Two-Tier Self-Improving System
Engineering depth, not a sales demo
Fast Loop
Minutes · Prompt evolution
Response generation
The LLM generates a draft response to the user query
Hallucination detection
Our sLM verifies factual consistency of the generated response in real-time
Reflection
Classifies what type of hallucination occurred based on detection results
Prompt evolution
Automatically revises the system prompt to match the classified pattern
Instant deployment
The revised prompt takes effect starting from the next request
Slow Loop
Days · Model training
Pattern accumulation
Hallucination patterns and correction data collected from Fast Loop accumulate daily
Domain evaluation
Accumulated data is automatically analyzed against domain-specific evaluation criteria
Reward modeling
Generates reward signals from correct-response and hallucination-response pairs
Fine-tune
Retrains the sLM with generated reward data, specializing it for the domain
Auto deployment
The trained model is auto-deployed to production once it passes evaluation
Hallucination Detection sLM — The Heart of the System
The sLM sits at the center of both loops. In the Fast Loop, it judges hallucinations in real-time for every response. In the Slow Loop, accumulated pattern data makes the sLM itself more accurate. Because both loops share a single model, the entire system operates as one feedback cycle.
94.25%
Hallucination detection accuracy
4B
Lightweight model parameters
~1/100
Cost vs frontier models
sLM — Hallucination Detection Pipeline
Works Inside Airgaps
Patronus / Predibase
SaaS-only architecture cannot be deployed in airgapped environments. Evaluation data is sent to external servers, and model training only works in the cloud. Unusable in finance, defense, and public sector environments where data sovereignty is required.
Fully on-premises operation
Self-hosted sLM + local fine-tune + zero data exfiltration. Every component runs within customer infrastructure. The full loop of hallucination detection, evaluation, and model improvement works without internet connectivity.
It must be integrated to close the loop
n8n + external services
Connecting external hallucination detection APIs, separate evaluation tools, and third-party training platforms via workflow tools like n8n can create a pipeline. But results from each stage don't automatically feed into the next. A closed loop where hallucination patterns revise prompts, and revised prompt results feed back into model training, is impossible without manual intervention.
One system
Builder + Verifier + Evaluator + Trainer are integrated in one system. When hallucinations are detected, prompts evolve. When patterns accumulate, the model trains. The system becomes more accurate on its own without human intervention.
Integration Options
Self-hosted sLM Mode
Deploy CT's 4B sLM directly on customer infrastructure. Ideal for organizations with airgap requirements and data sovereignty needs. All data stays internal, and fine-tuning is performed locally.
External Provider Mode
Integrate with external LLM providers like OpenAI, Anthropic, etc. CT's hallucination detection and self-improving loop operates as a layer on top of external models. Enables rapid adoption and flexible model switching.
Annual Product License
Adopt the full Two-Tier system on an annual license. Includes hallucination detection sLM, closed-loop infrastructure, and monitoring dashboard
1-month deployment support
FDE deployment support included with the annual license. Covers environment-specific deployment, domain evaluation set creation, and initial closed-loop setup