Prove it on paper before you prove it in a lab.
Our first milestones are entirely computational: demonstrate a real, measured advantage on public data the model has never seen — before commissioning a single blinded experiment.
We think the honest version of this field is unglamorous and powerful: map the data, respect the baselines, make predictions you can be wrong about, and only spend on experiments once the computation has earned it.
A computation-first path, with a real go / no-go.
If the models don’t consistently beat serious baselines on fully withheld data, we don’t advance to expensive experiments just because the output looks sophisticated.
- Stage 0
Audit the evidence
Build a verified, open inventory of public datasets — what exists, how good it is, and how the pieces connect. No modeling claims until the ground is mapped.
- Stage 1
Reproduce strong baselines
Re-run the best existing methods ourselves, honestly. If we can’t beat a serious baseline, we don’t get to claim a new one.
- Stage 2
Blind prediction on public data
Freeze predictions before looking at the answers. Measure against real, independently generated results — and count the misses.
- Stage 3
Simulate limited experiments
Replay dense published experiments in software: hide the outcomes, let the system “buy” a few measurements, predict the rest. This tests whether we’d really need fewer assays.
- Stage 4
Freeze one testable claim
Before any lab spend, pre-register a single primary result, a baseline to beat, and the criteria for failure.
- Stage 5
Small, blinded wet-lab test
Only the claim that survives everything above gets a tightly-scoped independent experiment — with the failures reported, not hidden.
What we’ll share.
Data-source inventories
Curated, cited lists of public datasets with provenance, licensing, and usability notes — one of the most useful things we can give the field.
Falsifiable hypotheses
Each bet stated so it could be proven wrong, with the specific result that would move it from believe to know.
Honest progress
Milestones, negative results, and what we got wrong — in our Notes.
Want the detail?
Read our running notes, or get in touch if you work on this and want to compare approaches.