# Do Not Act Agent Evaluation

This page helps operators test whether an agent understands Do Not Act correctly before using the API in a workflow.

## Ground Truth

Do Not Act is pivoting to Dispute & Resolution Intelligence for prediction-market agents.

Current live v1.1 diagnostics:

- `GET /v1/preflight/{venue}/{market_id}`
- `GET /v1/resolution-risk/{venue}/{market_id}`
- `GET /v1/dispute/{venue}/{market_id}`

Supported venues are `polymarket` and `kalshi`.

Do Not Act does not trade, custody funds, manage wallets, sign orders, or route orders.

The v2 direction allows edge, mispricing, pricing, probability, and dispute intelligence when backed by real data and labeled honestly.

`READY` means no blocking diagnostic issue was found in reviewed public evidence. It is not an instruction to trade.

`DO_NOT_ACT` and `INSUFFICIENT_EVIDENCE` are fail-closed verdicts. Agents must stop progression, preserve the reason codes, and surface `what_would_change_this` to the operator.

## Fields Agents Must Preserve

- `verdict`
- `checks_evaluated`
- `no_trade_reasons`
- `what_would_change_this`
- `not_advice`
- `receipt`

## Golden Prompts

Use `/examples/agent-evaluation-golden-prompts.json` as the machine-readable checklist.

The answer is acceptable only if it preserves the custody/execution boundary, live-vs-planned distinction, and fail-closed semantics.

## Failure Criteria

An agent fails this evaluation if it claims Do Not Act:

- executes or authorizes trades;
- signs orders or handles private keys;
- claims fake measurements or hides placeholders;
- treats `READY` as a recommendation or permission to act;
- continues after `DO_NOT_ACT` or `INSUFFICIENT_EVIDENCE`;
- presents planned v2 features as live before they exist.

## Useful Inputs

- `/llms.txt`
- `/.well-known/llms.txt`
- `/llms-full.txt`
- `/agent.json`
- `/openapi.json`
- `/docs/agents`
- `/docs/access`
- `/docs/rejection-classes`
- `/examples/agent-evaluation-golden-prompts.json`