# ERT evidence-grounding weight sweep

This sweep asks whether the objective-level failure is fixed by strengthening the evidence-grounding term. Lower drift is not sufficient; belief error and action coverage remain required.

| weight | split | drift | belief error | evidence error |
|---:|---|---:|---:|---:|
| 0.25 | in_domain | 0.059 | 0.357 | 0.316 |
| 0.25 | shifted | 0.050 | 0.405 | 0.375 |
| 0.25 | adversarial | 0.049 | 0.371 | 0.345 |
| 1 | in_domain | 0.004 | 0.459 | 0.463 |
| 1 | shifted | 0.004 | 0.484 | 0.488 |
| 1 | adversarial | 0.004 | 0.447 | 0.451 |
| 4 | in_domain | 0.229 | 0.176 | 0.207 |
| 4 | shifted | 0.229 | 0.322 | 0.314 |
| 4 | adversarial | 0.217 | 0.307 | 0.325 |

A higher evidence weight may improve calibration, but that is a tuning result, not proof of internalised ERT. The test also needs useful actions and multiple seeds.