The Label Survived. Which Case Did the Agent Check?

Provenance labels can remain intact during an agent handoff without determining what the receiving agent verifies. We test whether a model uses report-level scope to allocate one priced inspection. In matched synthetic worlds, two archived reports retain their facts, identifiers, order and payoffs while their tested-configuration labels exchange places. An analytic expected-loss oracle identifies the optimal action. The frozen main experiment evaluated DeepSeek V4.1 Flash through OpenRouter/PrimeIntellect in 704 cells: 448 decision episodes and 256 independent-context readbacks across 64 worlds, with payoff and information controls on 16 worlds. The mean paired allocation effect was 0.9375 in separate layout (95% world-bootstrap interval 0.875 to 0.984375) and 0.921875 inline (0.859375 to 0.984375); the oracle reference is 1. Readback was correct in every branch. The results meet the predeclared, bounded "no gap here" criterion, but do not imply perfect action: fully correct twin pairs numbered 57/64 and 55/64, and high-stakes optimality was 52/64. Sixteen invalid continuations and one missing continuation remain in the evidence. Withheld-scope interpretation is limited by first-reply truncation in 21/32 cells. The study contributes an auditable scope-binding intervention and shows strong scope-sensitive verification under explicit reliabilities, alongside residual payoff and execution failures. Research preprint; not peer reviewed. Research conducted during the Apart Research AI x Epistemics Research Sprint, November 2026. Code and data: https://github.com/araf81772/Beyond-the-label/tree/v1 LLM usage: Codex assisted implementation. Claude assisted repository review and staged execution. Experimental outcomes were scored by deterministic code.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-17
DOI
https://doi.org/10.5281/zenodo.24309950
Primary Topic
Ethics and Social Impacts of AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

The Label Survived. Which Case Did the Agent Check?

Abidur Rahaman, Araf Akhter, Tonmoy Sarker
2 links
Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
preprint

The Label Survived. Which Case Did the Agent Check?

Abidur Rahaman, Araf Akhter, Tonmoy Sarker
preprint en

Abstract

Provenance labels can remain intact during an agent handoff without determining what the receiving agent verifies. We test whether a model uses report-level scope to allocate one priced inspection. In matched synthetic worlds, two archived reports retain their facts, identifiers, order and payoffs while their tested-configuration labels exchange places. An analytic expected-loss oracle identifies the optimal action. The frozen main experiment evaluated DeepSeek V4.1 Flash through OpenRouter/PrimeIntellect in 704 cells: 448 decision episodes and 256 independent-context readbacks across 64 worlds, with payoff and information controls on 16 worlds. The mean paired allocation effect was 0.9375 in separate layout (95% world-bootstrap interval 0.875 to 0.984375) and 0.921875 inline (0.859375 to 0.984375); the oracle reference is 1. Readback was correct in every branch. The results meet the predeclared, bounded "no gap here" criterion, but do not imply perfect action: fully correct twin pairs numbered 57/64 and 55/64, and high-stakes optimality was 52/64. Sixteen invalid continuations and one missing continuation remain in the evidence. Withheld-scope interpretation is limited by first-reply truncation in 21/32 cells. The study contributes an auditable scope-binding intervention and shows strong scope-sensitive verification under explicit reliabilities, alongside residual payoff and execution failures. Research preprint; not peer reviewed. Research conducted during the Apart Research AI x Epistemics Research Sprint, November 2026. Code and data: https://github.com/araf81772/Beyond-the-label/tree/v1 LLM usage: Codex assisted implementation. Claude assisted repository review and staged execution. Experimental outcomes were scored by deterministic code.

Zenodo (CERN European Organization for Nuclear Research)
Ethics and Social Impacts of AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon