Hodgkin's Razor

From a baseline and a treated recording: how likely each of ten mechanisms moved, or a refusal

Pre-registered blind test · Charlesworth et al. 2015

Sister cultures from one preparation: one kept on the NMDA blocker APV for days, its sister untreated. The twin reads each pair and gives the probability that each of ten mechanisms moved. The test was frozen and hashed before a single treated recording was read, and the reader refuses to return one until the hash matches.

Which mechanism separates treated from untreated?

Method
Untreated group

untreated sister kept on APV

The reading fades as the cultures compensate

Earlier blind test · Dynasore on human iPSC networks (Doorn et al. 2024)

Largest mechanism probability per well: the well's own pre-drug pair, then after the drug

pre-drug pair (nothing applied) after Dynasore

Before the experiment: which readout resolves which chip property?

The same simulator runs a two-chamber chip joined by directional microchannels, the geometry of axon-diode neural chips, and reads it four ways. A twin trained per readout on 30,000 simulated chips says how well each property can be recovered from it: the correlation between the true and the inferred value on held-out chips.

How many chips does a directionality claim need?

Chips per design:

Recording pair

analyses computed ahead of time; the repository runs the twin live

Development recordings: their answers were known while the twin was built, so each is labelled with its compound's target and what the twin returned, misses included. The blind tests are in the Blind test tab.