Re-eval lc_long at max_new_tokens=2048 to test eval-truncation hypothesis

Parent: #260 (clean-result #295)

Goal

Re-eval lc_long merged model at max_new_tokens=2048 to test the eval-truncation hypothesis from #260's clean-result.

Hypothesis

lc_long was trained on ~1050-token completions with [ZLT] at the end. The original eval used max_new_tokens=512, cutting off ~360 tokens before the trained marker position. The 0.00 source-rate is most likely an eval-truncation artifact, not a real implantation failure.

H_truncation: at max_new_tokens=2048, lc_long source-rate will be substantially > 0 (predicted: 0.20–0.40 based on lc_short=0.29 / lc_medium=0.31 trend before the cliff).

Setup

• Model: superkaiba1/explore-persona-space/leakage_experiment/marker_librarian_asst_excluded_medium_seed42_lc_long (already on HF Hub from #260)
• Eval: 11 personas × 20 EVAL_QUESTIONS × 5 completions, T=1.0 top_p=0.95, vLLM batched
• Marker: [ZLT] case-insensitive substring (eval/trait_scorers.py::evaluate_markers)
• Only change vs #260's lc_long Leg-1 eval: max_new_tokens 512 → 2048

Success / Kill criterion

If source-rate ≥ 0.10 with non-zero bystander rates: H_truncation supported, lc_long REJECTED verdict in #295 should be downgraded to "REJECTED at gate due to eval truncation."

If source-rate stays < 0.05 across all personas: H_truncation rejected, lc_long collapse is real (not a truncation artifact).

Compute

~5 min on 1× H100. Pod epm-issue-260 is still running and has the merged model on HF Hub. No new training.

References

• Parent: #260 — clean-result #295: "Train-time conversation shape — turns, completion length, system-prompt length — fails to monotonically increase marker implantation; (a)/(b) collapse to zero, (c) peaks at parent recipe (LOW confidence)"
• Plan v3 §6 lc_long ceiling-vs-style alternative
• Interp-critic v1 finding 5 (eval-truncation explains lc_long collapse)