| --- |
| license: mit |
| pretty_name: Digbench Runs |
| --- |
| |
| # Digbench-Runs |
|
|
| Run traces for the EWM harness (schema world-model harness) and a basic harness on DiG-bench, |
| model `gpt-6-astra` via Codex, 21 games (P-1 … P-21) at three reasoning levels (medium, high, max). |
|
|
| ## Layout |
|
|
| - `index.csv` — one row per (game, reasoning level, harness). |
| Columns: `game, tier, reasoning, harness, levels_beaten, max_level, won, steps, path, source`. |
| `path` points into this dataset; `source` is the original local path the run came from. |
| - `ewm/<reasoning>/P-<n>/` — one EWM run per directory: |
| `events.jsonl` (full event trace), `run.json` (run config), `notes.md`, `sessions/`, `snapshots/`, |
| plus the world-model / planner code the agent wrote during the run. |
| `git.tar.gz` holds the run's own git repository (one commit per turn, tags `start` and `cleared_level_<k>`). |
| Restore it with `cd ewm/<reasoning>/P-<n> && tar xzf git.tar.gz`; the working tree then matches the |
| files in this directory (`events.jsonl` carries the final `run_finished` line as an uncommitted change, |
| as the harness leaves it). |
| - `basic/<reasoning>/P-<n>.jsonl` — the basic-harness run for the same game and reasoning level. |
|
|
| ## events.jsonl |
|
|
| One JSON object per line with a `kind` field: `run_started`, `turn_started` (observation, level, lives, |
| steps remaining, `surprise`), `thinking_delta`, `tool_started` (`name`, `args`, `call_id`), |
| `tool_finished` (`output`, `is_error`, `call_id`), `turn_committed` (`plan`, `reason`, `usage`), |
| `turn_fallback`, `action_taken` (`action`, `observation`, `level`, `levels_beaten`, `level_up`, `win`, …), |
| `model_mispredicted` (`predicted` vs `actual`, `surprise`), `run_finished`. |
| Pair `tool_started` / `tool_finished` by `call_id`. |
|
|
