Cross-model transfer: does base-geometry prediction of post-inoculation re-elicitation replicate on Llama-3.1-8B-Instruct?
Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.
No comments yet.
Goal
Replicate the inoculated-FT -> re-elicitation -> base-geometry-predictor pipeline on Llama-3.1-8B-Instruct: does the misalignment positive (Qwen pooled rho +0.708) and the capitalization sign flip (-0.747) transfer beyond the Qwen family, or are they family-specific representation quirks?
Context
Filed from #2474's park-time follow-up proposals (proposal 3, question_relation: substantially-different — a fresh Goal, so it cannot fold into the parent). Screened not-redundant by both follow-up critics (2026-08-23): no second-family replication of this pipeline exists in the corpus (ruled out: #544, #1426, #1336, #2502, #697, #2389).
Estimated 32 GPU-h. auto_run: no — model pick, layer selection, and map re-derivation for the new family need human scoping before planning (the parent's layer pins 16/27 and the #779 LMSYS map are Qwen-specific and must be re-derived, not reused).
Full proposal spec: the epm:follow-ups v1 marker on #2474 (2026-08-23T19:15Z), proposal 3.