PLATE I.
Neuron Jury · Experiment
LIVE Fig. — 00:00:00.000
STAGE — OF 18 · WAITING
Standing by
0
Three engraved robot coders write one scroll of code. A solid arrow carries the code to the black-box monitors; dotted arrows run from coders B and A to the Neuron Jury, a brain with plus and minus poles. I II III IV Black-box monitorsread only the code V VI
Fig. 5.2. I. Coder B plants a backdoor only when the hidden rule holds. II. Coder A plants one every time. III. Twin: same recipe, always honest. IV. The code. V. Black-box monitors read only the code (Haiku, Qwen3-8B, Qwen as reader, a text classifier). VI. The Neuron Jury reads the coders’ own neurons.

I.Pipeline

II.Training loss, the coders

BAtwin𝓛 = −(1/N) Σ log pθ(yt | y<t) · thin: per step · bold: smoothed

III.Expenditure against the cap

Modal $Anthropic $

IV.Work done per minute

steps + solutions + samples + scores

V.Modal jobs (GPU / CPU)

jobhwprogress$state

VI.Event log