sAIfe

About the Project

What happens when an AI agent is confidently wrong halfway through a task?

I built sAIfe, an agentic AI framework in which an external agent checks the main LLM agent’s answer for hallucinations and alerts the user promptly. It takes an AI answer, splits it into atomic claims and checks each claim with independent detectors.

Rationale

Language models sound just as sure of themselves when they are wrong. In an agent that is dangerous: a made-up fact in step two quietly becomes the input to step three, and by the final answer nobody can tell which sentence was invented.

Two things bothered me about the usual fixes. Self-review is often the same model with the same blind spots, and safety checks tend to fail silently, shipping the answer anyway when a verifier times out. So I set myself a Safety-by-Design goal: a guard whose safety lives in the architecture, not in a prompt asking the model to be careful.

How It Works

Each claim is checked by four kinds of detector:

• grounding — a judge returns exact quotes from the user’s sources, and code confirms each quote is really there;
• self-consistency — agreement across resampled answers, which can reach “Likely” but never “Supported”;
• web evidence — search excerpts for world-knowledge errors when no sources are given; and
• deterministic checks — pure code for wrong arithmetic, impossible dates, dead links and invented Python APIs.

Flagged claims are corrected (only with a verified quote), hedged or removed, and every edit is recorded. The answer ends as Verified, Caution or Blocked, with plain-language alerts about what was changed and what could not be checked.

Designed to Fail Closed

Eight principles shape the design, including fail closed, independent verifiers, bounded loops and no silent edits. A judge’s quote only counts as evidence if a normalised version of it, at least 12 characters long, is found in the source. The safety loop stops when it runs out of passes, budget or progress, and anything left unchecked becomes “Unverifiable”.

sAIfe comes as a CLI, a Python SDK and a FastAPI server, and runs on the user’s own keys. A model from a different lab always judges the answer, and every verdict streams to the sAIfe Monitor dashboard, which users host themselves on free tiers.

Project Outcome

In a preliminary five-item benchmark, the residual hallucination rate fell from 100% to 0%, with 9 of 9 planted errors detected and 0 silent failures. The full benchmark run is next, along with an MCP server so any MCP-capable assistant can call sAIfe directly.