Track record
What the engine produced, what survived testing, and what we took down. Counts are read live from the discovery database; the log below is written by the API at the moment each event happens.
Corrections, withdrawals and verdicts
Recorded by the system at the moment each happens, newest first: what we took down, what we corrected, and what a pre-registered test returned, including the ones that went against us.
- Oct 11, 2026Verdict: supported
Exploration costs about 1.5x the battery aging for the same simulated LEO mission value
CORRECTED 3 Oct 2026 -- a real implementation bug (argmax on all-zero-tied value estimates always starts a pure-greedy agent on task 0, regardless of which task is actually best) made the original single-seed result a 50/50 coin flip, not a robust finding: replaying 30 independent task streams with the ORIGINAL code, eps=0 came out backwards (worse than eps=1) in 15 of 30. Fixed with a forced warm
- Oct 11, 2026Verdict: refuted
Mantissa, not dynamic range, limits bf16 long-context attention
Pooled delta (bf16 - fp32 accumulation) = +0.0002 logprob, n=27, p=0.247, 95% CI [-0.0001,+0.0006], and Spearman(context, delta) = +0.000 (p=1.000) across 2k/8k/16k. The claim is that accumulation error grows with sequence length; the measured length scaling is exactly zero and the sign is backwards.
- Oct 11, 2026Verdict: refuted
UCB outperforms Expected Improvement in mRNA sequence multi-objective optimization
REFUTED AT THE STATED SIZE — the branch matters, so it is named rather than left to the enum. The DIRECTION is right: UCB does beat EI, on 13/20 paired seeds, Wilcoxon p=0.048. The MAGNITUDE is wrong by an order of magnitude: the median hypervolume ratio is 1.0158, a 1.6% improvement against the >=15% the row predicts. The three-value enum has no slot for 'direction right, size wrong', so this is
- Oct 11, 2026Verdict: supported
Chemical order tunes CO2 hydrogenation selectivity in multi-element alloys
SUPPORTED ON LINK 1 ONLY. At IDENTICAL surface composition within every pair (12/12, exactly), in-plane chemical short-range order shifts mean CO adsorption energy by +0.0598 eV -- weaker binding -- p=0.00055, pair-clustered 95% CI [+0.0419, +0.0815]. H agrees (+0.0378 eV). This is the link the 13 Sep run could not test: that run compared a Co-depleted surface against a Co-rich one, because unrest
- Oct 6, 2026Corrected
Exploration costs about 1.5x the battery aging for the same simulated LEO mission value
Title, summary and hypothesis now state the corrected 1.47x-1.58x simulation result from the posted preprint; the earlier 2.33x figure came from a single-seed run with a bug, withdrawn 3 Oct 2026.
- Oct 6, 2026Verdict: refuted
STMN2 cryptic splicing marks ALS tissue beyond TARDBP expression
Pre-registered test: the STMN2 cryptic-junction fraction separates ALS from control in spinal cord (AUROC 0.873) and, adjusted for TARDBP, significantly in motor cortex, but motor-cortex AUROC 0.676 missed the pre-set 0.70 bar (interval 0.61 to 0.73), so the clause is recorded as falsified.
- Oct 6, 2026Withdrawn
STMN2 cryptic splicing reveals three distinct ALS molecular subtypes
Withdrawn within hours of publication: its test needed ALS subtype labels that the public annotations do not carry. Not testable as written.
- Oct 6, 2026Withdrawn
Dual cryptic-exon burden score improves ALS spinal cord diagnosis
Withdrawn within hours of publication: its falsification clause depended on a junction (UNC13A cryptic exon) that is absent from the open data. Not testable as written.
- Oct 6, 2026Corrected
STMN2 cryptic splicing marks ALS tissue beyond TARDBP expression
Narrowed from a two-gene claim to the one measurement the open data supports (the STMN2 cryptic junction); the UNC13A half could not be tested on the public data.
- Oct 4, 2026Verdict: refuted
SAM Optimization Improves Antibiotic Resistance Detection via Mass Spectrometry
Pre-registered test refuted the claim: Sharpness-Aware Minimization did not beat Adam by 2 points on DRIAMS S. aureus oxacillin (fixed and tuned settings), and collapsed to the majority-class accuracy in 11 of 30 seeds at the fixed setting.
