PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Mitigating Fabricated Consensus: The Active Provenance Gate for Multi-Agent Debate Synthesis

Jakub Masłowski, Jarosław A. Chudziak

arXiv:2609.31422v1cs.MAcs.AIcs.CL

TL;DR

Fluent synthesis can turn unresolved debate into consensus that is not grounded in the debate record. The paper introduces a post-debate provenance gate that audits claims, attempts bounded repair, and blocks unsupported publication; across its testbed, the authors report improved provenance fidelity and a divergence mechanism when grounded agreement is unreachable.

  • Problem

    The paper addresses unsupported consensus at the synthesis stage, where fluent summaries may interpolate across partially aligned debate sources.

    Source: §I · p. 1

  • Method

    The Active Provenance Gate audits candidate synthesis claims against a closed debate corpus, applies bounded self-correction, and enforces publication constraints.

    Source: §I · p. 1, A. Principles of Active Runtime Auditing · p. 2, C. Bounded Self-Healing and Divergence… · p. 3

  • Results

    Across 90 adversarial runs, the APG increased the Provenance Fidelity floor and provided structured divergence when grounded agreement was unreachable in the testbed.

    Source: §VII · p. 5

  • Takeaways & Limitations

    The results present provenance as an active runtime publication constraint rather than a passive diagnostic record.

    Source: §I · p. 1, §VII · p. 5

    The evaluation uses synthetic simulations and hard-coded shocks, and its proxy trust study lacks a domain-expert versus crowd-operator distinction.

    Source: §VI · p. 5

Abstract

from arXiv · show

Large language model-based multi-agent debate (MAD) systems are being increasingly used as complex decision pipelines in distributed processes. However, their final synthesis phase still remains inadequately controlled. Even with detailed debate logs, summarizing models are prone to fabricating smoothly written debate consensus that is not grounded in the debate's history. To address this safety gap, this paper presents empirical research and studies if the introduction of active post-debate verification can mitigate the production of such factually unsupported summaries, while still providing valuable information. Furthermore, it is examined whether explicitly signalling divergence is preferable in the absence of a reliable compromise. The Active Provenance Gate (APG) is introduced as a post-debate verification layer that treats the source as a hard constraint, analysing the debate logs, auditing each claim, and applying self-correction. In crisis simulations, the self-healing mechanism more than doubles the average data Provenance Fidelity in difficult condition scenarios, before the strict gate blocks unsupported claims and generates divergence reports. In the human study, a vast majority of the users (over 75%) preferred a report explicitly stating failure in critical scenarios, despite most of them perceiving fabricated consensus from the baseline system as more fluent. Our main contribution is the transition of data origin tracing from passive logging to active conditional blocking before publication.

I. INTRODUCTION

Multi-agent systems rely on synthesis to compress debate logs, but fluent synthesis can blur unresolved conflict into unsupported consensus. The paper proposes the Active Provenance Gate (APG), which checks claims against the debate record and blocks publication when grounding constraints fail.

I. INTRODUCTION

  • Synthesis can dissolve unresolved conflicts into fluent, unsupported consensus, making the transition from deliberation to publication a key hallucination risk.The risk can increase as traces and knowledge artifacts grow and sources are only partially aligned.Source: §I · p. 1
  • The paper asks whether post-debate verification can reduce fabricated consensus while preserving practical utility in high-stakes decision support.The authors argue visible dissension is less harmful than fluent agreement that may miscalibrate operator trust in crisis settings.Source: §I · p. 1
  • The APG treats the debate record as a closed evidentiary world, anchoring major claims and blocking publication when grounding constraints fail.It publishes a verifiable synthesis or a Divergence Report showing conflicting stances and lack of evidence.Source: §I · p. 1, Figure 1 · p. 2
  • The contribution shifts provenance from passive tracing to controlled publication, framing it as an operational artifact for cooperative AI.The paper contrasts this with frameworks that mainly use logs for post-hoc diagnosis and self-refinement methods without hard synthesis-boundary constraints.Source: §I · pp. 1–2

II. CONCEPTUAL FRAMEWORK: ACTIVE PROVENANCE ENFORCEMENT

The framework moves provenance from passive logging to active runtime enforcement: each synthesis claim must connect to evidence in the debate log. This is motivated by the risk that dense, conflicting information can invite plausible but ungrounded connections.

