# CAIN42 — Effect Benchmarks (Prompt 13)

Measured on the host (Python 3.14, single process). These are engineering measurements, not a product benchmark;
they are not compared to a commercial baseline because none was run in this batch.

| Operation | Notes | Reproduce |
|---|---|---|
| Effect contract issue + verify | Ed25519 sign + canonical digest | `pytest tests/test_ironclad.py` |
| Effect fingerprint compute/compare | hashing + dict comparison | `pytest tests/test_ironclad.py` |
| Two-phase prepare/commit/finalize | in-process | `pytest tests/test_ironclad.py` |
| Clean-room receipt verification | recompute + verify (no runtime) | `python3 scripts/cain42_ironclad/cain_effect_verify.py <receipt>` |
| Fault-injection lab | 100,000 deterministic cases ≈ a few seconds | `python3 -c "from cain.ironclad.lab import run_effect_lab; print(run_effect_lab(100000))"` |
| Hosted pipeline probes | 48 wire probes against the live gateway | `python3 scripts/cain42_continuum/hosted_pipeline_deploy_verify.py` |
| Harness properties | 97/97 in the hosted worker | `python3 scripts/cain42_continuum/hosted_harness_worker.py` |

**No latency improvement over any baseline is claimed.** Authorization/effect latency separates the decision path
(live) from the effect observation path (library), so an end-to-end effect latency is only meaningful once a
customer system is bound.
