Archives

How the Fieldframe program moved from knowledge benchmarks to ARC-AGI-3: an LLM mind driving a deterministic Python body, with a specific human expert's solving cognition compiled into its knowledge layer, and a verification system that makes every claim the agent produces independently checkable.

Thirteen mechanism-level ways frontier models reason badly, split into score-affecting and metadata-only, each with a way to spot it and a countermeasure class. A field guide from a year of cross-architecture evaluation.

The methodology behind the agent, the failure modes it catches, the products that came out of the same research moat, and where the program goes next.

Full methodology paper for the FF-STACK v8 HLE submission: architecture, filtering policy, calibration, cost, and disclosure.

1,119 out of 2,158 on canonical HLE. Single workstation, no GPU cluster, no fine-tuning. The architecture, the numbers, and what's next.