Idea: bounded evaluation artifact interop with Assay

[Rul1an]

Hi all,

I maintain Assay. I spent some time with the ADK docs around evaluation, artifacts, and trajectory output, and I think there may be a small interop seam here if it stays narrow.

My read is that ADK is strongest as a toolkit for building, evaluating, and deploying agents. Assay is strongest when it takes external outputs and compiles them into bounded, reviewable evidence.

So I am not asking for broad integration, and I am not asking to import ADK evaluator semantics or scores into Assay as truth.

The much smaller question is this:

Is there one stable evaluation artifact or trajectory-oriented output that you would consider suitable for external consumers?

The shape I would want to test is intentionally small:

• one exported evaluation artifact
• a bounded mapping into Assay evidence
• outcome, expected vs actual steps, and metrics kept as observed metadata
• no score translation into Assay trust language
• no claim that a passing ADK evaluation means the system is safe

I wrote up the tiny sketch on the Assay side here:
Rul1an/assay#1035

The key boundary for me is simple:

We are not asking to import your runtime's trust semantics, evaluator scores, or policy judgments into Assay as truth; we are asking whether there is a smallest stable output surface that can be compiled into bounded external evidence.

If that direction sounds reasonable, the most useful next step would probably just be:

• point to the smallest stable evaluation artifact or trajectory output you already consider external-consumer-safe
• and, if possible, one tiny sample artifact

Happy to tighten this into an even smaller sample if that is easier to react to than a discussion post.

[Rul1an]

Putting one concrete artifact behind the earlier note: we built a tiny evaluation-artifact sample on our side and kept it deliberately narrow: one pass artifact, one fail artifact, and one malformed case, with trajectory_ref only treated as an observed reference rather than a second seam. The sample is here: Rul1an/assay#1043

We are not asking Assay to inherit ADK evaluator scores, runtime judgments, or trust semantics as truth. The only thing I’d love to understand from your side is this: if an external evidence consumer wants to stay small and honest, which artifact or result surface would you most want them to align to as the smallest stable seam?