Mantissa, not dynamic range, limits bf16 long-context attention
Est. first test: $45k · about 75 days
Source: AegisMind Research
Read full discoveryScience at the speed of computation
Most published findings are never re-run. solver.press is the public window into AegisMind, an engine built to change that. It generates hypotheses, stress-tests them in five-model debate, and settles what a computation can settle against a pre-registered protocol, in our own claims and in published papers. We re-test our own results on fresh seeds. Refutations, corrections and withdrawals sit beside what held, because a record that only shows wins isn't a record.
Everything here is a research lead; none has yet been confirmed in a wet lab. We are working with wet-lab partners to test selected leads, and we will report each result here, including when it goes against us. Where computation is itself the experiment, the numerical result stands on its own. In biology it does not, and we label it accordingly.
One claim per paper, tested with the authors' released code under a protocol deposited before the run. 18 listed: 13 replicated, 3 inconclusive, 2 did not replicate. Each outcome is the verdict under our protocol, not a verdict on the paper. 10 more papers have been run and are not listed: some are inside an author window, the rest have not been prepared for listing.
Yu and Zhou · arXiv 2611.02145
At 2.33M parameters on FineWeb-Edu with the repository's recipe, SiLU on the attention value projection lowers mean held-out perplexity by about 2.0% against standard attention, and every paired seed improves.
Pankka, Lehtinen and Roine · arXiv 2610.38558
Median inference latency of autoregressive WaveNet is about 52x REVBRAIN's at 1 sample per input and about 3170x at 100 samples per input (released latency script).
Crasmaru · arXiv 2610.08592
A fully complex FNet-style causal LM matches a param-matched real-valued FNet in under half the steps on tiny-shakespeare.
Last 30 days, since 2026-10-18
17
hypotheses published
Passed the publication gate. Drafts are not counted.
2
settled by computation
2 refuted, 0 supported. A negative counts. 9 settled to date.
Recently withdrawn, corrected or updated
last 30 days: 3 withdrawn, 4 corrected, 1 updated. Withdrawn = claim taken down; corrected = something we published was wrong, now fixed; updated = claim revised or narrowed, nothing false.
Read the full record → · We have also re-tested 4 of our own published results on fresh seeds: 4 agreed. See them →
Discoveries that have had a computation run against them, most recent first, whether it supported the hypothesis or refuted it. Where that computation is itself the experiment it is marked verified; where it is a structure-prediction score against a protein target it is marked as docking only, which is a screening result and not evidence of binding.
Est. first test: $45k · about 75 days
Source: AegisMind Research
Read full discoveryEst. first test: $9k · about 35 days
Source: AegisMind Research
Read full discoveryEst. first test: $18k · about 75 days
Source: AegisMind Research
Read full discoveryEst. first test: $1k · about 30 days
Source: AegisMind Research
Read full discoveryEst. first test: $9k · about 45 days
Source: AegisMind Research
Read full discoveryEst. first test: $45k · about 150 days
Source: AegisMind Research
Read full discoveryThe engine's newest output, running since March 2026. These are untested leads that survived adversarial debate. They are not findings, and most have no computed result yet.
Untested hypothesis
Est. first test: $45k · about 270 days
Source: AegisMind Research
Read full discoveryUntested hypothesis
Est. first test: $18k · about 75 days
Source: AegisMind Research
Read full discoveryUntested hypothesis
Est. first test: $45k · about 150 days
Source: AegisMind Research
Read full discoveryUntested hypothesis
Est. first test: $8k · about 35 days
Source: AegisMind Research
Read full discoveryUntested hypothesis
Est. first test: $18k · about 75 days
Source: AegisMind Research
Read full discoveryUntested hypothesis
Est. first test: $45k · about 70 days
Source: AegisMind Research
Read full discoveryThe full corrections and verdicts log, what has been settled, and the aggregate results, including those against our own numbers.
How a hypothesis is generated, debated and tested, and what each stage does not establish.
Send one. We test it with a pre-registered computation on named public data and report the result whether or not it supports you. Our own claims are tested the same way.
Request a test →Research teams and organisations interested in domain-specific discovery runs can get in touch via aegismind.app.