[Proposed] Special-token position ablation (prefix / suffix / middle)
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.
From EXPERIMENT_QUEUE.md, added 2026-04-16
Follow-up to leakage v3 + A3b. All marker experiments to date place the special token at one fixed position (typically prepended or appended to the persona's response). Question: does marker position matter for adoption vs containment?
Factorial: marker position ∈ {prefix, suffix, middle-inserted} × contrastive / non-contrastive training × 2 source personas.
Hypothesis: prefix marker has strongest identity-conditioning effect (consistent with in-context learning literature on early-token dominance); suffix is most prone to leakage (token can be "detached" from persona cue); middle is weakest signal.
Falsifier: position-invariant adoption rates.
Compute: ~6 GPU-hours (6 conditions × ~1h each). Reuses leakage infrastructure.
Gate-keeper priority: MEDIUM (narrow but clean question; could sharpen the marker-leakage mechanism story).