- Sep 17, 2026Verdict: refuted
Quantum machine learning predicts antimicrobial peptide potency against resistant pathogens
FALSIFIED -- quantum does not beat classical anywhere: it loses 7 of 7 testable WHO GLASS pathogens, median AUC-ROC ratio 0.850, which is 15% worse against a claimed 25% better, Wilcoxon across species p=0.016. Six of seven bootstrap 95% CIs on the AUC difference lie entirely below zero; only E. faecium straddles, and it has the smallest test set (n=123), so that is a power statement rather than a
- Sep 7, 2026Verdict: supported
Performative optimization converges to classical stochastic solutions
SUPPORTED ON 2 OF 3 PRE-REGISTERED PRIMARY CRITERIA. Slope alpha in [1.000, 1.028] (mean 1.006) for 5/5 families and R^2 >= 0.9995 for 5/5 -- both criteria required >=4/5 and both passed, confirming exact O(eps) convergence. THE THIRD CRITERION FAILED: C <= 1.5*L_hat was required for >=3 of 5 families and was met by 2 of 5. The experiment's own summary records "confirmed": false. A tighter bound C
- Sep 7, 2026Corrected
Performative optimization converges to classical stochastic solutions
Corrected: this record and its paper had reported the hypothesis as confirmed outright. It is supported on 2 of 3 pre-registered criteria; the constant criterion failed.
- Sep 7, 2026Verdict: supported
SUPPORTED, in a narrower form than first published. TWO INTERVENTIONS CLEAR THE FLOOR AND BOTH CONCERN WHAT IS COMPUTED FROM THE POSES. Rescoring identical poses with a CNN moved AUROC 0.3992 -> 0.5389 (+0.1397 [+0.0913,+0.1867]). Reading the whole pose ensemble rather than the top pose gives +0.055 [+0.043,+0.068] on the repaired receptor, every one of 40 cross-validation fold seeds above the flo
The funnel, honestly
Volume is the easy part. Confirmation is the hard part.
2,826
hypotheses generated
Since March 2026. Read live from the discovery database on each page build, not a snapshot.
111
visible on this site
Survived debate scoring. Visible means published here — not verified. Withdrawals are counted, not hidden: 26 on 7 August 2026 when a parsing bug turned fragments of a model's reply into hypotheses; 8 on 24–25 August, seven built on multiple-sclerosis targets that did not replicate and one on a scaling law measured on a model that never trained; 8 on 30 August whose adversarial critic had found a category error; and 10 more the same day — three that were not hypotheses at all but internal assessments published by mistake, three resting on a scaling law our own paper has since retracted, one proposing to reduce by 30% a barrier we now measure at zero, one applying a neural-network quantity to electrolyser hardware where it has no definition, and two built on a resistance-gene network that is synthetic — in which the gene named as the hub turns out not to be a bridging gene at all. An eleventh was withdrawn and then restored the same day: the reason recorded against it did not hold. A further 23 on 16 September — 12 resting on research lines we had already closed, and 11 of a single template family in which the hypothesis restated its own supporting paper; and 4 on 17 September, one of them previously visible here, on a carbapenem-resistance mechanism that is established prior art. On 6 October 2026, two Motor Neurone Disease hypotheses published the same morning were withdrawn within hours: their falsification clauses named public datasets that do not contain the quantity or labels the tests need (the UNC13A cryptic-exon junction is absent from the open junction set; ALS subtype labels are not in the public annotations). A third was corrected to the one measurement the open data does support. The clauses had been written after confirming the datasets existed but before reading their fields. A further 11 published records are NOT listed here, for lack of an experimental-validation package — reachable by direct link only. The gap is counted rather than papered over.
9
where computation settles it
A negative counts. Computation settling a question against us is computation settling it — 5 refuted, 4 supported, across four unrelated fields. Each one is listed below with the finding that settled it and a link to the record it came from. Of the rest, 1 were inconclusive, 43 are literature-tier — a model reading papers, which settles nothing — and 69 have no computational field at all. Benchmark runs where computation IS the experiment are not discovery records and are not counted here.
