# Experiment-run dataset

Subset/configuration: `experiment_runs`. Split: `test`. File:
`experiment_runs.jsonl`. Five rows summarize the five recorded runs of the exact
finite-family AXIOMESH experiments. Values are extracted from the unchanged
`results/results.json` and `results/replications/<seed>/results.json` artifacts.

The integer seeds `20261007` through `20261011` are RNG labels, not run dates.
Each row summarizes a run with many correlated tests; it is not an independent
real-world task, training example, or general-intelligence result. The architecture
and strong pooled learner tie the principal prediction benchmark.

| Field | Type | Meaning |
|---|---|---|
| `seed` | integer | RNG seed in the source run |
| `core_version` | string | Scientific implementation version |
| `learned_test_words` | integer | Held-out words in the supplied learned family |
| `weave_correct` | integer | Correct proposed-pipeline outputs for those words |
| `pooled_baseline_correct` | integer | Correct strong conventional pooled outputs |
| `pooled_baseline_tie` | boolean | Equality of the two correct-output counts |
| `erasure_trials` | integer | Total trials over the source erasure configurations |
| `false_erasure_certificates` | integer | Incorrect exact recovery certificates in those trials |
| `identifiability_trials` | integer | Exhaustive finite-fiber checks |
| `incorrect_identifiability_certificates` | integer | Incorrect certificates in the finite-fiber checks |
| `frozen_closure_test_words` | integer | Sum of words in frozen-closure checks |
| `recursive_fracture_test_words` | integer | Words checked after recursive reconstruction |
| `recursive_fracture_correct` | integer | Correct outputs in the recursive-fracture check |
| `source_result_file` | string | Repository-relative provenance path |
| `evidence_scope` | string | Scope reminder for each row |

The erasure trials include honest abstentions when insufficient symbols survive;
zero false certificates does not mean that every erased state was recovered.
Timing data and large program structures remain in the source artifacts rather
than being treated as comparable tabular measurements.

## Load locally

```bash
python -m pip install -r requirements-dataset.txt
```

From the repository root:

```python
from datasets import load_dataset

runs = load_dataset("json", data_files={"test": "data/experiment_runs.jsonl"}, split="test")
assert len(runs) == 5
```

After dataset publication, use the repository ID and explicit configuration as
shown in the root dataset card. No custom dataset loading script is required.

This is synthetic mathematical evaluation evidence produced by the bundled
reference implementation, not human training data. It covers supplied finite
operator families and does not establish neural RSI or unrestricted reconstruction.
The repository owner's existing unselected reuse-license status is preserved.