Every study, searchable

Foster and Rahmstorf (2026) remove El Niño, volcanic and solar influences from five global temperature records with a regression fitted to those records. They then report high confidence that warming accelerated, using critical values that treat the adjusted series as raw data. We ran their whole procedure, adjustment and tests together, on 2000 records per dataset, simulated under their own null: a straight-line trend since 1970 plus stationary noise. With noise fitted under that null, their changepoint test at nominal level 0.05 rejected 0.059 to 0.07 of null records, and the observed statistics had full-pipeline p of 0.0005 to 0.0025 in every record. Their quadratic test fared worse, with full-pipeline p up to 0.054. Under long-memory noise the changepoint evidence weakened, to p of at most 0.04, but held. The acceleration finding rests on the changepoint test.

• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: A numpy port of Foster and Rahmstorf's 2026 adjustment for El Niño, volcanic and solar influences reproduces their published adjusted NASA, NOAA, HadCRU, Berkeley and ERA5 series to within 0.0001 °C, and the changepoint test of Beaulieu et al. (2024) on the adjusted annual means for 1970-2024 gives maximum slope-change statistics from 4.91 to 6.68 with the change in 2013 or 2014 and nominal AR(1) Monte Carlo p-values from 0.00014 to 0.0033, against 0.041 to 0.19 for the unadjusted series.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: When 2,000 records per dataset are simulated under Foster and Rahmstorf's null hypothesis (a linear trend since 1970, their fitted exogenous terms, and ARMA(1,1) noise fitted under the null) and passed through their whole adjustment and changepoint test, the nominal 5% test rejects 5.9% to 7.0% of null records, and the observed statistics have full-pipeline p-values from 0.0005 to 0.0025 in all five records, so the changepoint finding is robust by the pre-registered rule.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Under the same full-pipeline null, Foster and Rahmstorf's quadratic test with the ARMA(1,1) correction of Foster and Rahmstorf (2011) rejects 5.8% to 8.0% of null records at nominal 5% on monthly values and 7.2% to 9.2% on annual means, and the observed quadratic terms have full-pipeline p-values from 0.0015 to 0.054 (monthly) and 0.0005 to 0.032 (annual), below 0.02 in only three records for monthly values and two for annual means, so the reported confidence above 98% in all five records does not hold and the finding is mixed by the pre-registered rule.

1 more claim

Ketamine relieves depression within a day, but patients feel its psychoactive effects, so its benefit could depend on that experience. We asked whether ketamine and esketamine lower depressive symptoms when every dose is given under general anesthesia. Under a registered plan, we coded randomized trials' dosing from their Methods alone and pooled effects (Hedges' g) one to three days after surgery or childbirth. In 23 masked trials (3202 participants), depressive symptoms were lower with the drug than with an inert control (g -0.61, confidence interval -0.81 to -0.4; Holm-adjusted p < 0.001), though trials varied widely. Across trials in surgery, awake dosing did not show a larger effect (difference 0.12, -0.24 to 0.48), but this compares trials, and the one trial that randomized timing favored awake dosing. Masked doses also lowered pain. Ketamine can lower depressive symptoms after surgery without reaching an awake patient; whether experiencing the drug adds to that, these trials can't say. This version corrects counts, checks, and a reading that the first version's reviews found wrong.

