EOV Receipt Benchmark: Sub-Millisecond Overhead, Linear Chain Scaling

Setup

The EOV I-D proposes a signed receipt at each action boundary. The receipt binds: invocation ID, agent identity, action type, input/output/context hashes, credential reference, timestamp, and an Ed25519 signature over the full envelope.

• Harness: receipt.py
• Chain benchmark: chain_attestation.py
• N=200 per latency measurement; 100 trials per chain depth

Generation latency

Median receipt generation is 67µs. P99 spikes to ~492µs, consistent with garbage collection at cold runs. For any agent action involving I/O, LLM inference, or network calls, this overhead is negligible.

Verification latency

Ed25519 verification runs ~168µs mean — roughly 2× generation cost, as expected for asymmetric verify vs sign. Still negligible next to any real agent action.

Payload size

516 bytes per receipt. At 10,000 receipts per minute, that is ~5 MB/min — manageable with standard log compression and entirely reasonable for auditability of a production agent pipeline.

Chain scaling (delegation depth 1–5)

Chain scaling is linear at ~190µs per hop with a 1.01× ratio — no observable super-linear overhead up to depth 5. Multi-hop delegation chains remain auditable in under 1ms total.

Published to Zenodo: 10.5281/zenodo.19423545.

Conclusion

The overhead objection to per-action execution receipts does not hold for typical agentic workloads. Ed25519 signing at 67µs median and linear chain scaling at ~190µs/hop make EOV receipts practical at the granularity the I-D proposes.

The remaining engineering question is not cost — it is the label registry and the COSE encoding path, which is covered in the COSE vs JOSE benchmark.

View all EOV experiments on GitHub →