Causal patching of context vectors across framings (chat-to-plain, assistant-to-story)
kind: experimentparent: #2094
track: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
Test whether the framing-transfer results (#2054 chat-plain, #1345/#1639 assistant-story) are causal: patch the context vector v_C across framings and measure persona/behavior transport into the target framing's generated answer vs matched within-framing patches and null controls.
Context
- Paper need: C4 (PSM section) of the context→answer mapping paper (
~/overleaf-6a59c927/plan.md;claims.mdGap 1). C4 is currently correlational-only; these two arms are what upgrade it to causal. - Method inheritance: #2094's null-separated patching grid (context-end patches, fraction-of-full-swap DV, bootstrap screen + independent temperature-1.0 confirmation). #2333's prefill-vs-patch decomposition (opening-token share of the patch effect) qualifies interpretation and should be carried as a control.
- Two arms: (a) chat-template ↔ no-template ("User:/Assistant:" plain form) patches; (b) assistant ↔ story-character patches on the constructed rigs from #1345/#1639.
- Deadline context: ICLR 2027 abstracts 2026-09-18.
Scheduling
NOT SCHEDULED — filed per user directive 2026-08-19 ("file these but don't run them right away"). Do not auto-run; user/PM dispatches.
Activity