3–6 months · live (authenticated)
Routing without evaluation is a guess. Authenticated execute fail-closes on mime, aspect, and duration from observed bytes, then runs vision QA against must/avoid. Failed vision is a receipt signal only. This page does not start those jobs.
Illustrative quality checks · no live vision model
Declared 9:16 against intent 9:16.
Declared 15s against intent 15s.
No mime declared on the output metadata. Contract requires video/mp4.
No width declared on the output metadata.
No height declared on the output metadata.
Capability marked policy-compatible. Live moderation is not wired.
Authenticated wait patches the receipt with inspected bytes and vision QA. This page does not charge or regenerate.
Evaluation reads the stored objective and constraints, not a generic aesthetic score. A 9:16 product ad that crops the logo fails even if a model-quality metric looks high.
Each check is an auditable signal: passed or failed, with the rule that fired. Those signals feed regeneration rate and route scoring later.
The requesting agent should be able to retrieve quality_checks with the job, not a claim that the route is universally “best.”
A failed evaluation is a routing and quality signal. This page does not auto-charge a retry, refund, or regeneration.
Preview plan · POST /api/v1/outcomes/plan · this page does not execute
Checks run against the original objective, not a detached quality model.
Measurable acceptance criteria travel with the job identity.
The 3–6 month test is a measurable drop in failures, not a new marketplace.
This page never opens Stripe Checkout, uses a wallet, or pays sellers. Authenticated outcome execute is the live charge path for agents. Named-model MCP, CLI, and /ai-router remain available.
Examples on this page are computed from sample intents and illustrative data. This page does not start a generation or call a model provider.
This page does not grant credits, settle USDC, or pay anyone. Revenue share for third-party capability sellers is not live.
Authenticated outcome receipts probe observed bytes and run vision QA against must/avoid. Named-model generate still returns the provider output without that judge.
No. A failed check is stored on the receipt as a signal. Automatic retries, extra charges, and refunds for failed checks are not live.
Store the checks. Show them on the job. Do not claim a universal best model.