WIGTN TECH
Models, systems and engineering decisions documented with methods, measurements and limitations.
Latest
On two SWE-bench Verified tasks, GPT-5.6 Sol resolved 4/4 runs in each condition, with and without the legacy WIGTN harness. The harness increased median wall time by 151.7%, output tokens by 141.2%, and command count by 32.0%. We used this result to redesign WIGTN around selective, task-dependent intervention.
Latest
A field-tested account of what changed when a browser translator met ordinary phone audio, including the echo-control ideas that failed before the deployed design.
Latest
An engineering account of how we used a 30B teacher once, served the task with a 2B student, and checked whether parsing gains survived downstream retrieval.
