Training against reward hacking
The model is trained to do the task as meant, not fake tests, results or reports.
Layer L1 Model · Owner Model provider · Family Model safeguards
No incident we have read names this control yet (45 of 780 incidents read for controls).
Frameworks
| Standard | Clause | Requirement |
|---|---|---|
| Research (arXiv) | 2406.10162 | To verify when the text is stored |