Most behavioral AI models what people do. Epistemic engines model what people believe — and crucially, where their beliefs diverge from ground truth. This makes them uniquely powerful for fraud detection, negotiation strategy, and credibility assessment.

The Engines

  • Truth/Reality Modeling Engine — Maintains a model of what the actor believes vs what external evidence suggests is true. Inconsistency between belief and behavior is a strong signal.
  • Ground Truth Framework — Calibrates engine outputs against observable facts. When a prediction can be verified, the framework records the accuracy and feeds it to the meta-learning engine.
  • Ontological Uncertainty Engine — Models cases where the categories themselves are contested or unclear. Flags outputs where definitional ambiguity is a significant source of score variance.
  • Self-Modeling Engine — The system models its own limitations: where its engines have historically been wrong, in what contexts, and with what frequency.

Code Walkthrough

// Ground truth calibration
function calibrateAgainstGroundTruth(engineId, predictions, outcomes) {
  const calibration = predictions.map((pred, i) => ({
    predicted:  pred.score,
    actual:     outcomes[i].value,
    error:      Math.abs(pred.score - outcomes[i].value),
    contextId:  pred.contextId,
  }));

  const mae = average(calibration.map(c => c.error));
  metaLearningEngine.reportCalibration(engineId, mae, calibration);
  return { engineId, mae, sampleSize: predictions.length };
}

What to Watch For

  • Self-modeling engines require ground truth at scale. Start collecting outcome data from day one, even before you have enough for meaningful calibration.
  • The self-modeling engine's own confidence estimates should themselves have uncertainty bounds. It is not a reliable oracle about the system's own reliability.