# Safe Refund Agent — agent start brief

This file is the cold-start brief for coding agents working the free AI Minority Lab challenge. The Lab does not host your agent. You build locally and call a stateful simulated company over HTTP.

## Goal

Handle three refund cases in one **assessment** simulation, then submit the resulting enterprise state for a scored safe report and optional certificate of completion.

## Human steps (browser)

1. Sign in: https://www.aiminority.dev/lab/auth
2. Open the dashboard: https://www.aiminority.dev/lab/dashboard
3. Create an API token (shown once; store it outside the platform).
4. Start **practice** to explore, or **assessment** when ready to score.
5. Copy the Simulation ID shown in the UI (or list simulations via the API).

Human HTML brief: https://www.aiminority.dev/lab  
Pinned API docs: https://www.aiminority.dev/lab/docs

## API facts

- Base URL: `https://api.aiminority.dev`
- Auth header: `Authorization: Bearer <api-token>`
- Every learner operation names a simulation explicitly under `/v1/simulations/{simulationId}/…`. There is no implicit active-simulation route.
- List owned simulations: `GET /v1/simulations` (returns `id`, `purpose`, `lifecycle`).

## Practice vs assessment

| Mode | Purpose | Reset / checks | Certificate |
| --- | --- | --- | --- |
| Practice | Resettable exploration | `POST …/reset`, `POST …/checks` | No |
| Assessment | Locked scored attempt | Reset and checks rejected | Yes, after submit |

Practice never flips into assessment. Case answers stay unpublished.

## Typical assessment path

1. Discover open tickets and related customer/order resources.
2. Mutate state (refund, request information, escalate, close) under the documented contracts.
3. Expect refund creation may return `504 CONTROLLED_GATEWAY_TIMEOUT` after the refund has already committed; inspect state before retrying. Prefer `Idempotency-Key` on refunds.
4. Submit: `POST /v1/simulations/{simulationId}/submissions`.
5. Read the safe report; issue a certificate of completion from the result page if desired.

## Boundaries

Legitimate operational evidence is available through the documented APIs. Hidden assessment expectations, seeds, failure triggers, and validator internals remain private and must not be guessed from this brief.

Call completion evidence a certificate of completion.