About solver.press

solver.press is the public window into the AegisMind research discovery engine. It publishes scientific hypotheses generated and adversarially debated by AI, together with the experimental packages needed to test them. They are research leads. None has been confirmed in a wet lab.

How a discovery is made

1
Paper ingestion: arXiv and Semantic Scholar papers across 15+ scientific domains are embedded into a vector store. Bridge-detection algorithms identify papers that span multiple fields — these are where novel hypotheses hide.
2
Hypothesis generation: A five-model ensemble (Claude, GPT, Gemini, Grok, Mistral) independently proposes cross-domain hypotheses based on the bridging papers. Near-duplicate hypotheses are filtered using semantic similarity.
3
Adversarial debate: Models take turns critiquing each other's hypotheses. A debate score (0–1) reflects how well a hypothesis survives adversarial scrutiny. Only hypotheses above threshold are published.
4
Logical consistency check: Z3 checks whether the hypothesis is internally self-contradictory. About 62% of hypotheses pass. This is not empirical proof: a pass means only that the stated claim does not contradict itself, and we do not treat it as evidence for the claim.
5
Novelty check: A novelty checker queries the vector store and patent/literature sources for prior art, and hypotheses matching existing work are scored down. This is a keyword and embedding search, not an expert freedom-to-operate opinion: it reliably catches direct restatements but routinely under-penalises work that sits inside a large established field under different terminology.
6
Experimental Validation Package (EVP): High-confidence discoveries receive a Claude-generated EVP: a full experimental protocol including methodology, required datasets, success/failure criteria, cost estimates, and dependency map.

Model roles in debate

Claude: Analytical evaluator — reasoning consistency and boundary conditions.
GPT: Skeptical opponent — challenges assumptions and prior art.
Gemini: Creative synthesiser — cross-domain connections and novel framings.
Grok: Contrarian — adversarial edge cases and failure modes.
Mistral: Pragmatic analyst — implementation feasibility and testability.

Available models rotate based on API availability. The circuit breaker pauses models that are rate-limited and resumes them automatically.

What these discoveries are not

Discoveries on solver.press are AI-generated hypotheses, not validated experimental findings. The EVP provides a protocol for empirical testing — that testing has not yet been done. Treat them as high-quality research leads, not established science.

Docking scores are not binding evidence. They are enrichment heuristics that correlate weakly with measured affinity. We measured this on SARS-CoV-2 Mpro: 753 compounds, measured actives against measured inactives, on a receptor whose redocking control passes at the screening protocol. AutoDock Vina returned AUROC 0.427 (95% CI [0.381, 0.471]) — significantly below chance, ranking the compounds that do not inhibit Mpro as the stronger binders. Molecular weight alone scores 0.620 on the same panel, and our own PDBbind-trained model is no better than Vina (0.417). An earlier pre-registered run (git 92c6f8bf, 14 July 2026) reported EF@1% = 0.00, but it prepared every ligand rigidly with no conformational search, so it was never a fair test of docking; it is superseded by the numbers above. We report docking as a way to order what to test first, never as validation.
A Z3 pass is not evidence. Roughly 62% of hypotheses pass. It tells you a claim is not self-contradictory, which is a very low bar and says nothing about whether it is true.
A high debate score is not truth. It measures survival under argument. Several hypotheses that scored well were later refuted.
Our preprints are not peer reviewed. Zenodo and Research Square deposits are self-published — citable and DOI'd, but not vetted.
Most hypotheses land in crowded fields. Efflux-pump inhibition, β-lactamase inhibitor analogues and cathepsin inhibitors all have decades of prior work. The value is a narrow angle inside a known field, not virgin ground.

Where computation settles the question

The distinction that matters most on this site is between domains where a computation is the experiment and domains where it is only a prediction.

In machine learning theory, quantum systems, and mathematics, a converged numerical result is the finding itself — convergence rates, ergotropy under Nash-equilibrium detuning, and verified combinatorial searches need no wet lab, and we mark those numerically established.

In biology it is the opposite: a docking score or an MM-GBSA energy is a hypothesis about the physical world that only an assay can settle. Those are marked prediction — awaiting wet lab, and they stay that way no matter how good the number looks.

Published research

The Precision Tetrahedron: Loss Landscape Topology Across Number Formats and Multi-Target Drug Discovery — John Goodman, AegisMind Research, May 2026.

A 42-phase empirical study characterising inter-precision Linear Mode Connectivity barriers across FP32, BF16, FP16, and INT8. Key findings: a model-size scaling law placing ~10M parameters as the practical basin-separator for mixed-precision training; an isosceles precision triangle (FP32↔BF16 ≈ 0.014 eV vs FP32↔FP16 ≈ 0.150 eV); and a four-vertex INT8 precision tetrahedron replicated across transformer, LSTM, and ResNet architectures. Applied to drug discovery with Nash equilibrium combination optimisation, which selected EPTIFIBATIDE as a KPC-3 inhibitor candidate (Spearman ρ = 0.827, six therapeutic targets).

Partial retraction (v2, July 2026).The originally reported “three-way convergence” identifying EPTIFIBATIDE as a dual-target KPC-3/MSH3 candidate is retracted. The MSH3 docking oracle targeted the wrong chain of PDB 3THW (chain A is MSH2, not MSH3), so the three analyses shared one oracle — pseudo-replication, not independent corroboration. The KPC-3 result above is unaffected, and the precision/LMC findings are unaffected. The DOI below is the concept DOI and always resolves to the corrected version.

DOI: 10.5281/zenodo.20363635 (concept — always latest) · All 42 phases on Google v6e-8 TPUs via Google TRC

Code & checkpoints →

Research teams: access the discovery engine at aegismind.app →