HumanEval: replace syntactic 90% with real pass@1 96.3% (158/164, full set, subprocess-exec)

README.md

@@ -83,10 +83,12 @@ Beyond the synthetic decode bench, a 15-prompt real-world suite (chat, coding, r
 | Benchmark | Sample | Score |
 |---|---:|---:|
 | **GSM8K** (T=0, COT-style, exact-match on `#### N`) | **100** | **93.0% (93/100)** |
+| **HumanEval pass@1** (T=0, greedy, subprocess-exec tests) | **164** (full) | **96.3% (158/164)** |
 | **MMLU** (T=0, mixed subjects, max_tokens=128 to leave room for reasoning) | **100** | **53.0% (53/100)** ¹ |
-| **HumanEval** (syntactic check; not pass@1 unit-test exec) | **50** | **90.0% (45/50)** |
 | Internal capability eval (math/code/reasoning/knowledge/instruction/longform/tools) | 14 | 13/14 (92.9%) |
 
+The HumanEval result is the **full 164-problem set with real unit-test execution** (greedy T=0, content-only extraction, 60 s test timeout, sandboxed subprocess). Of the 6 failures: 3 were `NameError` (model named the function differently from the prompt's `entry_point`), 2 were `AssertionError` (real wrong answers), 1 was `AttributeError` (used `list.add` instead of `list.append`). Total wall time for the full run: 11.3 min at concurrency 2 on this hardware.
+
 ¹ The MMLU sample mixed all 57 subjects uniformly (incl. hard categories like `formal_logic`, `college_*`, `professional_law`). The model's main forward path is identical to `pastapaul/DeepSeek-V4-Flash-W4A16-FP8`, so headline numbers track that base — refer to base-model evals for tight numbers across full leaderboards.
 
 ## How to run (production canonical, on this hardware family)