How It Works

Three steps to
trustworthy AI

01

Wrap any LLM

One line of code attaches the Verace metacognitive layer. The base model stays completely frozen — invisible until trained.

Works with Llama, Mistral, Qwen, and any open-weight model.

02

Train the adapter

5,000 steps of lightweight training teaches the nervous system to monitor the model’s internal states.

< 7% parameter overhead. 15–25% inference overhead.

03

Deploy with confidence

Every response now carries calibrated uncertainty. Route high-confidence directly, hedge medium, escalate low.

Zero-latency detection — caught during generation, not after.

agent.py
1from verace import enhance
2
3model = enhance('meta-llama/Llama-3-70B')
4response = model.generate(
5 "Delete user account #4523"
6)
7
8if response.confidence < 0.5:
9 ask_human() # Agent pauses
10else:
11 execute(response) # Confident

Know when to ask for help vs proceed. That's the entire product.

Validated Results

Numbers that speak

Tested on TinyLlama 1.1B5,000 training steps • <7% parameter overhead

0.0%
Hallucination Detection
AUROC
+0.0%
Intelligence Index
Composite improvement
-0%
Perplexity
Model gets better
+0.0%
Entity Recall
@500 tokens

Head-to-Head Comparison

Base LLM vs Verace-Enhanced

MetricBase LLMVeraceChange
Detection AUROC0.5000.917+83%
F1 Score0.0000.823from zero
Uncertainty Gap0.0001.878from zero
Selective Acc @30%0.4300.588+37%
NTP Accuracy47.7%51.2%+3.5%
Perplexity10.409.80-6%

The model doesn't just detect hallucinations — it actually gets better. When allowed to abstain on 30% of queries, selective accuracy reaches 58.8% vs 43.0% baseline.