Reports and write-ups
Writing
Reports and write-ups, hosted where the code lives.
-
IncidentGate: a verification receipt for every published number
A dated log of what each measurement actually establishes, including the first time a monitor was shown a covert action and what its rationale reveals about how it decided.
-
Digital Grimace or Decoder Artifact? A preregistered stress-test of nonverbal generation-instability markers as candidate welfare-relevant signals
The preregistered primary test failed and is published as a FAIL; a re-preregistered iteration found an answer-margin channel that survives being trained out of the model’s language.
-
ProofPack: known gaps, and what to do next
An engineering audit of my own prototype, priority-ranked, written so a team inheriting the repository knows exactly where the edges are.
-
Whose Voice Is This Corpus Written In? Blind principal attribution from covertly poisoned data
Hidden principals can be recovered from poisoned training data at 12–44% top-1 of 47 candidates in a dense regime, and not at the poison densities real attacks use.