++
INFERENCE

Routing is a loop,
not a lookup.

Five stages run on every call. Each one is scored, bounded, and signed — so the router earns trust instead of asking for it.

01 SCORE02 BOUND03 RECOVER04 LEARN05 PROVE
01 · SCOREA posterior, not a price listAinfera keeps a learned quality posterior per model, per task family — how likely each of the 233 models is to actually finish this kind of work. The match runs in one pass.
anthropic/claude-opus-4-80.97 ←
openai/gpt-5-50.94
deepseek/deepseek-v4-pro0.91
TASK: refactor-multi-file · POSTERIOR MATCHED IN 38MS
02 · BOUNDConstraints, not suggestionsCost ceilings and latency targets ride with every key. The router treats them as hard constraints — a candidate that would breach them is never called, whatever its score.
max_cost ≤ $0.004 / call✓ enforced
p95_latency ≤ 900ms✓ enforced
xai/grok-4-5 · $0.006excluded pre-call
BOUNDS RIDE THE KEY · NEVER BREACHED
03 · RECOVERFailures retry down your setWhen a provider degrades mid-call, the router retries down your candidate set — same bounds, next best posterior. The recovery is logged and signed like any other decision.
fireworks/llama-5-405b✗ degraded 503
└─together/llama-5-405b✓ recovered · 210ms
RETRY LOGGED · SIGNED · VISIBLE IN THE AUDIT LOG
04 · LEARNSharper with every outcomeCompleted tasks feed the posterior. Routing improves on your traffic — not a public benchmark — and models are quality-gated: they enter on evaluation and exit on regression.
PASS RATE / WEEK
05 · PROVEEvery hop on the signed audit logFour hops per call, each hashed and Ed25519-signed to an append-only public audit log. Anyone can verify the router did what it claims — no account, no API key.HOW VERIFICATION WORKS  ›
request.routed──✓ signed
provider.responded──✓ signed
receipt.created──✓ signed
ledger.settled──✓ appended
++

Route the next call by outcome.