What narrow AI task should the harness test next?

[fusiontechstrategies]

AI Accuracy Harness v5.0.0 currently demonstrates a deliberately narrow claim: bounded workers completed 39/39 qualification calls, while an open-ended security-review adapter was rejected because its evaluation was not deterministic enough.

I want the next example to come from a real task that people need, but it still has to meet a strict bar:

• one clearly bounded input and output;
• a deterministic oracle that does not ask another model to judge correctness;
• synthetic fixtures we can publish;
• tempting wrong answers the oracle must reject;
• explicit abstention behavior; and
• proposal-only authority with no tool calls, publishing, deployment, or self-approval.

Good candidates might include exact configuration extraction, choosing from a preregistered option set, or a transformation checked by a project-owned parser and invariants.

If you have a candidate, please share:

1. the exact task;
2. the smallest useful input and output;
3. how you would grade it mechanically; and
4. two plausible wrong answers that must fail.

The new deterministic-oracle issue form is also available for a more structured proposal:
https://github.com/fusiontechstrategies/AI-Accuracy-Harness/issues/new/choose

Roadmap:
https://github.com/fusiontechstrategies/AI-Accuracy-Harness/blob/main/ROADMAP.md

I am especially interested in examples where a fluent, schema-valid answer can still be wrong in a way that matters.