{"paper":{"title":"Ghost in the Context: Measuring Policy-Carriage Failures in Decision-Time Assembly","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Decision-time context assembly in LM agents is a measurable control-path element that can be partially hardened.","cross_cats":[],"primary_cat":"cs.CR","authors_text":"Igor Santos-Grueiro","submitted_at":"2026-05-02T18:07:42Z","abstract_excerpt":"LM agents do not act on raw interaction history; they act on a bounded decision state assembled by truncation, summarization, reordering, and rewriting. If directive-bearing state is dropped, weakened, or rebound during that step, an agent can cross a policy boundary without prompt override, model changes, or persistent-memory compromise. We study this failure mode over local Llama 3.1 8B, Qwen 2.5 7B, and Mistral 7B using judged exact constraint respect and direct audits of assembled-state visibility. We evaluate SafeContext, a control layer that pins control state, reuses retained control pr"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Decision-time context assembly is therefore a measurable part of the control path that can be partially hardened.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the specific context assembly policies tested (truncation, structured-compaction) are representative of those used in practical LM agent deployments, and that the constraint respect judgments accurately reflect policy carriage without evaluator bias.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Policy directives can be lost during context assembly in language model agents, leading to unprompted policy violations that SafeContext can partially prevent.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Decision-time context assembly in LM agents is a measurable control-path element that can be partially hardened.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"35b3b9aa4f594812f10590e8696064342c284f7095c5f2c43515d6a2155e1d73"},"source":{"id":"2605.12535","kind":"arxiv","version":1},"verdict":{"id":"403eb7ca-73eb-4d41-aeaf-1ce9e349e4e7","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T21:20:28.020793Z","strongest_claim":"Decision-time context assembly is therefore a measurable part of the control path that can be partially hardened.","one_line_summary":"Policy directives can be lost during context assembly in language model agents, leading to unprompted policy violations that SafeContext can partially prevent.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the specific context assembly policies tested (truncation, structured-compaction) are representative of those used in practical LM agent deployments, and that the constraint respect judgments accurately reflect policy carriage without evaluator bias.","pith_extraction_headline":"Decision-time context assembly in LM agents is a measurable control-path element that can be partially hardened."},"references":{"count":44,"sample":[{"doi":"10.18653/v1/2024.acl-","year":2024,"title":"Chen Qian, Wei Liu, Hongzhang Liu, Nuo Chen, Yufan Dang, Jiahao Li, Cheng Yang, Weize Chen, Yusheng Su, Xin Cong, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun","work_id":"f30bdb61-5994-46dd-a4d7-2d320a1917e0","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.18653/v1/2025.naacl-long.605","year":2025,"title":"Gormley, and Graham Neubig","work_id":"14e588c9-40f2-46ac-a49a-0d71e440562a","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.52202/079017-2636","year":2024,"title":"Agentdojo: A dynamic environment to evaluate prompt injection attacks and defenses for llm agents","work_id":"673a54cd-ec69-4997-9c2b-5d11a1480af6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Shen Dong, Shaochen Xu, Pengfei He, Yige Li, Jiliang Tang, Tianming Liu, Hui Liu, and Zhen Xiang. 2025. Memory Injection Attacks on LLM Agents via Query- Only Interaction. InAdvances in Neural Informa","work_id":"ea073470-cbe5-4f2b-ba41-f8c012ca799f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Cheng-Ping Hsieh, Simeng Sun, Samuel Kriman, Shantanu Acharya, Dima Rekesh, Fei Jia, Yang Zhang, and Boris Ginsburg. 2024. RULER: What’s the Real Context Size of Your Long-Context Language Models?. In","work_id":"ae761e92-66be-4a61-80e0-be3d65641cf0","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":44,"snapshot_sha256":"168f19816b93f88bdcaf17a0254594a19f1d6e30accb5f8dfaea9d7c59ad89cf","internal_anchors":5},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}