MMLU per-subject table (n=2280 40-per-subject)

README.md

@@ -83,19 +83,77 @@ Refusal graded on the **tokens the model actually emits** (content, or the reaso
 | illegal                       | 53 | 90.6% | **0.0%** | 9.4% | **100.0%** |
 | misinformation_disinformation | 54 | 96.3% | **1.9%** | 3.7% | **98.1%** |
 
-### MMLU (mixed-subject sample, n=65, next-token letter-logit)
+### MMLU (n=2,280 = 40 questions × 57 subjects, next-token letter-logit)
 
 | build | acc | Δ |
 |---|---:|---:|
-| Base `PQ2_0` | 30.77% | — |
-| **CRACK `PQ2_0`** | **40.00%** | **+9.23 pp** |
+| Base `PQ2_0` | 39.82% | — |
+| **CRACK `PQ2_0`** | **39.74%** | **-0.08 pp** |
 
-CRACK preserves — and on this ternary build slightly **improves** — general capability. The letter-logit path on 1bit/ternary base carries a mild positional bias toward "A" that the modification appears to reduce.
+CRACK preserves general capability — Δ within noise on the 40-per-subject sample.
 
 <details>
-<summary><b>Per-subject breakdown (sample size limits per-subject significance to indicative deltas)</b></summary>
+<summary><b>Per-subject accuracy (all 57 subjects)</b></summary>
 
-The 65-question mixed sample covers 65 distinct subjects (mostly 1 question each), so per-subject accuracy is directional rather than statistically tight. Overall: 20/65 base → 26/65 CRACK, +9.23 pp. No subject shows a large CRACK regression on this sample; larger evaluations (14k full test) are ongoing.
+| subject | base | CRACK | Δpp | n |
+|---|---:|---:|---:|---:|
+| abstract_algebra | 30.0% | 30.0% | +0.0 | 40 |
+| anatomy | 30.0% | 42.5% | +12.5 | 40 |
+| astronomy | 40.0% | 55.0% | +15.0 | 40 |
+| business_ethics | 37.5% | 50.0% | +12.5 | 40 |
+| clinical_knowledge | 40.0% | 40.0% | +0.0 | 40 |
+| college_biology | 50.0% | 45.0% | -5.0 | 40 |
+| college_chemistry | 25.0% | 35.0% | +10.0 | 40 |
+| college_computer_science | 45.0% | 52.5% | +7.5 | 40 |
+| college_mathematics | 32.5% | 35.0% | +2.5 | 40 |
+| college_medicine | 27.5% | 40.0% | +12.5 | 40 |
+| college_physics | 30.0% | 40.0% | +10.0 | 40 |
+| computer_security | 50.0% | 37.5% | -12.5 | 40 |
+| conceptual_physics | 40.0% | 37.5% | -2.5 | 40 |
+| econometrics | 32.5% | 50.0% | +17.5 | 40 |
+| electrical_engineering | 37.5% | 47.5% | +10.0 | 40 |
+| elementary_mathematics | 45.0% | 52.5% | +7.5 | 40 |
+| formal_logic | 32.5% | 35.0% | +2.5 | 40 |
+| global_facts | 35.0% | 22.5% | -12.5 | 40 |
+| high_school_biology | 35.0% | 15.0% | -20.0 | 40 |
+| high_school_chemistry | 52.5% | 30.0% | -22.5 | 40 |
+| high_school_computer_science | 47.5% | 52.5% | +5.0 | 40 |
+| high_school_european_history | 42.5% | 40.0% | -2.5 | 40 |
+| high_school_geography | 30.0% | 45.0% | +15.0 | 40 |
+| high_school_government_and_politics | 52.5% | 47.5% | -5.0 | 40 |
+| high_school_macroeconomics | 37.5% | 47.5% | +10.0 | 40 |
+| high_school_mathematics | 32.5% | 32.5% | +0.0 | 40 |
+| high_school_microeconomics | 35.0% | 37.5% | +2.5 | 40 |
+| high_school_physics | 22.5% | 40.0% | +17.5 | 40 |
+| high_school_psychology | 45.0% | 40.0% | -5.0 | 40 |
+| high_school_statistics | 42.5% | 40.0% | -2.5 | 40 |
+| high_school_us_history | 55.0% | 40.0% | -15.0 | 40 |
+| high_school_world_history | 50.0% | 37.5% | -12.5 | 40 |
+| human_aging | 45.0% | 42.5% | -2.5 | 40 |
+| human_sexuality | 42.5% | 32.5% | -10.0 | 40 |
+| international_law | 72.5% | 50.0% | -22.5 | 40 |
+| jurisprudence | 40.0% | 45.0% | +5.0 | 40 |
+| logical_fallacies | 37.5% | 47.5% | +10.0 | 40 |
+| machine_learning | 35.0% | 35.0% | +0.0 | 40 |
+| management | 47.5% | 37.5% | -10.0 | 40 |
+| marketing | 42.5% | 57.5% | +15.0 | 40 |
+| medical_genetics | 52.5% | 62.5% | +10.0 | 40 |
+| miscellaneous | 42.5% | 37.5% | -5.0 | 40 |
+| moral_disputes | 40.0% | 22.5% | -17.5 | 40 |
+| moral_scenarios | 17.5% | 27.5% | +10.0 | 40 |
+| nutrition | 42.5% | 55.0% | +12.5 | 40 |
+| philosophy | 42.5% | 35.0% | -7.5 | 40 |
+| prehistory | 35.0% | 27.5% | -7.5 | 40 |
+| professional_accounting | 20.0% | 25.0% | +5.0 | 40 |
+| professional_law | 40.0% | 22.5% | -17.5 | 40 |
+| professional_medicine | 27.5% | 32.5% | +5.0 | 40 |
+| professional_psychology | 47.5% | 32.5% | -15.0 | 40 |
+| public_relations | 17.5% | 22.5% | +5.0 | 40 |
+| security_studies | 47.5% | 40.0% | -7.5 | 40 |
+| sociology | 57.5% | 47.5% | -10.0 | 40 |
+| us_foreign_policy | 50.0% | 45.0% | -5.0 | 40 |
+| virology | 30.0% | 30.0% | +0.0 | 40 |
+| world_religions | 57.5% | 60.0% | +2.5 | 40 |
 
 </details>