Qualify ACE-Step

ACE-Step is a managed media profile, not an open generic API connection. The profile pins the runtime source, checkpoint tree, model identity, recipe, and hardware policy before the worker may advertise audio generation.

Install the Reviewed Runtime

Do not attach an arbitrary ACE-Step checkout to the public Grid. The media manager installs and verifies the exact source revision, dependency lock, checkpoint tree, model identity, and recipe bound by its reviewed profile. Use only the qualification artifact selected by the verified release gate on aipowergrid.io/run.

Detect the Endpoint

The managed profile supervises its own loopback ACE-Step API and validates the reported model inventory and health. An unrelated reachable audio endpoint is not sufficient. The manager must launch the bound runtime and keep its health check passing.

Compatible Models and Recipes

Compatibility is profile-defined, not inferred from a model directory. The current qualification path evaluates only the pinned ACE-Step audio model and recipe declared by the candidate profile. Other checkpoints, quantizations, or audio workflows require their own reviewed profile and evidence.

Qualification Flow

1. Open the current media qualification status on aipowergrid.io/run.
2. Download a qualification artifact only when that page verifies its manifest, checksums, SBOM, and immutable release identity.
3. Run the local hardware inspection, installation verification, and audio canary.
4. Submit only the privacy-reduced public report through the documented cohort path. Keep the private hardware report and all credentials out of GitHub.
5. Wait for a reviewed signed profile and supervised staging before treating the machine as a Grid worker.

After a signed serving release is enrolled and reports online, the manager also offers Run Grid test. That separate test asks Core to hard-target a small governed audio render to the exact rig and verify its upload and receipt. It does not replace qualification and creates no charge, den, payout, validator evidence, or quality claim. The unsigned qualification artifact cannot run it because it cannot enroll.

Operator Controls

The manager exposes always available, paused, or one bounded maintenance-window schedule. Media capacity remains one job at a time. A pause drains an active generation before the Grid connection closes. The manager does not infer spare capacity from unrelated local activity automatically.

Security Boundary

The managed runtime remains offline from model hubs while serving and must keep its exact source, dependencies, model files, recipe, and local health checks valid. The worker uses a funds-less signer and a separate payout-wallet delegation; it never asks for a payout-wallet private key.

Prompts, lyrics, and generated audio are plaintext to the community worker. Runtime logs stay at warning level so those inputs are not deliberately written to service journals. Output uploads use Grid-issued short-lived URLs rather than standing storage credentials.

Verify Qualification

Check the exact candidate version, checksums, manifest restrictions, and public evidence counts on /run. A successful local benchmark proves only that the candidate ran on that machine; it does not create a worker registration. Public verification in the worker registry becomes relevant only after a signed serving profile and supervised release exist.

Troubleshooting

• Manager refuses enrollment: that is expected for the unsigned public qualification artifact.
• Hardware check fails: compare the sanitized result with the published minimum, midrange, and datacenter classes; do not bypass the profile.
• Install or canary fails: rerun the manager’s local verification and keep the machine off the Grid. Share only the requested sanitized report, never credentials or full private hardware evidence.
• Runtime later loses health: the supervisor withdraws the capability and restarts the complete runtime/worker pair rather than leaving a dead route advertised.

Use the full media-worker guide for current evidence classes, availability controls, and troubleshooting.