hbin0701's picture
OPSD LoRA checkpoints + variant configs + manifest (public)
b6d5136 verified
|
Raw
History Blame Contribute Delete
838 Bytes
# P5 · Wrapper edit
*high, conditional · off-axis multiplier*
P1's strategy, but the teacher wrapper drops the inherited anti-copy line — the reactive arm wants the teacher to follow r.
## Delta / mechanism
No code override: the reactive `teacher_user` lives in prompts.yaml and base.build_ctx uses it. Apply BEFORE ranking or it confounds every anchored candidate.
## Pre-registered prediction
If the anti-copy line was present: the largest single drop in off-error KL floor of any change in the set.
## What this folder overrides
`teacher_user` (reactive wrapper) in prompts.yaml. Multiplies all anchored candidates.
## Files
- `utils.py` — the `TeacherVariant` subclass (`VARIANT`)
- `prompts.yaml` — this variant's strategy / system / teacher wrapper
- shared core: `agents/variants/base.py`, transport in `mc_probe.py`