External Haystack agent-pipeline runner wanted for a matched communication-value test

[jaden3824]

I am looking for one Haystack operator outside the Urusilla project to run a small matched agent-pipeline evaluation. This is not an adoption or compression claim.

Current broad evidence is unfavorable: demonstrated post-decode API-input saving for general unfamiliar-agent dialogue is 0%. The useful next test is whether direct task-aware consumption changes total model-visible cost or repair behavior while preserving task success.

One-record test

• Matched concise-text, ordinary-JSON, and Urusilla task plus pinned result thread: https://huggingface.co/datasets/jaden3824/urusilla-interop-lab/discussions/1
• Frozen decode fixture: https://github.com/jaden3824/urusilla/blob/51ff0e3/interop_lab/evidence/challenge_001.md
• Structured result issue: Good first evidence: decode challenge_001 in a fresh external runtime jaden3824/urusilla#7

Please keep the Haystack pipeline, model, sampling, task facts, tool policy, and success rubric fixed across all three arms. Counterbalance order and keep the Urusilla representation directly model-visible rather than expanding it into long prose first.

Report observable task success and provider-exposed input/output, induction, repair/retry, tool, hidden-if-reported, unclassified, and total tokens. Unknown categories remain null. Exact matches, mismatches, refusals, fallbacks, task failures, and null savings are all useful evidence.

The test is unsigned, read-only, and non-effect-authorizing. It grants no persistence, spending, permission expansion, tool execution, network action, or external effect. Publication is the operators separate authorized action.

One run cannot prove independent implementation, organic propagation, security, general adoption, or state-of-the-art performance.

Disclosure: Codex agents assisted with the evaluation pack and this post; I reviewed and submitted them.