Silent errors
A single model returns a plausible, wrong value with no signal that anything went wrong — and it flows straight into your data.
···
The end of silent LLM errors
Crosscheck sends the same prompt through two providers on the BTL gateway. Where they agree, it accepts. Where they disagree, a strong model judges it and flags it — so a cheap model's confident mistakes never reach your data.
89.7% → ~93% field accuracy
100% flag precision
~$0.004 real gateway cost / run
Powered by the BTL runtime · one gateway
OpenAI/DeepSeek/Google/OpenRouter
"Order: 3 cartons, 12 bottles per carton."
field · total_bottles
gemma-3-4b · cheap
12 flagged
crosscheck · resolved
36 verified
disagreed → gpt-4.1-mini judged: 3 × 12 = 36
The Problem
A confident wrong answer looks exactly like a right one
Shared blind spots
When two models make the same mistake, cross-checking can't see it. We measure that blind spot and report it — never hide it.
Flaky infrastructure
Providers 500, drop connections, and rate-limit. Crosscheck retries and fails over to the other provider so the pipeline keeps running.
Cost opacity
You can't see what inference truly costs. Crosscheck reads the real charge and cache savings straight off the gateway's response headers.
The Approach
Two providers, one gateway, zero silent errors
01 · FAN-OUT
Same prompt, two models
Every field runs on a cheap and a strong model in parallel — one OpenAI-compatible call each, both through the BTL gateway.
02 · RESOLVE
Agree → accept · disagree → judge
Matching answers are auto-accepted. Disagreements go to the strong model as judge and are flagged for a human to review.
03 · SURVIVE
A provider drops? Fail over
On a 5xx, 429, or dropped connection the call retries and fails over — and real cost is read straight off the response headers.
89.7%cheap alone
~93%with crosscheck
100%flag precision
$0.004real gateway cost
78tests, all green
See it run
Read the code →
The numbers