Back to Paper
cs.CR
Local ID: 2605.12535v2
AI Summary: gemma4:e4b
Ghost in the Context: Measuring Policy-Carriage Failures in Decision-Time Assembly
By Igor Santos-Grueiro
Revision History Timeline
v15/2/2026
5/2/2026
No submitter comment provided.
v25/19/2026
5/19/2026
No submitter comment provided.
★ Version indexed in Explorerv37/1/2026
7/1/2026
“19 pages”
Comparing v2 vs v3
Green = Added • Red = Removed
Title Comparison
Ghost in the Context: Measuring Policy-Carriage FailuresIntegrity in Decision-TimeLLM AssemblyAgents
Authors Comparison
No author changes.
v2 Comment
No comment for this version.
v3 Comment
“19 pages”
Abstract Word Diff
LLM agents do not act on raw interaction history; they actchoose onactions afrom bounded decision statestates assembled byfrom truncation,system summarization,policy, reordering,runtime andstate, rewriting.tools, Ifworkload directive-bearingcontent, stateand isthe dropped,final weakened,request. orWe reboundstudy duringpolicy-carriage thatintegrity: step,applicable antrusted agentpolicies canmust crossremain apresent, policysound, boundaryand withoutcorrectly promptbound override,in modelthe changes,decision orstate persistent-memoryimmediately compromise.before Weaction. studyUnder thiscontrolled failurepressure modereplay over local Llama 3.1 8B, Qwen 2.5 7B,AutoGen/tau3 and Mistral 7B usingOpenHands/SWE-bench judgedtraces, exactthe constrainttested respectprotected-placement andconfigurations directpreserve auditspolicy ofacross assembled-statethe visibility.pressure Wesweep, evaluatewhile SafeContext,task-local aplacement controlexhibits layereviction, thatweakening, pinsor controlover-budget state,continuation reusesdepending retainedon controlthe prefixes,context andmanager. optionallyWe injectskeep remindersthis underresult pressurestate-level: whilea keepingfixed-assembler modelbehavioral weightscalibration fixed.produced Unmitigated0/90 riskunsafe-action isproposals systematic,and but0/90 absoluteunguarded exactpolicy complianceviolations, remainsso low.policy Againstabsence truncation,alone SafeContextdid yieldsnot smallestablish gains;unsafe againstmodel abehavior. strongThe structured-compactionresulting policy,design mostguidance aggregateis liftsystems-level: disappears,assign leavingtyped residualprovenance benefitto mainlypolicy instate, overflowisolate evictioncontrol andbudget, selectedcheck aliasingbefore slices.assembly Replay-onlythat doesthe notcomplete explainactive thepolicy effect.set Afits, larger-modelfail extensionclosed on Qwen 14Boverload, and Llama 70Benforce showsstructured thepolicies sameat failurethe objectaction underboundary. largerWe models,present althoughControlCapsule signas anda magnitudereference remaindesign policy-conditional.pattern Decision-timefor contextthese assemblyrequirements; isexact thereforeactive-policy areplay measurable+ partpreflight ofremains the control path that can be partiallykey hardened.baseline.