When Unsupported Claims Become Working State: Measuring Ratification in Sequential Agents
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
The same unsupported claim can be challenged and disappear in one agent trace, yet be accepted as a premise and redirect retrieval or action in another. We call the latter transition ratification and represent it as a claim–consumer state change, with separate injection, endorsement, and operational-reuse events and event-specific risk sets. This representation supports matched policy tests and EndorseGuard, a plan selector whose epistemic critic predicts whether a candidate will carry unsupported content across the state-admission boundary. On 200 held-out Persuasion-Conflict scenarios, assertion-time and adoption-time guards yield 8.2% and 4.6% conditional propagation, respectively, under identical planner, detector, feedback, proposal, and refinement settings. The ordering survives blinded human adjudication (a 3.6-point gap), a component-disjoint DeBERTa/BGE evaluator (3.5 points), and direct annotation of all 500 TopiOCQA test dialogues (6.1% versus 4.1%). Under matched planner-call budgets, EndorseGuard reduces Persuasion-Conflict propagation from MARCH's 4.1% to 2.1% while improving success from 78.4% to 81.0%; on TopiOCQA it raises answer F1 from Utility-only's 55.6 to 60.4, with the largest propagation reduction at topic switches (9.8% to 3.7%). Treating state admission as an explicit reliability event therefore reveals consequential contamination that response-local factuality misses and supplies a concrete control target for sequential agents.
- Primary area:
- foundation or frontier models, including LLMs