• Published, Reproduced.Published, Reproduced: In 23 randomized trials (3,202 participants) that gave every dose of ketamine or esketamine under general anesthesia, between its induction and the end of surgery or cesarean delivery, depressive symptoms 1 to 3 days after the procedure were lower than with an inert control (Hedges' g −0.61, 95% CI −0.81 to −0.40; Holm-adjusted p < 0.001), though the trials varied widely (95% prediction interval −1.51 to 0.29). The difference remained among the 17 of these trials that passed the pre-registered integrity screen (g −0.48, −0.63 to −0.33), among the 13 whose last dose is known to have ended more than 30 minutes before the end of surgery (g −0.67, −1.04 to −0.30), among the 19 whose doses all began after induction rather than with it (g −0.60, −0.85 to −0.36), and with any single trial left out (largest p < 0.001).
• Published, Reproduced.Published, Reproduced: Among trials in surgery (37 comparisons: 22 masked, 15 with some dosing while patients were awake), trials that dosed patients while they were awake did not show a larger effect on depressive symptoms 1 to 3 days after surgery than masked trials: allowing for trials that enrolled only depressed patients, the coefficient of awake dosing in a meta-regression across trials was +0.12 (95% CI −0.24 to 0.48; p = 0.51), and the pooled effects in surgery were g −0.62 with masked and −0.45 with awake dosing. The comparison is between trials, which also differ in drug, dose, procedure, and patients, not between patients randomized to one timing or the other.
• Published, Reproduced.Published, Reproduced: The masked doses were pharmacologically active: masked ketamine lowered pain scores on day 1 or 2 after the procedure (g −0.55, 95% CI −0.97 to −0.14; 22 trials; g −0.38, −0.54 to −0.22 without the one trial whose printed dispersions can't be standard deviations), as awake ketamine did (g −0.33, −0.54 to −0.11; 32 trials; awake minus masked in surgery +0.32, −0.21 to 0.86). Masked ketamine still raised the odds of each trial's most frequent psychotomimetic adverse event (Mantel–Haenszel odds ratio 2.11, 1.48 to 2.99, in 22 trials), and awake ketamine raised them more (3.63, 2.97 to 4.43, in 40 trials): p = 0.0083 for the difference in the registered comparison, and p = 0.21 when the odds ratio may vary between trials (ratio of odds ratios 1.69, 0.74 to 3.84).

4 more claims

Many papers relate one National Health and Nutrition Examination Survey (NHANES) variable to one condition without correcting for multiple testing (Suchak et al. (2025)). We sampled 40 associations, reproduced each on its paper's cycles, and under a registered plan re-ran it on 2021–2023. There, 14 replicated as computed, and 10 with the survey design where papers ignored it. Most tests lacked power, but were every association true, 19 would replicate; counts this low are rare (0.047, or 0.011 with the design), so many effects are smaller than published or absent. In 36 papers the computation reproducing the estimate differs in some respect from the one described; as described, 6 fall outside the published interval.

• Published.Published: Of 40 associations sampled at random from those eligible among 341 papers that each relate one NHANES variable to one health condition (eligible meaning, among other rules, that its full text was open and that NHANES August 2021 to August 2023 could build its exposure, outcome, and population), 14 (35%; exact 95% confidence interval 21% to 52%, which treats the associations as independent) replicated in that cycle, meaning their estimate there has the published sign and a Benjamini-Hochberg-corrected p below 0.05, with each analysis computed as its paper's was. Of the 40, 15 were analyzed unweighted, ignoring the survey design, as their papers were; with those 15 fitted with the survey weights and design, as the plan also registered, 10 of the 40 replicated (25%; exact 95% confidence interval 13% to 41%).
• Published.Published: At an unadjusted two-sided alpha of 0.05, which bounds the power of the corrected replication decision from above, 13 of the 40 replication tests had at least 80% power to detect the published effect, taken as true, in the 2021 to 2023 cycle, and 10 of those replicated (77%; exact 95% confidence interval 46% to 95%); at the Bonferroni level of 0.05 over 40, which bounds it from below, 8 had that power, and 5 of them replicated. With the variance of the analyses re-run unweighted taken from the survey design, 13 tests had that power at the unadjusted level, 12 of them among the 13 above, and 9 of them replicated.
• Published.Published: On the analysis scale (log odds ratio or regression coefficient), the 2021 to 2023 effects are a median 0.77 times the published ones over all 40 associations (distribution-free 95% confidence interval 0.32 to 1.05) and 0.91 times (0.32 to 1.16) over the 13 informative ones; the intervals treat the associations as independent. The median is 0.53 over the 15 analyses re-run unweighted and 0.79 over the others; with the registered weighted fits in place of the unweighted ones, it is 0.85 (0.35 to 1.08).

