At a Glance
Enforce a post-debate verification gate that requires every claim to be tied to the debate record; doing so raises groundedness and either produces verifiable conclusions or clearly signals disagreement.
ON THIS PAGE
What They Found
When you force the final synthesizer to justify each claim against a frozen debate record, many fluent but unsupported 'consensus' statements disappear. The Active Provenance Gate raised measurable grounding while intentionally producing Divergence Reports when evidence was missing. Under heavy conflict, the system prefers to flag disagreement (and ask for repair) rather than publish a plausible-sounding but ungrounded answer. The approach keeps most of the original text while restructuring it around verifiable facts. This aligns with the Mutual Verification Pattern.
Key Data
1Provenance Fidelity rose from 0.288 to 0.617 in high-conflict runs, and from 0.183 to 0.586 in low-conflict runs after active auditing and bounded repair.
2The gate emitted Divergence Reports for 75.0% of runs under severe conflict and 60.0% under low conflict, preventing publication of ungrounded syntheses.
3Bounded self-healing retained 97.3% of the original text length while reworking outputs to match verified evidence.
Implications
Engineers building systems where multiple AI agents debate and produce recommendations should care because the gate prevents silent, hard-to-detect hallucinations at the publication step. Technical leaders and evaluators of decision-support stacks should care because the approach gives a clear, auditable signal when the AI cannot produce a verifiable consensus. See Multi-Agent Research Synthesis.
Need expert guidance?We can help implement this
Key Figures
Fig 2: Fig. 2: State machine architecture of the Active Provenance Gate. The system aggregates debate artifacts into a closed evidentiary corpus, decoupling generation (Executive Proposer) from verification (Global Auditor). A hard computational gate ( P F ≥ 0.95 PF\geq 0.95 ) enforces NLI auditing. If a synthesis is ungrounded, bounded self-healing is triggered ( I T E R < 3 ITER<3 ); exhausted budgets trigger a hard exit via a programmatic Divergence Report.
Fig 3: Fig. 3: Synthesis grounding under epistemic shock.
Fig 4: Fig. 4: Calibrated Trust and Groundedness Perception
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreYes, But...
Results come from a synthetic testbed (90 annotated trajectories) and a small user study (N=33), so real-world behavior with domain experts under time pressure may differ. The method assumes a closed evidentiary world — it verifies claims against the frozen debate trace, not against all external reality. Relying on language models for the final verification step risks bias and circularity; human or benchmark comparisons remain important for high-assurance deployments. This is consistent with Defense in Depth Pattern.
Methodology & More
Multi-agent debates often yield fluent final reports that hide unresolved disagreements by stitching together partially compatible facts. The Active Provenance Gate (APG) is a post-debate enforcement layer that freezes the debate artifacts into a closed corpus, requires the proposer to synthesize only by compression, conflict exposure, and deduction, and then runs an explicit verifier that checks each major claim against that corpus. If a candidate synthesis fails a strict evidential threshold (provenance fidelity ≥ 0.95), a bounded self-healing loop attempts limited repairs; if repairs exhaust the budget, the system emits a Divergence Report that documents conflicting stances instead of publishing a possibly fabricated consensus. This behavior mirrors the Accountability Diffusion approach in failure handling. In experiments on a 90-run, knowledge-annotated testbed and a blind A/B trust study with 33 participants, the APG substantially increased grounding scores and surfaced disagreement under epistemic stress. The approach raised terminal provenance fidelity, preserved most of the original wording via non-destructive repair, and intentionally rejected outputs that could not be reliably grounded (75% divergence under severe conflict). Practical trade-offs include extra verification cost, sensitivity to verifier model choice, and the closed-world verification assumption. The design is useful for anyone who prefers visible disagreement and verifiable claims over silently confident but unsupported recommendations, and future work focuses on moving some verification into the live debate loop and introducing interactive, claim-level adjudicators and human-in-the-loop arbitration. This aligns with Tree of Thoughts Pattern and Accountability Diffusion.
Need expert guidance?We can help implement this
Credibility Assessment:
Authors have low h-index (1 and 8), affiliation is a non-top university, and venue is arXiv with no citations — fits Emerging / limited info.