# Benchmark Summary

## ONNX q8

| Suite | F1 | Examples/s |
|---|---:|---:|
| Irish core | 0.9664 | 29.9676 |
| Edge | 1.0000 | 116.3977 |
| Finance | 1.0000 | 27.2093 |
| Finance boundary | 1.0000 | 64.7358 |
| User PPSN | 0.8571 | 47.0000 |
| GA weak PPSN | 1.0000 | 52.5895 |
| Multilingual PPSN | 0.9591 | 54.2219 |
| Hardening exact | 1.0000 | 64.5117 |
| UAT replay exact | 0.9032 | 31.7159 |

## Full checkpoint

| Suite | F1 | Examples/s |
|---|---:|---:|
| Irish core | 0.9664 | 25.9375 |
| Edge | 1.0000 | 35.0893 |
| Multilingual PPSN | 0.9591 | 60.4876 |
| Hardening exact | 1.0000 | 27.8705 |

## UAT Replay Exact Comparison

| Model | F1 | Precision | Recall | Examples/s |
|---|---:|---:|---:|---:|
| `IrishCore-DiffMask-135M-v1-rc1` q8 | 0.4545 | 1.0000 | 0.2941 | 238.6524 |
| `IrishCore-DiffMask-135M-v1-rc2` q8 | 0.8276 | 1.0000 | 0.7059 | 183.6675 |
| `OpenMed-mLiteClinical-IrishCorePII-135M-v2-rc8` q8 | 0.3636 | 0.3750 | 0.3529 | 110.7595 |
| `IrishCore-DiffMask-135M-v1-rc3` q8 | 0.9032 | 1.0000 | 0.8235 | 31.7159 |

## Notes

- `rc3` keeps the stronger `focusv3` checkpoint and applies a narrower published decoder profile for email continuation and q8 passport recovery.
- The remaining known misses on the UAT replay suite are `071 967 2616`, `R93 EC57` inside a longer centre block, and `EPStamp4@enterprise.gov.ie`.
- `user_raw_regression_cases_v1` is a legacy PPSN-only suite; its counted false positive is `0871234567`, which `rc3` now masks as `PHONE_NUMBER`.