3–6 months · partial
Route intelligence needs a fixed set of jobs. Authenticated GET /api/v1/outcomes/benchmarks returns the frozen seed corpus plus observed receipt counts. Mix shares stay the seed design. This page does not run the suite.
Frozen seed corpus · observed receipt overlay
Frozen jobs across the three launch workloads. Seed corpus ids stay fixed.
Image-to-video, product image, vertical product ad.
No inspected receipts yet. Overlay stays empty until wait+judge writes stats.
GET /api/v1/outcomes/benchmarks keeps the frozen 120-item seed corpus and overlays observed pass/fail from inspected receipts.
Each corpus item is an intent, inputs, and the evaluation checks that define success. Changing the jobs later makes historical scores incomparable.
Image-to-video under budget, ecommerce product image, and short vertical product ad. Depth on those three beats a shallow catalog of every media type.
Before a route is eligible for automated selection, it has to produce measurable results on the same items the current router is scored against.
Community anecdotes and one-off demos do not replace the corpus. Human or agent evaluation scores are stored as signals, not as marketing copy.
The same jobs, over time, across models, workflows, and later sellers.
Do not recruit broad supply until these jobs have a stable baseline.
Mix shares stay the frozen seed design. Observed receipt counts overlay when present.
Stripe Checkout, wallet env, and seller payouts stay off this page. Authenticated outcomes execute is the live charge path. Named-model MCP and /ai-router remain available.
Do not call FAL or other provider credentials. Candidate scores and outputs on these pages are illustrative.
Authenticated callers read the frozen seed corpus plus observed receipt overlays. The mix shares are still the seed design, not a public gating suite.
The plan is that new capabilities run a standard benchmark and policy check before sandbox traffic. How much of the corpus is disclosed to sellers is not decided here.
A router that cannot fail a known job cannot claim it beats a manual baseline.