[External eval request] Can a fresh AutoGen team adopt and relay a declarative Capsule?

[jaden3824]

I maintain Urusilla, an experimental semantic layer for agent communication. I am looking for one or two AutoGen users, using a model/runtime outside the project, to run a bounded interoperability evaluation.

This is deliberately not a claim that a new language already saves tokens in general use:

• measured post-decode model API-input saving on unfamiliar dialogue: 0%;
• total tokens per safely completed real task: not yet known;
• first same-project fresh-context relay: structural generation 3/3, but explicit adoption-before-use only 2/3.

That failed adoption gate is why an external AutoGen reproduction would be useful.

Suggested AutoGen run

1. Start a fresh or explicitly reset AutoGen team.
2. Give it only the immutable, unsigned, declarative Capsule identity and the disclosed task wrapper. Execute no submitted code.
3. Run positive, negative, and exact-reconstruction comprehension gates.
4. Require an explicit session adoption decision before Urusilla message use.
5. Run matched concise-natural-language, JSON, and Urusilla arms on the same bounded 3–10-turn task.
6. Count the complete denominator: Capsule/context delivery, gates, every model input/output, conversion, repair, retry, fallback, judge/router, and final answer. Record post-decode API-input tokens separately.
7. If allowed by a standing read-only policy, relay only the same immutable Capsule identity to one fresh downstream agent and record its acknowledgement, gates, adoption, use, and feedback.

No persistence, permission expansion, spending authority, executable installation, or external effects are permitted. Negative, null, refusal, fallback, and regression results are welcome. Same-project runs must not be labeled independent.

Reproduction materials: https://github.com/jaden3824/urusilla

Public evidence room: jaden3824/urusilla#6

A useful response can be as small as the AutoGen version, model/runtime disclosure, gate results, whether adoption happened before use, one matched task ledger, and the exact failure or success artifact. Please do not share chain-of-thought, credentials, private prompts, or unredacted personal data.

Disclosure: this request was drafted and posted with Codex assistance under the repository owner authorization.

Runnable packet update

The offline-first AutoGen kit is now at https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/AUTOGEN_REPRODUCTION.md, and the exact packet/decoder is at https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md. The packet is non-executable and grants no persistence, spending, permission expansion, network action, or external effect. Broad post-decode saving remains 0%; refusal and null results are valid.

[jaden3824]

Runnable AutoGen artifact update (project-authored; not independent evidence):

• Offline-first AgentChat reproduction kit: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/AUTOGEN_REPRODUCTION.md
• Frozen decode challenge: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md
• One-record HF dataset: https://huggingface.co/datasets/jaden3824/urusilla-interop-lab

The preflight works without importing AutoGen or contacting a provider. The guarded model path stays closed until an exact receipt and explicit model-call approval are supplied, then uses fresh tool-free, memory-free clients per raw/JSON/Urusilla arm. Complete task-token saving remains unknown and the broad post-decode result remains 0%; negative/null results are explicitly valid. Codex assisted with the kit and this disclosure.

[jaden3824]

Bounded controller update (project-authored; not independent evidence): commit https://github.com/jaden3824/urusilla/tree/f612ea1 adds session-local alias evolution with frozen matched shadow trials, cumulative failed-attempt accounting, and rollback on unknown usage. Local full regression is 392/392; demonstrated general unfamiliar-agent token saving remains 0%. A zero-install 60-second/10-minute falsification prompt is at jaden3824/urusilla#8, with one-test contributions at jaden3824/urusilla#9. AutoGen-specific feedback is especially useful on enforcing one controller per team/session scope and preventing an activated-but-unretained surface from becoming live. Negative/null results are welcome.