Watch it decide. Deterministic rules first, a calibrated 150 MB on-device model only for genuinely ambiguous text — and an escalation, never a silent wrong answer.
Governed memory — recalled, corrected, forgotten, with provenance.
Inference state reused; every call measured byte-for-byte.
The cheapest capable path — or an escalation, never a guess.
That is not the product failing — it is the product. Tap a probe for the full finding; every number carries its workload.
How to read these numbers: every figure above carries its workload and sample size, because a routing number quoted without its workload is marketing, not measurement — probe 4 is the proof. Nothing on this page comes from a demo box or an unnamed benchmark.
A gateway that knows when not to use a model is worth more than one that reaches for a model every time.
Cost per call bills a refusal, an error, a miss — as if they worked. RouterML’s verifier checks every response against its contract: a rejected outcome counts as cost, never as success.
This page carries the number the day it exists with a named workload attached.
Calibration runs on your traffic during onboarding, and the system enforces it. That cuts both ways — ours can’t be dropped in blind, and neither can a competitor’s. The moat is the measurement.
Our gate’s accuracy is real only on the traffic it was fitted to — we measured that the hard way, and built the system to refuse shortcuts. Onboarding fits the gate to your workload and hands you the calibration report. That’s the product.