An agent may see a command dispatch and later narrate successful work even though an interactive process never returned the result it needed. RECEIPT analyzes how the process wait target, command intervention order and tool-run state affect the agent’s ability to distinguish progress from an unresolved action. The tests are on selected web-service and…
Inspect the ClaimsCase · DiggingBeagle record
RECEIPT studied agents that reported completion when commands had not returned
Process and harness state can diverge when interactive operations are interrupted or silently remain pending.
Public disclosure: Nov 2, 2026
Scope
Empirical research on process-return assumptions and tool execution state in autonomous web service tasks.
Assessment profile
Assessment method- Setting
- evaluation
- Exploitation
- demonstrated
- Evidence
- Grade C
- Remediation
- not applicable
- AI role
- AGAINST AI
Basis and provenance
The authors used a command litmus suite and tested harnesses; false return and reduction measurements do not establish a production attack or universal fix. An impact band remains unassigned because the available evidence does not establish a comparable severity score or verified numerical loss for this assessment.
Assessed Nov 8, 2026 using diggingbeagle.assessment/1.
Evidence basis: CLM-RECEIPT-SILENT-COMMAND-RETURN-01 · CLM-ENR-RECEIPT-FALSE-RETURNS · RECEIPT: the return assumption of LLM agents
At a glance
A task summary is not an execution receipt. Persistent process identifiers and checked return conditions are necessary for trustworthy completion reporting.
Read the implicationsThe tests are on selected web-service and agent work environments. They do not demonstrate universal production outages or that every incorrect summary is deliberate deception.
Limits and uncertaintyFull account
An agent may see a command dispatch and later narrate successful work even though an interactive process never returned the result it needed. RECEIPT analyzes how the process wait target, command intervention order and tool-run state affect the agent’s ability to distinguish progress from an unresolved action.
The tests are on selected web-service and agent work environments. They do not demonstrate universal production outages or that every incorrect summary is deliberate deception.
An exit code of zero is not always an execution receipt. In the paper's 158-invocation Litmus set, closed standard input led 11.5% of tested invocations to skip the requested action yet return zero, while an open but unattended terminal could stall. Receipt watches system-call wait states and reports them to the agent; across five tested models it reduced false completion statements by 73.5% relative to closed input. Those results concern the selected benchmark and harnesses, not a universal shell guarantee.
Mechanism and trust boundary
- 01
Invoke a tool command that may not return normally.
- 02
Interrupt or alter the process/session lifecycle.
- 03
Check whether the agent has a real completion receipt before it reports success.
Claims & evidence
CLM-ENR-RECEIPT-FALSE-RETURNSIn the RECEIPT Litmus experiments, 11.5% of tested command invocations returned status zero without completing the requested action under closed stdin; across five models, Receipt reduced false success statements by 73.5% relative to closed input.supported
Basis: reported finding
- supportsRECEIPT: the return assumption of LLM agentsprimary disclosure
Abstract; Section 2 Commands That Do Not Return as Expected; Section 4 Evaluation
CLM-RECEIPT-SILENT-COMMAND-RETURN-01The RECEIPT paper identifies a class of agent workflow failures caused by assuming that an invoked command has completed and returned when process and session state do not support that assumption.supported
Basis: reported finding
- supportsRECEIPT: the return assumption of LLM agentsprimary disclosure
Commands That Do Not Return as Expected; Receipt: Returning the State
Implications
A task summary is not an execution receipt. Persistent process identifiers and checked return conditions are necessary for trustworthy completion reporting.
Unknowns and contradictions
- The tests are on selected web-service and agent work environments. They do not demonstrate universal production outages or that every incorrect summary is deliberate deception.
Sources and citation
Material revision history
- Nov 8, 2026 · Canonical change recorded · new in release · revision 229