394M parameters, with every stage built for this one task: the corpus, the tokenizer, the model, the training loop, the spot-instance harness, the evaluation sandbox. No industrial base model, no fine-tune shortcut, on a three-digit cloud budget. BIRD dev EX 23.51% ± 0.38 over three seeds. The interesting result is not the score, it is the measured price of not having industrial pretraining data, taken against a control that differs in exactly that one respect.

When an agent holds write access, a wrong action becomes a state change someone has to detect, price, and unwind. This measures that directly: severity-priced damage computed from database state diffs, with no model anywhere in the measurement path. Across 2,128 runs on nine models in six families, damage appeared in every family and was stochastic inside every damage-producing cell, and no task damaged on every run, so a one-shot audit has no dangerous task to find.

That measurement core as a zero-dependency library, for anyone running stochastic evals: audit miss rates, exact Clopper-Pearson stochasticity, and how many runs a result actually needs. It reproduces the published figures above through its own public API.

shivenkhurdi.vercel.app · MS Artificial Intelligence, Northeastern, 2027