calmer, and also less pleasant narrowed in part

Across 502 held-out music clips by 270 artists, sounds that ring on longer and lean toward low frequencies were rated lower in arousal, beyond loudness, genre and year. We forecast they would also be rated more pleasant; they were rated less pleasant instead.
F32 Strong signal held Method and statistics, Psychology and perception

Forecasters who skip questions score better, and the best filter reads the probability

In 410,451 forecasts from the Good Judgment Project, a filter that holds back forecasts by their stated probability cut the average error far more than one that reads the forecaster's own expertise rating. People who skip more questions also score better on their first answers, though a person who skips more in one year is not clearly better for it.
F27 Honest null honest null Psychology and perception

Mood that lingers and mood that bounces back were not shown to be separate traits

In four open phone-survey mood datasets, people whose mood carries over from one prompt to the next also tend to return more slowly after a spike. A two-trait model never beat a one-trait model, but the measures were too noisy for that test to have much power.

What it rules out. Split-half reliability ran from 0.03 to 0.47, so the factor test had little power; two separate traits are not ruled out.
F23 Signal held Plants, trees and earth, Biology and medicine

Fungal hub trees mostly track sampling, cedar heartwood rots slowest, wind damage evades prediction

In a public map of fungi shared between Douglas-fir roots, the hub trees are largely what the sampling and the number of links produce. In a 30-year log experiment, western redcedar heartwood lost density slowest of four species, and in a large forest inventory, wind damage ranked well but forecast almost nothing.
F14 Signal held History of science, Physics, Optics, colour and imaging, Computing and mathematics

Newton's grazing-angle ring table follows his fitted rule, which runs 27% low at 85°extended

Up to 60 degrees, Newton's table of ring thickness against viewing angle matches modern thin-film physics within half a percent. From 65 degrees to 90 degrees every row matches a rule he fitted himself, and at 85 degrees that rule sits 27 percent below what the physics gives.

Verified on 2026-11-13. The index and all 30 finding pages were read in full by AI model readers working for the author, looking for a stated result with its numbers and limits, nulls kept as nulls, local file paths, stale claims and broken text. A script followed all 7,245 internal links and anchors on the 32 pages of this section and found none broken. It sent one request to each of the 164 outside addresses the pages cite: 158 answered (124 of them are DOI redirects, which shows the DOI is registered and says nothing about the publisher's page), five arXiv links carried a stray full stop and failed, and one, the Atomic Mass Data Center, refused the script. The five links were repaired that day and answer now. The same pass repaired one table reference that had printed as raw markup on F329 and one chart label on F310 that had rounded p 0.995 to 1.00.

Not checked. No study was re-run and no number was recomputed from its result file in this pass; each finding's own Verified section says what was checked for it. The glossary was scanned by script and not read in full. The read also raised points about wording and scope on single pages that this pass did not settle; a correction joins a finding's Trail when it is accepted.

Corrected on 2026-11-14. A second pass took the points the read raised to each finding's result files. 25 findings now carry a dated correction in their Trail: numbers that disagreed inside a page were set to the value in the file, four outcome words changed from held to mixed where a kill rule had fired or most preregistered verdicts had failed, 13 titles or summaries were brought inside the limits their pages state, and staff shorthand was replaced with plain wording. Each header now shows the count of preregistered verdicts. Where the files could not settle a point the page says so. No result changed and no study was re-run. The PDFs and their BuildTeX sources are unchanged and still read as first published, so a corrected page can differ from its PDF until the paper is typeset again.

Looking for the shorter cards? Experiments lists the preregistered studies that came back null, one card each. Findings are listed with the other writing under Research notes and papers.