Notes, working papers, experiment reports, and validation write-ups from the Answer North research program. Every public numerical claim will be tied to a reproducible report before it ships — which is why the first entries below are staged, not yet published.
Building a quantitative research system designed to preserve the evidence that proves it wrong: preregistered experiments, frozen evaluators, evidence custody, fail-closed authority — and the cycle where the system withdrew its own recommendation to protect a sealed measurement.
PUBLISHED · AUGUST 2026 · ~14 MINToken counts are a cost metric, not a value metric. This paper proposes optimizing the ratio of task-relevant reasoning to total compute spent, and describes the compress–route–verify architecture TLTLL uses to pursue it without silent quality loss.
FORTHCOMING — IN PREPARATIONFluency is not accountability. The case for evidence custody, deterministic execution, verification layers, and governance gates as the missing infrastructure between capable models and consequential decisions.
FORTHCOMING — IN PREPARATION