14 September 2026 · connectome replication
Eleven days after the complete male fruit fly connectome was released, we built it
into a 108,392-neuron network and pressed the panic button. Two neurons reach 19% of the brain;
scrambled wiring reaches nothing. A prediction staked in advance failed, was fixed by balancing
the network, and then held against five controls — including size-matched and random populations.
What the wiring could not do is reported too: selectivity or propagation, not both.
Prediction staked first
Failed once
Five controls
Limit measured
12 September 2026 · Lean formalization
11,424 build jobs, exit code 0, thirty-eight minutes. No unproven gaps in the
proof path and no added axioms. The theorem lets the construction choose the external force —
Clay alternatives (C) and (D) — which is narrower than the headlines and exactly what OpenAI's
own README says.
Build passes
No gaps
Standard axioms
Narrower than reported
13 September 2026 · primer
Not a finding — a primer. Every belief a predictor could reach about a four-state
hidden source, computed from published equations with no model calls, and rendered as an object
you can turn. It also carries two corrections: the dimension we published was a property of our
sample rather than the set, and the formula we checked it against was the wrong one for three
dimensions. Both failures are kept on the page.
Reproducible
Primer, not a discovery
Dimension unresolved