An LLM judge is a second full model call. The probe is one linear read.
To score every response for deception, an LLM judge makes a second full model call. A white-box probe does one linear read of activations you already computed — AUROC 0.93 (0.96 best layer), including on the live serving forward pass. Slide to see the compute gap.
Drag the slider — or tap a preset — to set your monthly monitoring volume.
Responses monitored / month
1,000,000/ mo
1K1M100M
LLM judge
$12,000
per month
White-box probe
$0.02
per month
Both bars share one scale. The probe's cost is so small its green bar is enlarged just to stay visible.
600,000x cheaper
≈ $12,000 saved every month vs. an LLM judge
Architectural estimate (order of magnitude): a frontier LLM judge ≈ $0.012 per response — a whole extra model call (~1.5k input + 0.5k output tokens). The probe reads activations the model already computed: one linear read over 62 layers, no extra call, no added latency. These dollar figures illustrate the architectural gap; they are not a billed benchmark.
Same accuracy, one linear read instead of a second model.
On google/gemma-3-27b-it, the FAR diverse linear probe (last-token-generation) reaches AUROC 0.93 averaged / 0.96 at the best layer — matching black-box LLM-judge detection. Reading activations directly off the live vLLM serving forward pass scores AUROC 0.928: essentially the same accuracy as offline, with no second model call.
Why it's cheap
An LLM judge re-reads the response and reasons about it with a second frontier model. The probe is a small linear detector reading the internal activations from the pass that already produced the response — nothing extra to run. Needs open weights.
Parity means the probe matches LLM-judge detection AUROC at a fraction of the compute — measured on the same forward pass. Not a claim that the probe is "better" than a judge.
Catch deception in your own models
FAR.AI builds white-box probes that flag lying, sandbagging, and sycophancy in real time.
Real FAR validation · google/gemma-3-27b-it · diverse linear probe · AUROC 0.93 (0.96 best layer), 0.928 on the live serving path · scores are live probe outputs.