Reference: Qwen2.5-1.5B (exact lineage). Probes: 199. Candidate pool: r1_orig, qwen3_32b, gemini, r1, gpt_oss_120b, llama33_70b.
Best reasoning-family mean alignment: +0.159 (r1) vs gate +0.05. Passed. Matched control: -0.005.
Family r1 wins 65% of probes (chance 17%); anytime-valid evidence log10 E = 48.88, threshold crossed at probe 23; style probability mass on this family 0.75. Passed.
Members in pool: r1_orig.
Likelihood: one member in pool.
Style member probabilities: {'r1_orig': 0.174, 'r1_openr1': 0.575}.
Note: one member in the likelihood pool; within-family attribution by style only (pipeline/source sub-styles)
Candidate pool: r1_orig, qwen3_32b, gemini, r1, gpt_oss_120b, llama33_70b; style classifier pool: r1_orig, r1_openr1, qwen3_32b, gemini, r1, gpt_oss_120b, llama33_70b. A strong preference for a wrong near-relative is what a missing true teacher looks like; the pool must include the suspect's plausible teachers and their pipelines.