# E72-L — Benchmark on held-out, externally sourced tasks

Small samples with Wilson 95% intervals (`artifacts/e72/benchmark.json`). No aggregate "intelligence" score.

| Family | n | Correct | Accuracy |
|---|---:|---:|---:|
| Repository defect (detector) | 5 | 5 | 1.00 |
| Test generation (detector-authored) | 5 | 5 | 1.00 |
| Data analysis (deterministic recompute) | 5 | 5 | 1.00 |
| Memory retrieval (deterministic) | 20 | 20 | 1.00 |
| Planning (Qwen2.5-0.5B) | 5 | 5 | 1.00 |
| Citation extraction (Qwen, real arXiv) | 5 | 4 | 0.80 |
| Arithmetic (GSM8K test split, MIT) | 10 | 1 | 0.10 |
| Data analysis model-only (Qwen) | 5 | 1 | 0.20 |

Local model total: **11/25 exact (0.44)**. Mean E71 keyword score on the same outputs: 0.38. Keyword coverage is not
completion and does not track it; it is reported only to show what the old metric would have said.

Sources: `openai/grade-school-math` test split (MIT), arXiv Atom API, public Iris CSV. Sample sizes are small; the
benchmark is not a general-intelligence claim and was not independently administered.