II. CONCEPTUAL FRAMEWORK: ACTIVE PROVENANCE ENFORCEMENT

  • Active Provenance shifts multi-agent debate from passive logging to enforcement of provenance in the final synthesis.Source: §II · p. 2

A. Principles of Active Runtime Auditing

  • Under active runtime auditing, each final-synthesis claim must connect directly to a particular proof in the debate log.This grounds claims in evidence rather than agent rhetoric.Source: A. Principles of Active Runtime Auditing · p. 2
  • Dense Knowledge Graph and RAG axioms can facilitate false connections, while conflicting information can yield plausible answers without source grounding.The authors therefore characterize active runtime auditing as a structural requirement for fidelity.Source: A. Principles of Active Runtime Auditing · p. 2

B. Research-by-Construction Methodology

The study constructs a crisis-management debate testbed with epistemic shocks to expose unsupported synthesis while keeping debate traces knowledge-grounded. It measures grounding through Provenance Fidelity (PF), the sentence-level share of output supported by strict entailment with the corpus.

B. Research-by-Construction Methodology

  • The study engineers a high-fidelity-trace testbed under systematic epistemic stress rather than selecting a benchmark post hoc.Source: B. Research-by-Construction Methodology · p. 2
  • The crisis-management committee simulation uses fixed budgets and conflicting telemetry shocks to create logical deadlocks.Restricting debate to knowledge-grounded traces helps attribute unsupported syntheses to the publication-stage synthesis module.Source: B. Research-by-Construction Methodology · p. 2

C. Formalization of Provenance Fidelity (PF)

  • Provenance Fidelity is the ratio of supported semantic units, evaluated sentence by sentence, to total output sentences.A(c) represents strict NLI entailment between generated sentence c and the corpus; PF = 1.0 indicates perfect traceability.Source: C. Formalization of Provenance… · p. 2

III. SYSTEM ARCHITECTURE: THE ACTIVE PROVENANCE GATE

The Active Provenance Gate is a post-debate loop that governs the transition from synthesis to publication. Its state-machine architecture separates proposal from auditing, permits bounded repair, and exits with a Divergence Report when the retry budget is exhausted.

III. SYSTEM ARCHITECTURE: THE ACTIVE PROVENANCE GATE

  • The APG wraps the synthesis boundary with evidentiary admissibility, NLI auditing, bounded self-healing, and explicit failure signalling.It is implemented as a deterministic state machine between multi-agent debate and publication.Source: §III · p. 2, Figure 2 · p. 3
  • The stated model-version note dates the versions to May 2026.Source: §III · p. 2
  • The architecture aggregates debate artifacts into a closed evidentiary corpus and decouples generation by the Executive Proposer from verification by the Global Auditor.A PF ≥ 0.95 gate enforces NLI auditing; when repair retries are exhausted, the system exits with a programmatic Divergence Report.Source: Figure 2 · p. 3

A. Evidentiary Corpus and Assembly Phase

The APG builds a bounded evidentiary corpus from debate artifacts, then uses a proposer–validator gate to check synthesis claims. Unsupported claims can trigger repair or rejection under a fixed provenance threshold.