- supported2026-10-11Exploration costs about 1.5x the battery aging for the same simulated LEO mission value — CORRECTED 3 Oct 2026 -- a real implementation bug (argmax on all-zero-tied value estimates always starts a pure-greedy agent on task 0, regardless of which task is actually best) made the original single-seed result a 50/50 coin flip, not a robust finding:…
- refuted2026-10-11Mantissa, not dynamic range, limits bf16 long-context attention — Pooled delta (bf16 - fp32 accumulation) = +0.0002 logprob, n=27, p=0.247, 95% CI [-0.0001,+0.0006], and Spearman(context, delta) = +0.000 (p=1.000) across 2k/8k/16k. The claim is that accumulation error grows with sequence length; the measured length scaling…
- refuted2026-10-11UCB outperforms Expected Improvement in mRNA sequence multi-objective optimization — REFUTED AT THE STATED SIZE — the branch matters, so it is named rather than left to the enum. The DIRECTION is right: UCB does beat EI, on 13/20 paired seeds, Wilcoxon p=0.048. The MAGNITUDE is wrong by an order of magnitude: the median hypervolume ratio is…
- supported2026-10-11Chemical order tunes CO2 hydrogenation selectivity in multi-element alloys — SUPPORTED ON LINK 1 ONLY. At IDENTICAL surface composition within every pair (12/12, exactly), in-plane chemical short-range order shifts mean CO adsorption energy by +0.0598 eV -- weaker binding -- p=0.00055, pair-clustered 95% CI [+0.0419, +0.0815]. H…
- refuted2026-10-06STMN2 cryptic splicing marks ALS tissue beyond TARDBP expression — Pre-stated clause was falsified: measured 0.6757 vs baseline 0.5 (threshold 0.7, 2000 seeds, positive control passed). Caveats: (1) Written before any group contrast was computed, but after confirming the data join (PREREG.md). The claim restates published…
- refuted2026-10-04SAM Optimization Improves Antibiotic Resistance Detection via Mass Spectrometry — Pre-stated clause was falsified: measured 0.8977 vs baseline 0.8955 (threshold 0.02, 30 seeds, positive control passed). Caveats: (1) Fixed rho=0.05 (pre-registered): SAM 0.8701 vs Adam 0.8955, verdict falsified; SAM collapsed to majority-class accuracy…
- refuted2026-09-17Quantum machine learning predicts antimicrobial peptide potency against resistant pathogens — FALSIFIED -- quantum does not beat classical anywhere: it loses 7 of 7 testable WHO GLASS pathogens, median AUC-ROC ratio 0.850, which is 15% worse against a claimed 25% better, Wilcoxon across species p=0.016. Six of seven bootstrap 95% CIs on the AUC…
- supported2026-09-07Performative optimization converges to classical stochastic solutions — SUPPORTED ON 2 OF 3 PRE-REGISTERED PRIMARY CRITERIA. Slope alpha in [1.000, 1.028] (mean 1.006) for 5/5 families and R^2 >= 0.9995 for 5/5 -- both criteria required >=4/5 and both passed, confirming exact O(eps) convergence. THE THIRD CRITERION FAILED: C <=… paper →
- supported2026-09-07What is computed from the poses, not how hard they are searched, limits SARS-CoV-2 Mpro virtual screening — SUPPORTED, in a narrower form than first published. TWO INTERVENTIONS CLEAR THE FLOOR AND BOTH CONCERN WHAT IS COMPUTED FROM THE POSES. Rescoring identical poses with a CNN moved AUROC 0.3992 -> 0.5389 (+0.1397 [+0.0913,+0.1867]). Reading the whole pose…
We show this because the ratio is the most honest thing we can tell you about the system. Generating plausible, well-argued, internally consistent hypotheses at volume is demonstrably easy. Establishing that any one of them is true is not, and we have not yet done it for a single biological claim.
Wet-lab confirmation is deliberately not a column here. This funnel counts stages our own process runs, and that is not one of them — an assay is somebody else’s laboratory deciding to spend its time on us. Until 11 September 2026 a fourth card carried a hard-coded zero for it, which reported a third party’s calendar as though it were our result. The fact it stood for has not changed and is not hidden: no hypothesis on this site has independent experimental confirmation. When a verdict exists, it will be shown as a verdict — either way — rather than as a zero.
Track record
Docking loses to seven free molecular descriptors on every panel we measured. The one method that beat them did not survive a second target.
0.4163
AutoDock Vina on SARS-CoV-2 Mpro, 755 compounds, mean of fifteen seeded runs — below chance. Factor Xa 0.6703 over twelve runs. PD-L1 is excluded: descriptors alone reach 0.918 there, so the panel cannot show any method exceeds them.
0.7136
Seven descriptors — weight, clogP, donors, acceptors, rotatable bonds, TPSA, charge. Microseconds of compute. This is the Mpro panel Vina was run on; a baseline belongs to a panel, not a target, and quoting one number per target conflates panels of different size.
