Grand Diomande Research · Full HTML Reader

Edit-Candidate Ranker Report

| condition | CER | delta pp | changed | better/same/worse | |---|---:|---:|---:|---:| | baseline | 0.3421 | +0.00 | 0 | 0/0/0 | | oracle_any | 0.2832 | -5.89 | 274 | 274/0/0 | | ranker | 0.2957 | -4.63 | 274 | 255/19/0 | | ranker_preserve | 0.3199 | -2.22 | 140 | 124/16/0 |

Language as Infrastructure experiment experiment writeup candidate score 32 .md

Full Public Reader

Edit-Candidate Ranker Report

Headline

Train/valid/test rows: 828/276/277
Validation-tuned threshold: 0.6500

Held-Out Test Conditions

condition	CER	delta pp	changed	better/same/worse
baseline	0.3421	+0.00	0	0/0/0
oracle_any	0.2832	-5.89	274	274/0/0
ranker	0.2957	-4.63	274	255/19/0
ranker_preserve	0.3199	-2.22	140	124/16/0

Candidate Classifier Metrics

Validation AUC: 0.9305039335757725
Validation AP: 0.8784072602172582
Test AUC: 0.9391017940743587
Test AP: 0.8967641446536138

Interpretation

This is a held-out test of the deterministic candidate-ranker architecture. The oracle remains non-deployable because it uses references. The ranker uses only candidate features.

Artifacts: `experiments/acoustic_gate/overnight/edit_candidate_ranker_v1/ranker_report.json`, `experiments/acoustic_gate/overnight/edit_candidate_ranker_v1/test_decisions.jsonl`.

Promotion Decision

Attach run IDs, datasets, metrics, and reproduction commands.

Source Anchor

nko-brain-scanner/experiments/acoustic_gate/overnight/edit_candidate_ranker_v1/RANKER-REPORT.md

Detected Structure

Method · Evaluation · References · Code Anchors · Architecture