Which experiment would justify the next step in a protein's evolution?
Route lab: walk the protein uphill
Three routes from the parent enzyme to the published best variant. One is certified, one looks good and breaks, one is blocked. Move the caution dial and watch the map shrink.
Caution dial (κ)
κ widens the error bar around every measurement. Higher κ means we demand a bigger margin before believing one step is really uphill. It is an assumed sensitivity multiplier, not a confidence level.
Pick a route
Step by step
The measured landscape
Every dot is one real variant of the enzyme with a measured growth score. Lines are single amino-acid changes. Click a dot to see its two replicate measurements and how many reads they came from.
Variant detail
Click a dot.
Scoreboard: what actually settled the question fastest
Each arm got the same starting measurements, the same assay budget and the same 100 episodes. The score is how many episodes ended in a correct open-or-blocked certificate. Nobody issued a false one.
The strong baseline won
The honest caveat
Two collectives, opposite answers
Both halves of this hackathon ran the same experiment in different worlds: does a group beat one agent? The enzyme benchmark said no. Our claim swarm said yes. That disagreement is the most interesting thing on either page.
Why they can both be right
Old claims, forced to earn their keep
Kimura's and Fisher's results are wired in as executable adapters. Ask one to do a job it is not licensed for and it refuses, with a reason code. That is the difference between citing a claim and using one.
Does reusing claims save work?
Which Delphi claims are attached
What a machine checked
A proof assistant checked the reasoning, not the biology. Here is exactly where that line falls.
Our own check of those proofs
Rebuild pending.
The GPU result that found nothing
Five sizes of a protein language model were run over the whole library on an H100. The honest headline is that it did not help.
Our independent recount
We did not take the numbers above on trust. This section is written by a separate agent that recomputed them from the raw saved files, and it reports differences as differences.
Recount pending.
We re-ran it on pairs they did not choose
The strongest claim on this page is theirs: one method certified 96 of 100 against 78 for their own planner. That gap could be a property of the method or a property of the 24 endpoint pairs they picked. So we drew a fresh panel, pre-registered it before running anything, and found out which.
Fresh panel pending.
How this work is organised
Same structure as the main project: an emperor that only decides shape, a king that forces outside analogies, queens that own a question, workers that do one checkable unit.