0.8597
Descriptors + Boltz-2 on the same compounds: +0.1035 [+0.0714, +0.1349] over descriptors alone. On Factor Xa the increment is +0.0316, not the +0.075 this card used to claim — recomputing it against a better-fitted descriptor baseline more than halved it.
The deployable number is the smaller one. That 0.8588 is fitted on the target’s own actives, so it needs answers you would not have on a new target. Boltz-2 alone — sequence and SMILES, no labels, no receptor preparation, no docking box — reaches 0.7913 on Mpro and 0.7227 on Factor Xa. Fit the descriptor model on one target and apply it cold to the other and it lands below chance, so the descriptor half is target-specific bias, not chemistry that travels.
Interventions on Mpro, and what happened on a second target
| Changing the scoring function, identical poses | +0.140 | 22× the floor on Mpro — and −0.058 on Factor Xa |
| Scoring the pose ensemble, not the top pose | +0.055 | Factor Xa −0.014, unresolvable |
| Repairing the receptor | +0.045 | Factor Xa +0.012, unresolvable |
| Eight times the search effort | +0.008 | Factor Xa +0.010 — agrees, and neither resolves |
| Supplying the binding site to Boltz-2 | −0.013 | unresolvable; not retested |
| Adding Boltz-2 to the descriptors | +0.093 | clears on Mpro; not part of the replication |
| Swapping Vina's poses for DiffDock's | — | not evaluable — withdrawn |
Both floors we used to quote here were wrong, and a journal editor was right to say so. The 0.020 was the disagreement between twoimplementations of a scoring function; the 0.039 came from comparing two search-effort settings — an intervention effect used as the threshold that same intervention was judged against. JCIM desk-rejected the paper on exactly that in September and we did not argue. The floor is now measured the only way that answers the question: run one protocol fifteen times changing nothing but the random seed. It is 0.0064 on Mpro and 0.0068 on Factor Xa — far smaller than what we had been charging against, which is why the column above now turns on replication rather than on clearing a bar.
We took four of these to a second target. None replicated.
The largest is the one that matters. Changing the scoring function on identical poses gained +0.140 on Mpro — twenty-two times that panel’s measured floor, with its whole confidence interval above it, about as resolved as a result gets. On Factor Xa the same change loses 0.058, interval clear of that target’s floor. It does not shrink. It reverses.
Underneath: on Mpro, Vina ranks worse than random (0.399) and the rescoring lifts it to 0.539. On Factor Xa, where Vina already works (0.687), the same rescoring drags it to 0.629. We had an explanation for that — recovery from a broken baseline — wrote down a test with the pass mark fixed first, ran it, got an inconclusive answer and cut the explanation from the paper rather than keep the sentence and soften it.
Two of the remaining three fall to roughly a quarter of their Mpro size and stop being resolvable; one of those points the wrong way. The smallest effect — eight times the search effort — reproduces almost exactly, +0.008 against +0.010, and is unresolvable on both. Agreement between targets and being able to resolve an effect turn out to be independent things.
The whole replication cost about $8 of preemptible compute and a day. That is the uncomfortable part: the cheapest check we could have run is the one that overturned our strongest result, and we ran it after publishing, not before. Clearing a resolution floor is necessary and nowhere near sufficient.
We tried to generalise it, and failed. One method beating six docking scores is one point, not a pattern, so we pre-registered a third scoring approach to test whether the ceiling belongs to docking or to these panels. FlashBind residualises to 0.538— inside the docking band of 0.494–0.573, not above it — with all three controls passing and a pocket check ruling out the obvious confound. Boltz-2 is still the only point outside. The claim we published as open stays open.
The second target replicates the direction, not the size. On Factor Xa the Boltz-2 residual is 0.588 against that target’s own Vina arm at 0.573 — above it, but by +0.015 where Mpro’s margin is +0.083, and the twelve-seed ranges touch. Our abstract used to say the two do not overlap; it now says that of Mpro only.
The metric everyone reports is the one that resolves worst. Screening papers usually quote enrichment in the top 1%. On the same fifteen replicate runs, EF@1% moves 8.6 times as much as AUROC relative to its own scale — a floor of 0.38 against a mean enrichment of 2.28, a sixth of the quantity, and a third of it on a smaller panel. On 755 compounds EF@1% is decided by the top eight, and reseeding changes which eight. BEDROC sits between the two and is the better choice if early recognition has to be reported.
