Skip to main content

The approach

How Helix reasons about ambiguous sensitive content.


A strong reviewer asks five questions

A strong reviewer asks five questions, in order.

  1. What does this look like?
  2. What else could it be?
  3. What context is it in?
  4. Is it actually risky here?
  5. Am I confident enough to act?

Helix mirrors that process: each step maps to a subsystem in the engine.


Design principles

Confounders are first-class.
Most systems focus on finding positives. Helix is exceptional at rejecting convincing negatives.
Uncertainty is structured, not a softmax score.
Predictive confidence, disagreement, novelty, validator conflicts, and surface-quality degradation are separate signals.
Abstention is mandatory.
A high-trust system without abstention will eventually bluff. Abstaining is a real product behavior — not a low-confidence variant of suppress.
Calibrated, not confident.
Raw model scores are not trustworthy enough to drive action policy. Every score that crosses a decision boundary is calibrated; every release reports calibration deltas.