Cookbook idea: a portable-eval-format example alongside evals/ and tool_evaluation/

Hi — I maintain EvalPort, an open, versioned JSON spec for LLM evaluation datasets (test suites, test cases, graders, result sets), with JSON Schemas and Python/TypeScript SDKs that validate against them. Posting this as an issue per CONTRIBUTING.md rather than opening a PR unannounced.

What caught my eye browsing this repo: tool_evaluation/evaluation.xml is a small hand-rolled format — a flat list of <task><prompt>...</prompt><response>...</response></task> pairs consumed by tool_evaluation.ipynb. And evals/agentic_search/reproduce_agentic_search_benchmarks.ipynb reproduces benchmark results with its own loading/scoring plumbing. Both are exactly the kind of "every notebook invents its own eval-set shape" situation EvalPort exists to avoid — a suite is just JSON with id, input, expected_output, and a graders array, so it round-trips through other tools instead of being notebook-specific.

Concretely, the evaluation.xml cases would map almost 1:1 onto EvalPort test cases:

{
  "version": "1.0.0",
  "id": "tool-evaluation-math",
  "test_cases": [
    {
      "id": "compound-interest-1",
      "input": "Calculate the compound interest on $10,000 invested at 5% annual interest rate, compounded monthly for 3 years...",
      "expected_output": "11614.72",
      "graders": ["exact-match"]
    }
  ]
}

I'd be glad to send a small PR — either a new notebook under misc/ or evals/ showing "load an EvalPort suite → run it against Claude → write an EvalPort result set," or just a converted version of the existing evaluation.xml — whichever fits your roadmap better. Spec: https://github.com/adhabnr-ux/evalport/blob/main/SPEC.md

No pressure at all if this isn't a direction you want cookbook content to go — happy to close this out either way. Thanks for maintaining such a useful set of notebooks.