Your Agent Lies About What Tools It Called, and

Abhinaba Basu proposes a straightforward fix for one of the most dangerous failure modes in agentic systems: the agent says it called a tool, but it didn't. NabaOS introduces HMAC-signed tool execution receipts — every tool call returns a cryptographic receipt that the orchestrator can verify. No receipt, no trust. The system detects 91% of tool-call hallucinations with less than 15 milliseconds of overhead per verification.

The paper benchmarks receipts against four alternatives: zero-knowledge proofs, self-consistency voting, RAG-grounded verification, and hybrid approaches. Receipts win on detection rate, latency, and scalability. The epistemic classification framework draws on Nyāya Śāstra — the classical Indian logic tradition — to categorize knowledge claims by their evidentiary basis. It is an unusual theoretical foundation for systems work, and it holds up. arXiv:2603.10060