FinalityBench: an effect-level benchmark for agent decisions under delayed and conflicting financial finality
Year: 2026
Type: software
Abstract: An executable benchmark in which a hidden canonical event log is projected into four operational systems - payment processor, double-entry ledger, ERP and bank feed - through independently faulted delivery streams, so that the systems temporarily disagree about the same order. Agents are graded on executed monetary effects rather than stated conclusions, and the corpus includes verified paired counterfactuals: task pairs whose observations are identical at the decision instant and whose correct actions are opposite.
Author: Sharma, Abhishek
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7206212697