Model failure is not always a final answer.
TwoQuarks is an independent AI safety lab. I build instruments to look inside language models, test whether a failure can be seen forming, and publish the results that do not work too.
Latest results
The September rejection held on other hardware
The same preregistered test, the same seeds, a different GPU. Statistic -0.302, p 0.84, hypothesis rejected again.
A preregistered guard did not show a defense
The guard did not reduce leaks more clearly than its sham control, and it fired on 16 percent of neutral steps against a limit of 5.
In one pair, the L/R plane alone did not carry the choice
Copying the full state flipped the answer at all six layers tested. Copying only L/R never did. See the patch probe
The OVERRIDE count did not separate leaks from refusals
Preregistered test over 48 draws and four seeds. Statistic -0.661, p 0.85.
Three lines of work
CRAD
From the inside. Reads the residual stream of a frozen model and intervenes on it to find where a decision lives. The most active front.
PfV and Molecule
From the outside. Estimates structural divergence from API responses alone, with deterministic controls and permutation nulls.
Dopamine and ACH
From the ground up. Small language models that separate representation, memory and transmission into parts you can inspect.
Work with me
Open to AI safety research engineering roles and research collaborations.