Injection and reversal behaviour of the deterministic baseline
MeasuredEV-004
Produced by running code in this repository and preserved as an artifact.
- Scope
- The four embedded-instruction cases and the four decision-reversal cases across both splits.
- Observation
- The baseline produced no item derived from an embedded instruction and never presented a reversed decision as current, because it reports every tagged decision as unresolved and never reads untagged text.
- Limitations
- This is safety by inability, not by judgment. A method that reads nothing cannot be tricked into reading something wrongly, and the same property is why its recall is 0.24.
- Source
experiments/runs/tagged-transcript-baseline--held-out--2026-09-07.1/evaluation.json- Source date
- 2026-09-07
- Retrieved
- 2026-09-07
- Reviewer status
- Pending review. Written down, not yet checked by a person.