A. Evidentiary Corpus and Assembly Phase

  • The assembly phase combines the public transcript, deduplicated KG/RAG documents, and private traces into a bounded corpus with stable addresses.The resulting corpus serves as a closed evidentiary world rather than a loose context window.Source: A. Evidentiary Corpus and Assembly Phase · p. 3
  • The Proposer may only compress, expose conflicts, or deduce within the frozen corpus, making unsupported bridges auditable violations.The paper identifies plausible bridges between partially compatible fragments as a common hallucination pattern.Source: A. Evidentiary Corpus and Assembly Phase · p. 3

B. The Verification Gate: Adversarial Constraints and NLI Auditing

  • The gate pairs a cost-effective Executive Proposer with an advanced-reasoning Global Auditor.This heterogeneous pairing is presented as the architecture’s approach to balancing cost and safety.Source: B. The Verification Gate: Adversarial… · p. 3
  • The Proposer uses consensus-oriented prompting without an explicit no-consensus mode, reflecting decision-support settings where recommendations are expected under time pressure.The paper says this pressure can produce unsupported compromises under epistemic shock.Source: B. The Verification Gate: Adversarial… · p. 3
  • The Auditor checks sentence-level entailment and rejects synthesis below the programmatic provenance threshold, even when a local verdict is positive.The example hard gate is PF = 0.95; rejected claims are captured for repair.Source: B. The Verification Gate: Adversarial… · p. 3

C. Bounded Self-Healing and Divergence Reporting

The APG retries failed syntheses through bounded self-correction, then reports divergence if the publication threshold remains unmet. The experiments use a knowledge-grounded crisis testbed, a blind user comparison, and a targeted validator ablation.

C. Bounded Self-Healing and Divergence Reporting

  • After failed validation, a bounded self-correction loop uses unsupported sentences to prompt the Proposer to remove or neutralize assertions, with at most three retries.The retry cap is intended to prevent infinite hallucination loops in unresolvable conflicts.Source: C. Bounded Self-Healing and Divergence… · p. 3
  • If retries still fail to reach PF ≥0.95, the APG exits with a JSON Divergence Report listing unresolved claims and conflicting sources.The report is produced instead of an ungrounded compromise.Source: C. Bounded Self-Healing and Divergence… · p. 3

IV. EXPERIMENTAL SETUP

  • The grounding test uses 90 knowledge-annotated trajectories from the kg_cfr_full corpus, with prompt templates and verification code, under consensus-oriented prompting.The corpus is integrated into the open-source Resilient MAS Framework.Source: §IV · p. 3
  • The blind A/B study (N = 33) compared a fluent baseline synthesis with APG reports, including Divergence Reports for S026, after participants reviewed raw debate traces.Models across the experiments used Temp: 0.5 and Max Tokens: 8192.Source: §IV · p. 3
  • The study used convenience-recruited participants, mostly higher-education students or graduates from STEM and technical fields, with varied self-reported LLM literacy.Reported literacy groups were high (21.2%), intermediate (48.5%), and elementary (30.3%).Source: §IV · pp. 3–4
  • A targeted 30-run ablation compared same-class Proposer–Validator deployments with an efficient Proposer paired with a high-reasoning NLI-focused Validator.It examines verification capacity and the trade-off between permissive thresholds that risk ungrounded synthesis and aggressive thresholds that risk inhibiting beneficial syntheses.Source: §IV · p. 4

V. RESULTS

The results describe a trade-off between synthesis grounding and operational divergence. APG combines self-healing with low-latency models while isolating the publication boundary.

V. RESULTS

  • Passive evidentiary density alone cannot prevent hallucination under epistemic stress, according to the analysis.The paper frames the findings as a trade-off between synthesis grounding and operational divergence.Source: §V · p. 4
  • The APG isolates the publication boundary as the point where synthesis is governed.The analysis describes this as a way to address hallucination that persists despite dense evidence.Source: §V · p. 4
  • A self-healing mechanism and low-latency models allow the system to maintain low computational costs while keeping operation efficient.The paper cites prior work for the self-healing mechanism and low-latency models,,.Source: §V · p. 4

A. Impact of Active Auditing on Synthesis Grounding

Active auditing raised terminal Provenance Fidelity before the strict gate, which then reported outputs still below the publication threshold. Divergence was more frequent under severe conflict than under low conflict.