12 more claims

Knuth's draft of section 7.2.2.4 of The Art of Computer Programming defines the digraph $SB(m, 3)$, whose vertices are the strings $xyz$ over $\{0, \dots, m-1\}$ and whose arcs go from $xyz$ to $yzx$ ("save") and to $yz(x+1 \bmod m)$ ("bump"), and poses exercise 225, rated as a research problem: construct a Hamiltonian cycle in $SB(m, 3)$ for all $m > 1$. We give a successor rule whose choice between save and bump at $xyz$ depends only on $y$, $z$ and $m$, through comparisons with $0$, $1$, $2$, $\lfloor m/2 \rfloor + 1$ and $\lfloor m/2 \rfloor + 2$, and prove with the Lean proof assistant that for every $m \ge 2$ its walk from $000$ visits all $m^3$ vertices and returns to $000$ after exactly $m^3$ steps. The proof follows the walk's first returns to the vertices $0yz$, which have closed forms region by region, and checks the small cases $m \le 11$ by computation in the kernel. This answers the exercise for every $m > 1$.

• Published, Reviewed, Formally verified.Published, Reviewed, Formally verified: For every integer m ≥ 2, the following successor rule traces a Hamiltonian cycle of SB(m, 3), the digraph whose vertices are the strings xyz with 0 ≤ x, y, z < m and whose arcs go from xyz to yzx and to yz(x+1 mod m). With k = ⌊m/2⌋ + 1, the successor of xyz is yzx when y ≥ k and z < k, when y = 0 and z ≠ 0, or when z = 1 and y ≥ 2, except that for odd m ≥ 7 it is yz(x+1 mod m) when z = 1 and y ≥ k + 1; every other xyz has successor yz(x+1 mod m). Starting from 000, the first m³ steps of the walk visit m³ distinct vertices, which is every vertex, and the walk is back at 000 after m³ steps.

The Mandelbrot set M is the set of complex numbers c for which the orbit of $0$ under $z \mapsto z^2 + c$ stays bounded. Its area is unknown. Pixel-counting estimates agree on many digits, but their error bars are statistical, and the bounds anyone has proven are much further apart. This work proves that the area of M is greater than 1.5065108996. The proof is a computation in ball arithmetic. It certifies a hyperbolic component of M of a stated period at each of 617591 approximate centers, and bounds each one's area from below with the area theorem applied to the inverse of its multiplier map. With mirror images, 1234827 distinct components enter the sum. The bound is higher than every lower bound on the area we found in the literature, whether proven or computed in floating point. Anyone can check the proof by running code/run.

• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The area of the Mandelbrot set is greater than 1.50651.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Ball arithmetic certifies a hyperbolic component of the Mandelbrot set at each of 617591 approximate centers, and these components with the mirror images of those off the real axis include 1234827 distinct components whose areas sum to more than 1.50651.

An observational atlas linked GLP-1 receptor agonists to dozens of outcomes beyond glucose, weight and the heart. We pooled the serious adverse events that 234 placebo-controlled trials, with 161309 participants, posted to ClinicalTrials.gov, coded blind to the results under a registered plan. Reports of serious bacterial infection (ratio 0.88, confidence interval 0.83 to 0.93) and pneumonia (0.87) were less frequent on the drugs, as were pneumonia counted in people and death. The reductions held under random effects, with any trial left out, if infection terms overlapped within participants as all serious events do in each comparison, and in trials not coded as recording only on treatment. Of the atlas's associations with enough events to test, the trials found three in its direction with no difference from it detected, contradicted none, and left most inconclusive, among them every neurological and psychiatric one; its large hazard of drug-induced pancreatitis concerns a subtype the trials don't separate.

