Factor panel for behavior-implantation strength: system-prompt vs message length, completion length, on-policy vs off-policy

Cross-cutting followup on the mentor agenda Q1 — "what controls the strength of behavior implantation?" — pulling together axes-to-vary that came up across several useful results.

Axes to include in the panel:

1. System prompt vs message length (followup on #337). #337 shows longer persona system prompts make markers more persona-localized. Is the effect specifically about system-prompt length, or just about total token count of distribution-shifting text? Run matched-token-count conditions with the same text placed in the system prompt vs. in the message.

2. Completion length effects + a predictive metric (followup on #295). Ask questions with naturally different completion lengths (rather than artificially stretched ones, as in #295 which didn't amplify uptake). Look for a quantity-of-x that predicts implantation/leakage rate. Candidate: a divergence (KL or similar) between persona-conditioned and base-conditioned next-token distributions, summed across all token positions ( t ). If summed divergence predicts leakage rate, we have a tractable per-input proxy for "how strongly will this implant?"

3. On-policy vs off-policy training data (from mentor agenda card). Whether implantation strength depends on whether training completions are sampled from the model itself vs. a fixed off-policy corpus is currently uncharacterized.

Protocol. Standard 11-persona × 20-question × 5-completion setup on Qwen2.5-7B-Instruct (matches the agenda's Q1 plan).

Source comments (mentor update, 2026-05-11):

System prompt vs message length (from #337)

Ask questions with naturally different completion lengths

• is there a quantity such that that quantity going up predicts implantation/leakage rate going up
• some total divergence summed across all t (from #295)

On-policy vs Off-policy other factor to vary (from mentor agenda card "Questions / next steps")

From mentor update — cross-cutting on #295, #337, and the mentor agenda Q1 panel.