File size: 1,020 Bytes
b20ddef
 
 
 
1
2
3
4
- Atlas read: 36 parallel attention/MLP layers, width 2048; edit writer columns only, preserve template, no embeddings/final norm edits. Reference is pre-safety base, not a better checkpoint. Reserve request IDs across all languages for validation.
- Localization restricted to 20 English originals (96 translations); remaining 84 originals reserved from localization. Visible mharm has incomplete translation coverage, so pair each translation with its exact English original.
- Paired activation deficit grows in layers 20–26, strongest jv/bn/sw; English direction also underestimates safe Korean refusals. Four inspected primary-only flags include legitimate legal information, so do not target those labels alone.
- Custom multilingual refusal detector is invalid here: reports 0 even on clear refusals. Use task_eval bilingual judges, not English 'I' margin. CircuitKIT discovery OOM; no evidence taken from failed discovery. Steering +8 at layer24 raises diagnostic safe refusals; reject that broad intervention.