What question did this study set out to answer?

This work aims to develop a governance framework for managing hallucinations in large language models.

May 19, 2026Open Access

Truthfulness Assurance for Large Language Models: An End-to-End Architecture for Hallucination-Risk Governance, Calibrated Abstention and Safe-Default Routing

Key Points

This work aims to develop a governance framework for managing hallucinations in large language models.
Developed an end-to-end architecture for hallucination-risk governance.
Utilized evidence discipline, knowledge-boundary control, and runtime truthfulness gates for supervision.
Included the Orion Veritas Evidence Pack v1.0 for structured review and local overhead context.
Proposed methods to govern unsupported expressions within large language models.
Introduced mechanisms that enhance auditing and conflict arbitration in AI outputs.

Abstract

This restricted technical paper presents truthfulness assurance as an end-to-end pre-expression governance architecture for large language model systems. The central claim is that high-stakes AI systems should not merely detect hallucinations after an answer has formed. They should govern whether a candidate answer is allowed to cross the expression boundary at all. The paper develops a technical architecture for hallucination-risk governance, calibrated abstention and safe-default routing. It formalizes how large language model systems can use admissible evidence discipline, knowledge-boundary control, claim-to-evidence mapping, evidence-conflict arbitration, runtime truthfulness gates, receipts, manifests and replayability to make unsupported expression harder to emit, easier to stop and easier to audit. The restricted Technical paper is accompanied by Orion Veritas Evidence Pack v1.0, a redacted reviewer-facing evidence package for qualified technical review. The package supports bounded inspection of conservative route behavior, schema validity, identical-input replay determinism, zero-ALLOW disclosure, ALLOW-readiness analysis and evidence-package consistency verification. A separate operational-overhead addendum provides bounded local overhead context. The current evidence posture is deliberately bounded. Evidence Pack v1.0 should be treated as Level-2 candidate evidence with bounded empirical route-behavior support, not Level-3 institutional validation. The route-behavior run uses a controlled structured fixture, not a broad natural-language prompt diversity benchmark. This paper does not claim universal truth, solved hallucination, deployment certification, regulatory approval, production readiness, governed runtime ALLOW utility, live retrieval evidence, full production latency or unrestricted institutional validation. It is a technical architecture and restricted evidence-review release intended to support qualified technical and institutional review. This paper is part of the Orion Veritas Program, the verification-first track within Project Orion and the broader Auren research arc. Project Hub & updates: Canonical papers and DOIs are archived on Zenodo. For the Project Orion and Orion Veritas research overview, publication roadmap and verification snapshots, see www.auren.one. Disclaimer line: Restricted technical research preprint. Not a compliance filing, certification claim, deployment approval, regulatory determination or commercial offering. Not affiliated with, endorsed by or performed on behalf of any employer or institution. Access note: Restricted access. Redacted Technical evidence package and operational-overhead addendum are included for qualified technical review. Full Master evidence archive is reserved for controlled institutional review.

Read Full Paperexternally

Perguntar à IA

Bookmark

View Full Paper