A. Impact of Active Auditing on Synthesis Grounding

  • Terminal Provenance Fidelity rose to 0.617 under high conflict and 0.586 under low conflict, from baseline scores of 0.288 and 0.183, respectively.These terminal scores are measured before the final gate evaluation; the baseline values are initial Provenance Fidelity scores.Source: A. Impact of Active Auditing on… · p. 4
  • Even perfectly grounded RAG/KG traces can leave unsupported interpolation in unconstrained synthesis, and denser evidence can extend that interpolation surface.The hard gate enforces a grounding floor; the paper contrasts conventional summarization’s performance under mild conflict with its unreliability when conflict is severe.Source: A. Impact of Active Auditing on… · p. 4
  • 75.0% of severe-conflict outputs and 60.0% of low-conflict outputs failed the PF ≥0.95 gate and triggered Divergence Reports.The gate rejects outputs that fail the strict publication threshold rather than silently publishing unsupported consensus.Source: A. Impact of Active Auditing on… · p. 4

B. Operational Efficiency and Parameter Sensitivity

Operational tuning exposed weaknesses in symmetric validation and favored a PF threshold of 0.95. In high-shock stalemate, divergence reporting earned higher trust, while evaluation remains bounded by synthetic settings and closed-world auditing.

B. Operational Efficiency and Parameter Sensitivity

  • Symmetric Gemini 3 Flash wrongly validated structurally unsafe compromises 85.6% of the time; asymmetric Gemini 3.1 Pro validation reduced fidelity to 0.300 and rejected 63.3% of irreconcilable outputs.Source: B. Operational Efficiency and… · p. 4
  • A preliminary sweep favored PF = 0.95 over PF = 0.77 and PF = 0.99; sessions failing after two correction hints usually oscillated, making further retries inefficient [1].The threshold sweep used n = 10 per threshold, and the system capped synthesis iterations at three.Source: B. Operational Efficiency and… · p. 4

C. Calibrated Trust Dynamics

  • In high-shock scenario S026, 75.8% identified APG output as more faithful, although 66.7% rated the Baseline more fluent and diplomatic.Source: C. Calibrated Trust Dynamics · p. 4
  • In the zero-sum stalemate, divergence Report B earned higher calibrated trust than the Baseline’s fabricated compromise (M = 3.27, SD = 1.21 vs. M = 2.45, SD = 0.90; p = 0.0039).Trust levels were comparable in the energy allocation scenario.Source: C. Calibrated Trust Dynamics · p. 5

VI. DISCUSSION AND EVALUATION

  • Recovered sessions retained 97.3% of text length by restructuring directives around verified axioms rather than truncating, while Divergence Reports identified deliberative deadlocks.In serious conflicts, explicit failure reporting was a significant factor in enhancing calibrated trust.Source: §VI · p. 5
  • Evaluation used synthetic simulations and hard-coded shocks; PF checks the aggregated corpus rather than external reality, and zero-shot NLI auditing remains subjective and prompt-sensitive.The proxy study included N = 33, and auditor-based evaluation risks metric circularity.Source: §VI · p. 5

VII. CONCLUSIONS

The paper identifies fabricated consensus at the publication boundary as a failure of fluency-optimized synthesis. It presents APG as a post-debate enforcement layer and points to real-time enforcement and operator-in-the-loop arbitration as future directions.

VII. CONCLUSIONS

  • Fluency-optimized synthesis can turn irreconcilable grounded positions into fabricated consensus, a failure passive provenance logging does not reveal.Source: §VII · p. 5
  • Across 90 adversarial runs, APG enforced a PF ≥0.95 publication threshold, bounded repair, and structured divergence reporting when grounded agreement was unreachable in the testbed.Source: §VII · p. 5
  • Future work proposes active provenance enforcement during real-time deliberation, dynamically calibrated claim-level judges, and interactive Divergence Reports for operator arbitration.Source: §VII · p. 5
Loading 2609.31422v1…