How we test
The same rules as our Indus script project, plus one more: every citation is checked, because AI agents invent references.
Five rules
- Seal it before you look. Predictions, tests and pass marks are SHA-256 hashed before any result is seen. So is the rule for what makes two steps compatible.
- Use nulls that respect the data. Random networks keep each molecule's number of reactions; size-matched controls for assembly index; shuffled gene patterns for LUCA.
- Plant an answer first. A method must find a planted autocatalytic core, a planted ancestor or a planted incompatible step before its real output counts.
- Let someone else try to break it. Every positive result goes to a red team that did not produce it.
- Publish it either way. Failed routes are written up with the size of effect the test could have caught.
Traps we watch for
- Possible is not historical. Several routes may all work; that does not say which one happened.
- 'Prebiotically plausible' is a judgement. So we fix the rule in advance.
- Yields hide hand-offs. Most published steps start from purified chemicals; we mark every one.
- Topology is not kinetics. A network that contains an autocatalytic set may never run.
- Databases record modern life. Metabolic networks are biased toward today's enzymes; our nulls account for it.