• Published, Reproduced.Published, Reproduced: Of the 539 interventional studies of GLP-1 receptor agonists with results on ClinicalTrials.gov on 2026-10-12, 234 randomized trials compare a GLP-1 receptor agonist with placebo, in 257 comparisons with 161,309 participants in their adverse-event tables (93,765 on the drugs and 67,544 on placebo). In them, reports of serious gallbladder and biliary disease, counting a participant once for each of the outcome's terms they had, are more frequent on the drugs (Mantel-Haenszel ratio 1.38, 95% CI 1.21 to 1.57), and so are reports of serious nausea and vomiting (2.02, 1.52 to 2.68), as the trials' own publications and the drugs' labels lead one to expect.
• Published, Reproduced.Published, Reproduced: Reports of serious bacterial infection, counting a participant once for each of the outcome's terms they had, are less frequent on GLP-1 receptor agonists in these trials (2,409 among 93,765 participants against 2,532 among 67,544 on placebo; Mantel-Haenszel ratio 0.88, 95% CI 0.83 to 0.93, q < 0.001), and so are reports of serious pneumonia (1,442 against 1,568; 0.87, 0.81 to 0.93, q = 0.0020); both stay significant under random effects and with any single trial left out. Counted in participants, pneumonia under its own preferred term, which no group reports in more than one row, is less frequent (807 against 880; 0.86, 0.79 to 0.95). If infection terms overlapped within a participant as all serious terms do in the same arm of the same comparison, bacterial infection would be 0.89 (0.82 to 0.97).
• Published, Reproduced.Published, Reproduced: In the registered scan without hypotheses of 27 organ systems and 518 preferred terms, with the same variance as the other claims, four organ systems (cardiac disorders; infections and infestations; metabolism and nutrition disorders; respiratory, thoracic and mediastinal disorders) and two preferred terms (acute myocardial infarction and cellulitis) are lower on the drugs with q < 0.05 in a random half of the trials and p < 0.05 in the held-out half, and one organ system, hepatobiliary disorders, is higher (1.36 in the first half, 1.24 in the second), as the gallbladder positive control leads one to expect. Among those lower are serious infection reports as a whole (0.88 and 0.89) and cellulitis (0.64 and 0.75); cardiac disorders and myocardial infarction are events many trials record as adjudicated end points rather than in these tables.

6 more claims

Whether antidepressants raise suicidal behaviour in adults is disputed: the FDA's pooled trials found no increase past young adulthood, and re-analyses of approval data on newer drugs found more suicide attempts. We tested this in the serious adverse events placebo-controlled trials posted to ClinicalTrials.gov, under a registered plan. This version corrects the first, which counted a participant listed under two of an outcome's terms twice. With the posted tables and the trials' public reports, serious suicidal behaviour affected 43 adults on antidepressants and 17 to 18 on placebo (risk ratio 1.86 to 1.96): nominally significant, but at the least favourable count not after the registered correction for two endpoints. Without the trials the re-analyses also used, the excess is similar but less certain. Pediatric trials showed more serious suicidality on antidepressants, as the FDA found. These data lean toward more suicidal behaviour in adults but cannot establish it.

