3–6 months · live (authenticated)
Keep the signals that made the pick.
Eligible routes must fit capability, policy, and budget. Outcomes already rank with quality, latency, cost, and reliability. Inspected receipts write those signals back. Named-model generate is a different ranker and stays that way.
- Gate
- Fit + policy + budget
- Score
- Four weights, two penalties
- Moat
- Task-specific data
Strategy formula · not the live ranker
eligible = capability_fit AND policy_fit AND provider_policy AND spread_policy AND price <= budget AND latency <= deadline
score =
quality_weight * expected_quality
+ speed_weight * latency_score
+ cost_weight * cost_score
+ reliability_weight * success_rate
- regeneration_penalty
- policy_risk_penalty- Expected quality0.25 × 0.8025 → 0.201
- Latency score0.25 × 0.6 → 0.150
- Cost score0.30 × 0.96 → 0.288
- Success rate0.20 × 0.986 → 0.197
- Regeneration penalty— × 0.08 → -0.040
- Policy-risk penalty— × 0 → 0.000
Score only what is eligible
Outcomes rank with the weighted formula. quality_source becomes observed after five inspected receipts. Named-model generate still uses reliability-first within band.
- 01
Filter, then rank
A route over budget or outside policy is not scored. Eligibility is a hard gate so the ranker cannot “almost” violate a spend cap.
- 02
Weight by workload
A 30-second SLA job and an ecommerce hero still do not share one universal quality number. Preserve per-intent weights.
- 03
Penalize hidden dissatisfaction
Regeneration rate is a quality signal even when the first output was delivered. Policy risk is a separate penalty, not a footnote.
- 04
Show the breakdown
The requesting agent — and later, operators — should see why a route won. Opacity is how a router becomes untrustworthy.
- GateFit + policy + budget
Ineligible routes never enter the weighted score.
- ScoreFour weights, two penalties
Quality, speed, cost, reliability, minus regeneration and policy risk.
- MoatTask-specific data
Learning which path wins for an intent is harder to copy than a model list.
This page does not charge.
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.
- TODO(generation)
Do not call FAL or other provider credentials. Candidate scores and outputs on these pages are illustrative.
Plan FAQ
Is this the formula the live router uses today?
Outcomes use this weighted formula. quality_source becomes observed after five inspected receipts. Named-model generate still prefers reliability among routes within about 10% of the cheapest eligible cost.
Can I tune the weights per agent?
The strategy says weights should evolve by workload. Per-agent preference knobs are not specified on this concept page.
The formula can change. The signals cannot vanish.
If you cannot replay why a route was chosen, you cannot improve it.