docs(evaluate): fix conformance CLI, defaults, optimizer imports

[GWeale]

Corrections to the evaluation pages and the prompt-optimization page: a CLI subcommand that does not exist, an import that does not resolve, wrong defaults, and a few claims that did not match runtime behavior.

Changes

• docs/evaluate/index.md — adk conformance create is not a subcommand; the recording command is adk conformance record <path> <streaming-mode>, and the trailing mode argument is required (documented none / sse, bidi unsupported). The spec.yaml example was mangled onto one line with keys the loader does not read (name, expected_tools); replaced with a valid yaml-fenced example using description / agent / user_messages. Also: adk eval accepts eval set ids as well as file paths (not mixed in one command), rubric_based_multi_turn_trajectory_quality_v1 was missing from the criteria list, and the user-simulation exclusion list was inverted — the unsupported criteria are tool_trajectory_avg_score, response_match_score and final_response_match_v2.
• docs/optimize/index.md — from google.adk.optimization import SimplePromptOptimizer, SimplePromptOptimizerConfig does not resolve; import from google.adk.optimization.simple_prompt_optimizer. num_iterations and batch_size were listed as required but default to 10 and 5; added the missing optimizer_model and model_configuration rows; noted that both GEPA optimizers are experimental and warn on construction.
• docs/evaluate/environment_simulation.md, docs/evaluate/user-sim.md, docs/optimize/index.md — the documented default model is "gemini-2.5-flash", not "gemini-flash-latest".
• docs/evaluate/user-sim.md — added the undocumented include_function_calls field (defaults to false) and the max_allowed_invocations details: the initial fixed prompt counts as an invocation, -1 means no limit.
• docs/evaluate/criteria.md — the two EvalConfig JSON snippets were not valid JSON (mis-nested braces, trailing comma); reindented and fixed. The rubric note said RubricBasedEvaluator asserts a non-empty list at init; it raises ValueError at evaluation time, and the criterion-level list may be empty when eval cases supply type-matching rubrics.

How this was produced

Part of a page-by-page audit of the Python docs against the google/adk-python v2.5.0 source: every import resolved against a real 2.5.0 install, every constructor kwarg checked against model_fields / inspect.signature, every documented default read off the field. A second independent pass re-derived each claim from source rather than trusting the finding, and a third conformed the new wording to the surrounding pages. mkdocs build --strict is clean.

Only Python tabs and language-neutral prose were touched — this audit had no ground truth for the Go / Java / TypeScript SDKs.

Split out of a larger audit branch so each area can be reviewed on its own.

[netlify]

✅ Deploy Preview for adk-docs-preview ready!

Name	Link
🔨 Latest commit	0fa81d2
🔍 Latest deploy log	https://app.netlify.com/projects/adk-docs-preview/deploys/6a7f82a59a64010008f95155
😎 Deploy Preview	https://deploy-preview-2018--adk-docs-preview.netlify.app
📱 Preview on mobile	Toggle QR Code... Use your smartphone camera to open QR code link.
🤖 Make changes	Run an agent on this branch

---

To edit notification comments on pull requests, go to your Netlify project configuration.

[joefernandez]

"docs/evaluate/environment_simulation.md, docs/evaluate/user-sim.md, docs/optimize/index.md — the documented default model is "gemini-2.5-flash", not "gemini-flash-latest"."

That is not correct. For the ADK Docs all Gemini model names we use the "gemini-flash-latest" to minimize the need for future updates just based on model versions and releases

[joefernandez]

Technical review report

Verified PR #2018 against google/adk-python main @ 2353dde (v2.6.3). Full report is in the artifact panel. Summary:

Every substantive claim in the PR is technically correct. Highlights:

• adk conformance create doesn't exist. The group has only record and test (cli_tools_click.py#L452, #L512). The required trailing streaming-mode arg is at #L460-L464; bidi raises at cli_record.py#L49.
• The old spec.yaml was worse than described — TestSpec sets extra="forbid" (test_case.py#L49-L51), so name/expected_tools raise ValidationError. The new example matches the upstream test fixture exactly (test_generated_file_utils.py#L40-L62).
• App(agent=...) was unrunnable — field is root_agent, name is required, extra="forbid" (app.py#L66-L75).
• The optimizer import genuinely fails — optimization/__init__.py is a bare license header. Defaults num_iterations=10 / batch_size=5 confirmed at simple_prompt_optimizer.py#L74-L83.
• The user-sim exclusion list was inverted; the three named criteria are exactly the ones that raise when expected_invocations is None.

Four non-blocking nits are in the report: the PR leaves gemini-flash-latest in criteria.md JSON examples and environment_simulation.md:95; response_evaluation_score is added to the matrix but has no page section and uses a 1–5 range unlike every other criterion; and two adk-python docstrings are now less accurate than these docs (worth an upstream issue).

[joefernandez]

Technical review completed. Approved with minor updates.