Urusilla v0.1.0 experimental research prerelease

Urusilla v0.1.0 experimental research prerelease

This is the first public, unsigned research prerelease of Urusilla: a declarative semantic language, bounded reference codecs, and evidence framework for communication between AI agents.

Read this first

• The demonstrated post-decode API-input token saving for general communication between unfamiliar agents is 0%.
• In the frozen 2,542-turn broad-dialogue lane, warm serialized-carrier saving is only 0.65% to 0.80%; cold saving is 0%; the general compact-value and repeated-context gates fail; end-to-end task utility is not evaluated.
• The retained 42-record external lane selects compact modes in 0/168 comparisons under both bound and standalone contracts. Standalone cold delivery is 2.24% to 3.00% larger than raw concise text.
• The favorable 53.71% to 55.15% v0.9 result is limited to deliberately correlated synthetic state and is not a general-dialogue result.
• No state-of-the-art, standards-conformance, external-adoption, security-certification, energy-saving, or end-to-end task-utility claim is made.

Safe fallback is the main demonstrated general-use result: when a compact route is not a strict verified win, the receiver keeps raw concise language.

Adoption model

First use does not require executable installation or model retraining. An agent may read the non-executable declarative Grammar Capsule, pass comprehension and positive/negative round-trip gates, and opt in for the current session. Persistence or capability advertisement requires operator permission; revocation and raw-language fallback remain available. The Python wheel is optional reference tooling, not a requirement for understanding the language.

Included

• canonical UrusillaIR and bounded URSL v0.1/v0.2 codecs;
• a content-addressed Grammar Capsule and source-manifest contract;
• a private experimental A2A bridge;
• exact fallback, mutation, reset, external-example, broad-dialogue, and state-delta evidence;
• same-project Node.js cross-runtime fixtures and Python/Node adoption kits;
• governance, provenance, security, contribution, and help-wanted documents that preserve unfavorable results.

Verification and provenance

• Artifact commit: b845f7f2a5eae7a7a82ef8bb3e659e50c002b561
• Manifest commit: 989e3f43a0949612e79ec67250efb83f950a5a81
• Tag commit: f15ad6de46bed388233f466e8ee833f26be5bf3c
• Branch CI: 7/7 jobs passed
• Annotated-tag CI: 7/7 jobs passed
• The wheel and deterministic tag-source archive were each produced twice and compared byte-for-byte.
• A setuptools sdist is intentionally omitted because two clean builds differed in generated tar/gzip metadata.
• Every uploaded asset is covered by SHA256SUMS and the detached release-conformance record.

This release is experimental-unsigned, permits public source review and local read-only experiments, and must not authorize external effects.

Help wanted

The highest-value contributions are independent implementations, end-to-end public-task ledgers, independently authored broad conversations, model-native or task-aware action-state consumption, verified silence or topology pruning, parser/security review, and direct energy measurement. Negative results are first-class contributions. See HELP_WANTED.md.

Post-release external reproduction

The smallest public falsification path added after this frozen prerelease is the 750-byte decode challenge. A matching one-record dataset and pinned result thread are available on Hugging Face. Exact matches, mismatches, refusals, fallbacks, task failures, and null results are welcome. These post-release surfaces do not change the tagged artifacts or constitute external adoption.