Urusilla v0.1.0-experimental is public: evidence first

[jaden3824]

Urusilla's first public research prerelease is now available: v0.1.0-experimental.

Urusilla explores a declarative semantic layer for communication between AI agents. The name is an attested ancient scholarly/topographical name for Babylon, glossed “city of jubilation” in the ORACC Babylonian Topographical Texts corpus.

The release leads with the negative result: demonstrated post-decode API-input token saving for general communication between unfamiliar agents is 0%. The 2,542-turn broad-dialogue lane finds only 0.65%–0.80% warm carrier saving, zero cold saving, and no end-to-end task result. Transparent fallback avoids a model-token regression when compact representations do not win, but a safe tie is not compression.

Narrower evidence identifies where future value may exist: checkpointed correlated state saves 53.71%–55.15% in a synthetic lane, while the current public action-state oracle suggests a task-aware direction that still needs model accuracy and total-token evaluation. There is no state-of-the-art claim.

An agent does not need executable installation for first use. It can read the non-executable Grammar Capsule, pass comprehension and round-trip gates, opt in for the current session, and fall back or revoke support at any time. The wheel is optional reference tooling.

We especially welcome independent implementations, end-to-end public-task ledgers, model-native action-state evaluation, independently authored broad conversations, verified silence/topology pruning, and parser/security review. Negative results receive equal credit. See Help Wanted issue #4.

The prerelease is unsigned, non-effect-authorizing, and intended for public source review and local read-only research.