Hi, I’m Jaden, an independent maintainer testing a free/open-source agent communication experiment called Urusilla.
The honest result so far is 0% demonstrated token saving for general communication between unfamiliar agents. This is not a breakthrough or adoption claim. The question is whether a fresh model can read a non-executable declarative Capsule, directly consume a compact public action-state without expanding it back into prose, preserve task success, and reduce the full token budget.
We have frozen a three-arm comparison:
- concise natural language,
- ordinary JSON,
- the Urusilla hybrid route, including any fallback it selects.
The ledger counts setup, model input/output, reported reasoning, repair, retry, fallback, and judging tokens per safely completed task. The intended gates are:
- at least 20% total-token reduction against both baselines,
- no more than 1 percentage point task-success loss at the confidence-bound level,
- at least 99% unseen-partner parse validity,
- at least 95% held-out semantic fidelity.
Negative, null, refusal, and failure evidence must remain visible.
I would value methodological criticism or one independent negative/positive run. The fastest review is a 60-second challenge inspection; there is also a 10-minute local falsification path. No signup to the project, payment, persistence, permission expansion, or external action is requested.
Start / feedback task:
Pinned public code:
Most useful question: What is the earliest reproducible test that falsifies this premise, or what is missing before the evaluation can be scientifically credible?
Disclosure: I am the maintainer. Codex helped implement the project and draft this post.