# V3 Verifier Ablation: 70-Row Comparison

Primary evidence uses one matched replicate over ten cases. The earlier 70-cell run is secondary stability calibration.

## Conditions

| Condition | Cells | Gate mean +/- sigma | >=4 | Verify calls | Mean minutes |
| --- | ---: | ---: | ---: | ---: | ---: |
| L1 Direct | 10 | 2.226 +/- 0.383 | 0/10 | 0 | 1.2 |
| L2 Base | 10 | 2.127 +/- 0.432 | 0/10 | 0 | 1.8 |
| L2 Scoped Verifier | 10 | 2.252 +/- 0.320 | 0/10 | 20 | 3.5 |
| L2 Detail Verifier | 10 | 2.186 +/- 0.226 | 0/10 | 43 | 10.3 |
| L3 Full v-Harness | 10 | 3.845 +/- 0.685 | 6/10 | 0 | 119.1 |
| L3 Hidden Gate Feedback | 10 | 3.409 +/- 0.611 | 1/10 | 0 | 146.0 |
| L3 No Self-Check Tools | 10 | 3.677 +/- 0.584 | 2/10 | 0 | 79.3 |

## Paired Contrasts

| Contrast | Mean delta | 95% bootstrap CI | W/T/L |
| --- | ---: | ---: | ---: |
| Scoped - Base | +0.125 | [-0.057, +0.356] | 4/3/3 |
| Detail - Base | +0.059 | [-0.110, +0.249] | 4/2/4 |
| Detail - Scoped | -0.066 | [-0.145, +0.008] | 2/2/6 |
| Full - Hidden | +0.436 | [-0.119, +1.026] | 6/0/4 |
| Full - No Self-Check | +0.168 | [-0.342, +0.749] | 6/0/4 |

Matched controls fall inside the old per-case r1-r3 range for 4/20 control/case rows.