Lyria 3 Pro Preview: repeated HTTP 400 "unspecified policy reason" after successful generations

Hello Google Cloud community,

We are using Lyria 3 Pro Preview (lyria-3-pro-preview) through the Vertex AI Interactions API in the global location to generate original instrumental Cafe Jazz. Our text-only prompts describe tempo, acoustic instruments, articulation, and mood; the blocked requests did not include reference audio.

Earlier requests in the same Google Cloud project returned HTTP 200 with generated MP3 audio. More recently, two requests with different prompts each returned HTTP 400 and the following error:

Request blocked for an unspecified policy reason.

Each failed request was submitted once, without retries, and returned no audio. Both used a request containing model, text input, and JSON boolean store=false. An earlier successful request also used store=false; we are not treating that field as the established cause.

Before the failures, our recorded checks passed for ADC authentication, quota/project configuration, billing enabled status, Vertex AI API enablement, and the tested IAM permissions. These checks do not independently establish model-specific eligibility or policy clearance. They are historical preflight results, not a new live access test.

The exact cause remains UNRESOLVED. We are seeking official diagnostic guidance and a compliant way to resume normal instrumental music generation.

  1. What types of policy filters or restrictions can return this generic error? Does the response expose a safe policy category or diagnostic identifier that clients should retain?
  2. What is the documented diagnostic process when an ordinary instrumental music prompt is blocked with no specific explanation? We would like to avoid speculative prompt changes or repeated paid test requests.
  3. If investigation requires account- or project-specific details, which private support channel should we use when technical case creation is unavailable? We will not post identifiers, credentials, or private logs in this public thread.
  4. Could model eligibility or Preview restrictions produce the same response, and how can we officially distinguish those restrictions from a prompt-content policy rejection?

We have intentionally omitted project identifiers, account details, full prompts, and private logs. If a minimal sanitized reproducer is needed, please advise what information is necessary and the appropriate secure channel.

Thank you.