H4 first cut: probe and generation budgets

From existing Phase 0 / H3 data by subsampling (400 random draws per size), no new compute. Script budget.py, output results/budget_S.json. Student S = DeepSeek-R1-Distill-Qwen-1.5B, reference = Qwen2.5-Math-1.5B, 2048 tokens, five teachers.

Likelihood attribution (paper's statistic)

Probes"true-teacher style" pair {r1_orig, qwen3} wins vs chance 2/5: median log10 pP(p < 1e-6)r1_orig alone vs chance 1/5: median log10 pP(p < 1e-3)
5-2.00.00-0.60.02
10-4.00.00-1.50.11
20-8.01.00-2.00.29
40-15.91.00-3.50.76
60-23.91.00-5.50.95
80-31.81.00-7.00.99
160-63.71.00-13.11.00

The pair wins on 199/199 probes, so "this student learned the R1-family reasoning style" is provable at p < 1e-6 with 20 probes. Naming the single true teacher by likelihood alone needs 60 to 80 probes for p < 1e-3, entirely because of the near-relative tie.

Style classifier on the student's own generations (H3)

GenerationsP(mean p(r1_orig) - p(qwen3) > 0.20)5th percentile of the margin
10.90-0.63
30.97+0.32
50.99+0.43
81.00+0.54
201.00+0.64

Five to eight generations from the suspect model settle the near-relative question that 160 likelihood probes cannot. Practical consequence: a provider-side test can be cheap, black-box (samples only, no logprobs) for the style stage, and reserve the likelihood stage for checkpoint-level attribution.

Caveat: budgets measured on one student; the Phase 1 testbed must confirm they hold across students and teachers.