AIHA Office Agentic AI & Hyperautomation Implementation Office · demo

Architecture & solutions

Hallucination control

Concept and policy: three verification gates (grounding ≥ 0.90, independent critic model, deterministic validators), ten controls mapped to risk tiers, a claim-checker demo and a policy generator.

Controls

  • Grounding check (≥ 0.90): Every claim must be supported by retrieved context; similarity and entailment scored 0–1.
  • Independent critic model: A second model, from a different family, compares extracted facts against source records and flags contradictions.
  • Deterministic validators: Regex, JSON schema and business rules (IBAN, amounts, dates, limits) — no model involved.
  • Retrieval quality: Curated sources, access-controlled retrieval, freshness dates, chunking tested on evaluation sets.
  • Constrained generation: Structured output, closed vocabularies, tool results instead of model memory for facts and numbers.
  • Citations required: Answers show the source for each claim; no source → the claim is removed or marked.
  • Abstain rule: If evidence is insufficient the agent says so and routes to a human instead of guessing.
  • Confidence calibration: Confidence scores calibrated on labelled cases so thresholds mean what they say.
  • Evaluation & red-teaming: ≥ 200-case evaluation set per use case, regression gate on every change, adversarial tests.
  • Production monitoring: Grounding, override and complaint rates tracked; drift alerts; sampled human review.

Public demo with fictional data (Example Group). © 2025–2026 Russo-Got Marius Mihail & PIR2-IT. All rights reserved.

AIHA Office

More in this demo