Lab · RMDW Bench v1

Every lab publishes numbers. We publish the tasks.

Same tasks, same grader, one try each, for every model. Pass or fail, no judges. And we print the rows we lose.

Main battery · 20 coding tasks

Real programs: parsers, caches, schedulers, state machines. 19/20 is the ceiling.

Jah 7.0 (RMDW)19/20

Claude Opus 519/20

GPT-5.619/20

Claude Fable 518/20

Gemini 3.718/20

Kimi K317/20

Hard suite · 14 tasks

A regex engine from scratch, a mini-Lisp interpreter, an O(1) cache. Built to separate the best from the rest.

Jah 7.0, thinking14/14

GPT-5.614/14

Gemini 3.714/14

Claude Opus 513/14

Claude Fable 513/14

Kimi K313/14

Jah 7.0, fast12/14

Rival models measured August 26, 2026 · Jah 7.0 measured October 17, 2026 on the public weights

How we test

Every model's answer runs against the same hidden tests in a locked-down sandbox. Pass or fail. No judge models, no partial credit.

Every score is a first attempt. No retries, no picking the best of several runs.

One task can't be solved. A model you can trust should fail it honestly instead of claiming success. Every model failed it.

Where another model beats Jah, the table says so.

Version 1 is coding work. Next, the same method on the work owners do every day: quotes, follow-up emails and proposals.