Auditing Action Settlement in LLM Agent Environments: Order, Progress, and Replay

Haotian Chen; Bowen Ye; Yuning Zhang; Jingkun Yu

cs.AIOct 1, 2026

Auditing Action Settlement in LLM Agent Environments: Order, Progress, and Replay

Authors: Haotian Chen, Bowen Ye, Yuning Zhang, Jingkun Yu

Organizations: School of Cyber Science and Technology, University of Science and Technology of China · School of Education, Shanghai Jiao Tong University · SWJTU-Leeds Joint School, Southwest Jiaotong University

Abstract

Concurrent actions in large language model (LLM) agent environments require arbitration even when each proposal is individually valid. We implement a typed snapshot-settlement contract and audit three distinct properties: order sensitivity, useful progress, and replay consistency. Five settlement policies are tested in 28,800 exhaustive permutation trials and 2,160 scripted multistep episodes. Joint policies are spatially order-invariant conditional on fixed priorities, yet conservative rejection completes only 31.25% of agents in a six-agent doorway task versus 90.28% for random tickets; the paired improvement is 59.03 percentage points (95% bootstrap interval: 50.00-68.06). All policies preserve the tested spatial constraints, and priority arbitration still misses the independent small-instance optimum. A separate full-state journal audit exactly replays 156 checkpoints and rejects 1,332 constructed corruptions with a retained terminal anchor. The evidence concerns execution semantics, not human realism or long-run fairness.

Figures & tables

Figure 1: Snapshot–settlement contract. Upper panels illustrate joint rejection, fixed ID, and sequential controls; A has lower ID and A,B/B,A gives sequential order. Below, snapshot-based proposals undergo validation $V$ , commitment $O$ , bounded subjective updates $B$ , and signal derivation $D$ . Dashed arrows denote logging, not execution. Journal coverage starts before pending-action/process advancement. Invariance covers positions only.

Policy	Door ( $n=6$ )	Chain ( $n=6$ )	Hub ( $n=4$ )
Joint rejection	31.25	3.167	0
Sequential	89.58	5.208	100
Fixed ID	91.67	3.167	100
Random tickets	90.28	3.167	100
Rotating slots	88.89	3.167	100

Table 1: Standard-layout high-load results, 24 seeds per cell. Door/hub: mean agent completion (%). Chain: restricted mean per-agent completion time (ticks). All policies satisfy tested spatial constraints.

Figure 2: Three-policy one-tick diagnostic: eight six-agent motifs

\times

720 orders (17,280 scripted trials). (a) Distinct final position states; (b) agents moving in one tick (%), not multistep goal completion; (c) contender success over all orders, not long-run fairness. All policies have zero tested spatial violations and exact position-delta replay.

Explore similar work

Auditing Emergent LLM-Agent Collaboration through Cooperation-Obligation Coupling
Jul 29, 2026Zuyuan Zhang, Hanqing Yang, Carlee Joe-Wong +1Model Auditing Passive
Department of Electrical and Computer Engineering, The George Washington University · Department of Electrical and Computer Engineering, Carnegie Mellon University
Same Winners, Different Success Rates: Evaluating How LLM Agents Recover from Failures
Sep 28, 2026Dong Xu, Zhangfan Yang, Jiantao Wu +5Task Success Rate Large Language Model Agents
School of Artificial Intelligence, Shenzhen University · EasternDawn · School of Computer Science, University of Nottingham Ningbo
How Strongly Should Task State Influence an LLM Agent?
Sep 22, 2026Chenyu Zhang, Wonbin Kweon, Jiawei HanLarge Language Model Agents State-Tracking
University of Waterloo · Sungkyunkwan University · University of Illinois Urbana-Champaign

Jul 29, 2026, cs.MAJ/K move · Enter open · S save

Auditing Emergent LLM-Agent Collaboration through Cooperation-Obligation Coupling

Zuyuan Zhang, Hanqing Yang, Carlee Joe-Wong, Tian Lan

Department of Electrical and Computer Engineering, The George Washington University · Department of Electrical and Computer Engineering, Carnegie Mellon University

Model Auditing Passive

Open paper AskPDF arXiv

LLM-agent systems can solve complex tasks through dynamic self-organization and emergent cooperation. Auditing this process is essential because plausible intermediate or final outputs can conceal incomplete or unsupported work and poorly allocated responsibility, ultimately compromising response quality. While existing approaches may record messages, tool calls, provenance, or task dependencies, an auditability gap exists as they do not jointly represent what work remains, who is responsible for it, and what evidence justifies each work-state transition. We address this auditability gap by proposing \emph{Integrated Cooperation-Obligation REpresentation} (iCORE). It creates a unified encoding

X=(G,Q,Π)

integrating observable interactions as a cooperation graph

G

, evolving work and assignments as an obligation graph

Q

, and the audit map

Π

linking them with verifiable properties and evidence. This iCORE representation enables the auditor to certify two complementary properties: {Work soundness}, where every active decision-relevant work assertion must have a finite justification through

G

and

Π

; and {Agent-assignment stability}, which requires that no feasible alternative agent improve the declared contribution value for an evaluated obligation by more than

ε

. We establish local-to-global soundness and assignment-regret guarantees and a performance bound under stated conditions. iCORE is an instrumentation layer over workflows. Numerical results show that the full coupled state exactly reconstructs soundness and assignment defects in two execution modes and that, relative to passive full-state observation, iCORE-Audit yields absolute trajectory-quality improvements of

11.5\%

and

26.4\%

in controlled and real-LLM execution, respectively, with corresponding absolute terminal-performance improvements of

15.1\%

and

31.0\%

Abstract

Explore similar work

Auditing Emergent LLM-Agent Collaboration through Cooperation-Obligation Coupling

Same Winners, Different Success Rates: Evaluating How LLM Agents Recover from Failures

How Strongly Should Task State Influence an LLM Agent?