| # Experiments | |
| This directory stores compact experiment records. It should make it possible for another Codex thread or a human researcher to answer: | |
| - What was tested? | |
| - Which code commits were used? | |
| - Which exact remote command/config ran? | |
| - What were the key metrics? | |
| - What did we learn? | |
| - What should be tried next? | |
| ## Artifact Rules | |
| Commit: | |
| - `manifest.yaml` | |
| - `summary.md` | |
| - `metrics.csv` | |
| - exact `repobridge.config.json` | |
| - small tokenizer/training configs | |
| Do not commit: | |
| - raw remote output trees | |
| - model checkpoints | |
| - tokenizer binaries | |
| - huge logs | |
| - duplicated historical output folders | |
| Raw outputs may remain in TaoTrain local `results/` or on the remote server. This repo keeps the distilled evidence. | |
| ## Legacy Configs | |
| `legacy_repobridge_configs/` contains exact copies of old RepoBridge root experiment configs. These are intentionally separate from curated run folders: | |
| - curated runs: interpreted, summarized, and often have metrics | |
| - legacy configs: exact historical command/config snapshots that have not all been fully interpreted yet | |
| ## Run Folder Contract | |
| Each run folder should contain: | |
| ```text | |
| manifest.yaml | |
| summary.md | |
| metrics.csv | |
| repobridge.config.json | |
| ``` | |
| If a run failed before producing metrics, keep `manifest.yaml`, `summary.md`, and the config, and mark `status: interrupted` or `status: failed`. | |