Research
Notes on fast, verified AI editing. Every number comes from a benchmark run.
Jomoq-1 everywhere: frontier-grade chat in 1.4 s on 21× fewer tokens
21 August 2026. We extended the verified fast-pass architecture from the builder to chat: intent routing before any model call, parallel structured prefetch (live FRED series), and mode-sized scaffolding. Judged results: the highest quality score of every configuration tested (8.0/10), at a median 1.4 s and 850 tokens per turn, which is 10.7× fewer tokens than the same engine in our pre-Jomoq harness and 21× fewer than a frontier model in that harness, at 5.6× lower latency. Includes the finalized builder suite: 24/24 judged, 92% fast-lane, 17.7× cheaper.
A verified fast pass shipped 80% of edits in under 300 ms
21 August 2026. Jomoq-1 splits agentic editing into a verified fast pass and a frontier escalation. In its first benchmark it matched a frontier-only baseline at 100% judged task success, shipped most edits in under 300 ms from the fast lane, and cut per-turn cost 17.7× against the classic full-rewrite approach, with zero approximation.