> Canonical guide: https://developers.foxlight.ai/build/guides/qualification/
> Contract snapshots: Skulk 2.0.0 (b0af39c79b6b7102b2478062904f1d7cc8619975); SDK 0.4.0 (31bb090b8f64689f87514e48385e22b6ad94a6c2). Check the installed runtime when versions differ.

# Qualify your capability

Verify your application at three boundaries before calling it supported. Unit tests alone do not prove a signed installation, routing or restart behavior.

## Application behavior

Test schema-valid input, invalid and oversized input, output validation, known backend failures and unexpected failures. For streams, test all used I/O modes, half-close, slow consumers, malformed frames, terminal ordering and cleanup. Keep model quality checks separate from protocol correctness.

## Managed process

Exercise inherited startup, matching identity and manifest digest, failed authentication, health while work runs, one active invocation, bounded deadlines and supervision after a crash. A child must not log its startup token or credential configuration. Do not infer readiness from a live PID.

## Installed Fabric integration

Install through the supported signed path. Confirm permission review, setup failure, configuration edits, activation, peer discovery, descriptor negotiation and unavailable-provider errors. Test API access from another node, then cancellation and restart with actual resources. If your application declares a UI surface, verify browser reachability and authorization independently from capability calls.

## Recovery and release record

For durable work, test a lost submit response and a restart after a backend reference is saved. Verify neither causes duplicate effects. Exercise artifact download failure and retention. Record exact SDK, host, Skulk and package versions with the platform evidence used to support the release.

A release must state qualified platforms explicitly. The package existing on a CPU architecture does not prove that every model or accelerator works on that architecture. [Product documentation](https://docs.foxlight.ai/) owns model installation and engine compatibility.