Grand Diomande Research · Full HTML Reader

Edit-Candidate Oracle Report

| condition | CER | delta pp | changed | better/same/worse | |---|---:|---:|---:|---:| | baseline | 0.3514 | +0.00 | 0 | 0/0/0 | | oracle_any | 0.2951 | -5.63 | 1367 | 1367/0/0 | | oracle_preserve | 0.3234 | -2.80 | 669 | 669/0/0 | | acoustic_gate | 0.3497 | -0.18 | 213 | 119/52/42 | | acoustic_preserve_gate | 0.3496 | -0.18 | 194 | 113/47/34 | | acoustic_featural_preserve_gate | 0.3496 | -0.18 | 194 | 113/47/34 |

Language as Infrastructure experiment experiment writeup candidate score 24 .md

Full Public Reader

Edit-Candidate Oracle Report

Headline

Rows: 1381
Baseline CER: 0.3514
Rows with any bounded local-edit improvement: 1367/1381 (99.0

Conditions

conditionCERdelta ppchangedbetter/same/worse
baseline0.3514+0.0000/0/0
oracle_any0.2951-5.6313671367/0/0
oracle_preserve0.3234-2.80669669/0/0
acoustic_gate0.3497-0.18213119/52/42
acoustic_preserve_gate0.3496-0.18194113/47/34
acoustic_featural_preserve_gate0.3496-0.18194113/47/34

Acoustic Selection

Acoustic-selected improvements: 119/1381 (8.6
Hit rate on rows where the candidate set had a true improvement: 8.7

Interpretation

The oracle rows are not deployable because they use clean references to choose the best candidate. They measure ceiling only. The acoustic rows are the deployable selection test.

Artifacts: `experiments/acoustic_gate/overnight/edit_candidate_oracle_v1/oracle_report.json`, `experiments/acoustic_gate/overnight/edit_candidate_oracle_v1/oracle_cases.jsonl`.

Promotion Decision

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

Source Anchor

nko-brain-scanner/experiments/acoustic_gate/overnight/edit_candidate_oracle_v1/ORACLE-REPORT.md

Detected Structure

Evaluation · References · Code Anchors