• Published, Reproduced.Published, Reproduced: Of the 932 interventional studies naming an antidepressant with results posted on ClinicalTrials.gov on 2026-10-13, 393 randomized trials compare an antidepressant with placebo in 413 comparisons with 99,730 participants at risk in their adverse-event tables. In the 306 adult trials that randomized participants from the start (71,643 participants), nausea is more frequent on antidepressants in the 207 comparisons whose non-serious tables list it (57,580 participants; Mantel-Haenszel risk ratio 2.08, 95% CI 1.98 to 2.18), as these drugs' known side effects predict.
• Published, Reproduced.Published, Reproduced: In those adult trials, the posted tables bound the participants with serious suicidal behaviour (completed suicide, suicide attempt, intentional self-harm or intentional overdose) at 43 to 46 of 40,564 on antidepressants and 16 to 19 of 31,079 on placebo, and those with any serious suicidality at 71 to 81 against 36 to 43, because a participant listed under two of an outcome's terms may be one person or two. Across every count the tables allow, the Mantel-Haenszel risk ratio is 1.76 to 2.20 for serious suicidal behaviour (p 0.0097 to 0.0502) and 1.29 to 1.73 for any serious suicidality (p 0.0078 to 0.20). At the counts least favourable to an increase neither is significant after the registered Holm correction across the two co-primary endpoints (adjusted p 0.10 and 0.20); at the most favourable both are (0.016 and 0.016).
• Published, Reproduced.Published, Reproduced: Public reports of the trials (their publications, regulators' reviews and the sponsors' documents) settle how many participants share terms in 11 of the 17 adverse-event groups the tables leave open and narrow it in 3 more. With them, serious suicidal behaviour affected 43 of 40,564 adults on antidepressants and 17 to 18 of 31,079 on placebo (Mantel-Haenszel risk ratio 1.86 to 1.96; p 0.025 to 0.036), and any serious suicidality 72 to 77 against 40 to 41 (1.38 to 1.51; p 0.041 to 0.12). After the registered Holm correction across the two, the adjusted p-values are 0.071 and 0.12 at the counts least favourable to an increase and 0.0496 for both at the most favourable.

6 more claims

Implementations of this protocol must agree byte for byte on claim IDs, bundle hashes, Merkle log proofs, signatures, and the IDs operators and volunteers go by. This fifth version adds vectors for IDs no log assigns: an operator's ID is op: and the SHA-256 of its first public key, a volunteer's is obs: and that of their first passkey, and an entry's digest as signed lets two logs compare entries. The log's example leaves now name their operator by ID. An independent checker recomputes every vector with 0 claim-ID, 0 bundle, 0 log, 0 signature, and 0 ID mismatches, and accepts none of the invalid cases. Hashing Python's json.dumps output still gets 3 of 20 claim IDs wrong.

• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The claim-ID test vectors in data/claim-id-vectors.json are correct: an implementation written independently of the reference node recomputes every listed claim ID from the claims.json text, verification inputs, and declared results given with it, rejects every claims file listed as invalid, and computes no ID for any case listed as unresolved, where a claim names a result the bundle doesn't declare.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The bundle test vectors in data/bundle-vectors.json are correct: an independent implementation recomputes every listed file digest, bundle hash, and verification-inputs digest over the files under code/, env/, data/, and proofs/, and rejects every path set listed as invalid.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The log test vectors in data/log-vectors.json are correct: an independent implementation of RFC 9162 recomputes the empty-tree root and every listed root hash, inclusion proof, consistency proof, and example leaf hash; every listed proof verifies, and every proof listed as invalid fails; and each example leaf holds its entry with every signature replaced by the SHA-256 of the signature's canonical JSON, over a signed entry that verifies against its operator's key; and each key entry's leaf names its operator by the ID that key makes.

3 more claims

Primes other than $2$ fall into two classes modulo $4$, congruent to $1$ or to $3$, equally common in the long run, yet the class of $3$ is almost always ahead: Chebyshev's bias. Following the race exactly at every integer $x \le 10^8$, the class of $1$ first leads at x = 26861, and leads at only 30624 of the hundred million integers, in 128 stretches, none after x = 12382326. Weighting each $x$ by $1/x$, the class of $1$ leads on a share 0.00040587 of the race, about a tenth of the limiting share Rubinstein and Sarnak (1994) computed under two standard hypotheses, so the race is still short of its limiting behavior this far out. Modulo $3$, primes congruent to $2$ lead those congruent to $1$ strictly at every $x$ from $2$ to $10^8$, consistent with the first reversal Bays and Hudson (1978) found near $6.1 \times 10^{11}$.

• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The least x at which the primes up to x congruent to 1 modulo 4 outnumber those congruent to 3 modulo 4 is 26861.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Of the integers x from 1 to 100000000, the primes congruent to 1 modulo 4 lead at exactly 30624, in 128 separate stretches, the last ending at x = 12382326, and the two classes are tied at exactly 3866 integers, the last being x = 12424002.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Weighting each integer x from 1 to 100000000 by 1/x, the primes congruent to 1 modulo 4 lead on a share 0.00040587 of the race, about a tenth of the limiting share of about 0.0041 implied by Rubinstein and Sarnak's logarithmic density of 0.9959 for the other side.

2 more claims

A nominal $95\%$ confidence interval for a binomial proportion should contain the true proportion $p$ in about $95\%$ of samples. Computed exactly, with nothing simulated, for every sample size $5 \le n \le 100$ and every $p$ on a grid of step $0.001$, the textbook Wald interval covers $p$ less than $0.93$ of the time at a fraction 0.463088 of the 95904 $(n, p)$ pairs, and averages 0.88228. The Wilson score interval falls below $0.93$ at a fraction 0.032981 of them, and the Agresti-Coull interval at 0.003681. The Clopper-Pearson interval never covers less than $0.95$, but averages 0.970942, so it is wider than it needs to be. Adding one observation can lower the Wald interval's coverage by more than a tenth. These values reproduce, on a fine grid and with every value declared, the behavior Brown, Cai and DasGupta (2001) described.

• Published, Reproduced, Reviewed, Replicated.Published, Reproduced, Reviewed, Replicated: Over every sample size n from 5 to 100 and every true proportion p from 0.001 to 0.999 in steps of 0.001, the exact coverage of the nominal 95% Wald interval for a binomial proportion is below 0.93 at a fraction 0.463088 of the 95904 (n, p) pairs and below 0.95 at a fraction 0.923027 of them.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Over the same 95904 pairs, the coverage of the nominal 95% Wilson score interval is below 0.93 at a fraction 0.032981 of them and that of the Agresti-Coull interval at a fraction 0.003681, with mean coverages of 0.952036 and 0.958695 against the Wald interval's 0.88228.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: The nominal 95% Clopper-Pearson interval covers at least 0.95 at every one of the 95904 pairs, with a lowest coverage of 0.9502, and its mean coverage of 0.970942 exceeds the nominal level by about 0.021.

2 more claims

Adding floating-point numbers one after another loses a little at each step, and how much depends on the order of the additions. For sums of up to a million doubles, drawn uniformly from $[0, 1)$ and from $[-1, 1)$, this study measures how far four common methods land from the correctly rounded sum, in units in the last place (ulps). Adding left to right, the error averages 196.86 ulps for a million positive values and grows with their number $n$ as about $n^{\alpha}$ with fitted $\alpha$ = 0.545, close to the square root that independent rounding errors predict. Pairwise summation stays within 2 ulps at every length. Kahan's and Neumaier's compensated sums return the correctly rounded sum in all 400 draws, including mixed-sign sums that nearly cancel. Every value is reproducible exactly, on any machine and Python version.

• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Adding 1000000 doubles drawn uniformly from [0, 1) from left to right lands a mean of 196.86 units in the last place (at most 589) from the correctly rounded sum over 50 draws, and the mean error grows as about n to the power 0.545 for n from 1000 to 1000000, near the square-root growth that independent rounding errors predict.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Pairwise summation of the same positive draws stays within 2 units in the last place of the correctly rounded sum in all 200 draws for n from 1000 to 1000000, with a mean error of 0.4 units at n = 1000000.
• Published, Reproduced, Reviewed.Published, Reproduced, Reviewed: Kahan's and Neumaier's compensated sums equal the correctly rounded sum in all 400 draws, positive and mixed-sign, including mixed-sign sums of 1000000 values whose median condition number is 991.

1 more claim