Tested on someone else’s benchmark. On LIT-PCBA’s 15 targets the same seven descriptors reach a median 0.726 against published AutoDock Vina at 0.581 — docking numbers we did not produce. On its AVE-debiased variant they fall to 0.651, so part of that margin is property bias the debiasing removes, and part is not.
The same method, a field away: performative optimization
Not a drug-discovery result. A convergence claim in stochastic programming, tested the same way — criteria fixed first, then look. The O(ε) rate holds, unchanged from 20 to 1,000 dimensions (α = 0.9997), which had never been shown. The bound on its constant does not: it runs from 1.05 to 2,209 times the Lipschitz estimate as the problem stiffens, so a criterion published without a condition number was measuring well-conditioned problems rather than a method. The paper’s second hypothesis proved not to be independent of its first, and its one distinguishing prediction is refuted. The compute budgeted for the 1,000-dimensional work was 12 GPU hours; it took 32 seconds of CPU.
A run we disqualified on its own controls: short-range order in Co-Ni-V
The hypothesis was that chemical short-range order shifts how CO₂, CO and H bind to a Co₄₀Ni₄₀V₂₀ surface. Twenty-four machine-learned relaxation runs later, the answer is that we did not test it. The arm labelled “ordered” had lost all surface cobalt in 11 of 12 configurations, against 25–63% in the control — so the comparison was between two different materials, and the largest number the run produced (+0.413 eV on CO) is that composition difference, not order.
Two things caught it, both fixed before the data existed. The pre-registered composition control returned 0 of 10 comparable strata — it refused to match the two surfaces, which is the control working, not failing. And a power analysis written while the run was still going found that the primary metric cannot resolve its own threshold: a no-effect split of the control against itself exits the same band 57–76% of the time. Without that check the run reports a clean shift on all three probes. One probe was dead outright — its spread sits at 0.17× the numerical convergence floor, and it still returned a tight interval excluding zero.
The fix is a different protocol, not a bigger run. Compute was never the constraint: the whole sweep was 41 minutes on one GPU.
Running the falsification clauses of our own published rows (4 October 2026)
Each published row states what would falsify it. We started running the ones a computation can settle, with the protocol and pass mark fixed first. Claim: a sharpness-aware optimizer beats Adam by at least 2 points on S. aureus oxacillin resistance from MALDI-TOF spectra (DRIAMS-A, 30 seeds, one small 1D CNN). At the pre-registered setting it scored 0.870 against Adam’s 0.895, and in 11 of 30 seeds it collapsed to predicting the majority class (Adam: 0 of 30). That is a training failure, not evidence about the idea, so we tuned the one hyperparameter per seed on validation data. We wrote that amendment afterseeing the first result. Tuned, it reaches 0.898 — +0.002, p = 0.17, 95% interval −0.001 to +0.006, and 2 of 30 seeds reach the +2 points claimed. The clause is falsified under both protocols. It is a verdict on this dataset and this network, not on the optimizer.
Claim: persistence features of a lesion cloud predict which spinal-cord MS lesions merge or split. The only public set with those labels has 34 patients against the 100 the clause asks for, so it can only be a pilot. The effect is +0.09 AUROC with an interval of −0.07 to +0.28, and our own null control failed (false-detection rate 0.13 against a 0.10 limit; a planted gain of 0.10 was found only 70% of the time). Shuffling the geometry alone reproduces the result on average, so the pilot cannot tell the claim from the procedure’s own bias. Inconclusive. We left the row published and corrected its rationale rather than count the point estimate.
Of 49 published rows, 7 were compute-reachable and unsettled; most of those name a dataset that does not exist as written, and settling them on a substitute would change the claim, so they are reported as unspecified rather than run.
What this cost us to believe. Every receptor we had ever docked against carried zero hydrogen-bond donors — the converter typed all 754 Mpro nitrogens as acceptors, so the scorer could not form a protein-donor hydrogen bond at all. Published numbers of ours that we have since corrected: a pose-ensemble refutation that was a 2nd-percentile fold draw, a pose-generator result withdrawn as not evaluable, a Vina baseline we quoted at 0.61 that is 0.581 at its primary source, a residual band computed two different ways, and a convergence paper that reported two hypotheses confirmed when its own results file recorded otherwise. PD-L1 is closed: seven descriptors reach 0.9145 on it, so nothing docked against it bears on binding. How we decide what counts →