Governed routing agent
The cheapest model
that clears the bar.
rig-route dispatches every prompt to the cheapest model in your fleet whose quality clears the prompt's own floor and whose latency clears its budget — with a sealed ProofPacket on every batch, not a dashboard you have to trust. 88% cost savings, 1.5% quality regression, faster p99 than the incumbent default.
[higgsfield: three faint stone nodes converging into one bright gold node on a black field, engraved circuit line art]
rig-route · 6-model fleet · 500-prompt batch, seed 42
Why teams install
What rig-route earns back
88% cost savings, not a cherry-picked demo
Real batch of 500 prompts across 5 complexity tiers: priced after a short conversation baseline (always-gpt-4) down to priced after a short conversation routed — measured, reproducible across 6 seeds.
‹3a9ad340›Quality floor is a gate, not a suggestion
Average quality regression holds at 1.50%, comfortably under the 2% ceiling — every prompt still clears its own floor, it just doesn't overpay for it.
‹7fdfc177›Faster at p99, not just cheaper
Routed p99 latency is 1500ms vs. the incumbent's 1800ms — most traffic never touches the slowest, most expensive model in the fleet.
‹85e50da1›How it works
Three steps. No faith required.
01
Floor
Every prompt gets a complexity read: a quality floor and a latency budget, not a flat SLA.
02
Route
The cheapest fleet model whose quality ≥ floor and latency ≤ budget wins; no match falls back to the strongest model, never dropped.
03
Seal
Cost savings, quality regression, and p99 latency are gated and sealed into a ProofPacket per batch — not self-reported.
Proof band
One packet. Every claim accounted for.
packet rig/rig-route · batch 500 prompts claim "beats always-gpt-4 baseline by ≥40% cost saved, ≤2% quality regression, no p99 regression" baseline gpt-4 · priced after a short conversation · p99 1800ms · quality 97.00 routed 5-model mix · priced after a short conversation · p99 1500ms · quality 95.54 savings 88.41% · floor 40% quality 1.50% regression · ceiling 2% p99 PASS — 1500ms ≤ 1800ms baseline themis SEALED — 01259648
Every rig-route batch emits a packet like this — hash-chained, verifier-signed, and queryable from your own dashboard. How ProofPackets work →
Objections
Answered before you ask
Isn't this just LangSmith or a model gateway with pricing rules?
A gateway logs decisions to a dashboard you have to trust. rig-route seals a runtime ProofPacket per batch — cost savings, quality regression, and p99 latency independently re-verified against the gate thresholds, not self-reported.
Won't routing to cheaper models tank quality?
The gate exists specifically to catch that: quality regression is capped at 2% and checked on every batch. The planted-failure test — route-all-to-cheapest, ignore the floor — drives regression to 3.6% and the gate goes red, on purpose, so you know the guardrail actually works.
Why trust the cost-savings number?
It's computed against the real baseline (always the incumbent default model) on the same prompt batch, not a synthetic comparison — and it's sealed into the same hash-chained packet as the quality and latency gates, so you can't cherry-pick one metric.
Buyers
Who runs rig-route
Platform Lead
AI SaaS at scale
Head of FinOps
LLM spend owner
CTO
seed-stage, usage-based pricing
Pricing
Start free. Pay when it pays.
Generalist ships today. Domain Expert adds vertical memory + priors. Custom is built to your stack.
Free
Priced after a short conversation
100K tokens a month routed, 2-model fleet.
Team
Priced after a short conversation
Unlimited routing, full fleet, per-request ledger.
Pro
Priced after a short conversation
Dedicated verifier, custom quality floors, SSO.
Automated routing decisions that affect what model answers a user are exactly the kind of consequential AI decision SB 26-189-style disclosure rules target. rig-route's per-request ledger and sealed ProofPackets are built to answer that audit before it's asked.
