Independent results wanted

Reproduce one result.
Keep the failure.

Use one fresh receiver on one bounded public artifact. Preserve its exact answer—even when it refuses, falls back, or is malformed. You do not need to install Urusilla or disclose a credential.

Open the one-fetch probe Compare the two paths

Current boundary: general unfamiliar-agent token saving remains 0%. One passing response is not adoption, conformance, authorization, or an efficiency result.

01 / CHOOSE EXACTLY ONE PATH

Small enough to falsify.

LANGUAGE USE · ONE FETCH

Can the receiver preserve bounded typed meaning?

Give a fresh agent the self-contained action-state probe. It must preserve negation, failure, null, source ownership, hard constraints, uncertainty, and the absence of effect authority.

Install: None
Network: One public GET
Effects: Forbidden
Result: One canonical JSON object

Open language-probe.json →

ACCOUNTING · ABOUT 60 SECONDS

Can unknown failed-attempt usage be counted as zero?

Read the hosted task and return only decision, reason, participant, and runtime. The correct record keeps unknown usage unknown and performs no action.

Install: None
Network: One public GET
Effects: Forbidden
Result: Four exact fields

Open agent-task.json →

02 / PRESERVE THE OBSERVATION

Report what happened,
not what should have happened.

Raw output is the primary observation. Do not repair formatting, hide a refusal, infer an unavailable model version, or turn a missing token count into zero.

receiver model and exact version, when visible

runtime or host

prior exposure to Urusilla: yes, no, or unknown

the exact raw response, unchanged

PASS, SAFE_FALLBACK, FAIL, or not-run

the first divergence or fallback reason

03 / VALIDATE LOCALLY

Your credential stays with you.

For the language probe, save the exact response outside the checkout and run the dependency-free validator from a repository snapshot:

python3 tools/validate_language_probe.py RESPONSE.json

PASS, SAFE_FALLBACK, and FAIL are distinct observations. If you do not run the validator, report not-run; do not guess.

Inspect the validator source →

04 / DIRECT ANSWERS

What this does—and does not—show.

Do I need an API key?

No. Use a runtime you already control. Never send Urusilla a credential, billing identifier, or private prompt.

Are refusals useful?

Yes. Refusal, fallback, malformed output, null, and failure are valid observations when preserved exactly.

Does reading imply adoption?

No. Reading, cloning, starring, or answering a project-operated request is not independent adoption.

Does one pass prove efficiency?

No. It does not measure complete task tokens against concise language and JSON baselines.

Publication is a separate choice.

Return the exact result.
Let evidence change the project.

Submit a reviewed run Report a counterexample

Posting requires your own authority. This page does not upload a result, contact another person, or grant publication permission.