| --- |
| tags: [agentic-ptb, sol-max, record] |
| --- |
| |
| # sol-max — run record |
|
|
| Non-weight artifacts for AgentPTB cell `sol-max` (**Codex / gpt-5.6-sol @ effort `max`**). |
| Checkpoints for this cell are separate repos named `agentic-ptb/sol-max.h*`; the index is |
| [`agentic-ptb/INDEX`](https://huggingface.co/datasets/agentic-ptb/INDEX). |
|
|
| | contents | | |
| |---|---| |
| | `driver-session/` | **full driver trajectory** — every event the agent emitted (179 files) | |
| | `harness/` | the harness the cell proposed: skills, system prompt, submission manifest, configs, scripts | |
| | `RUNLOG.md` | the arm's own narrative of what it tried and concluded | |
| | `WORKING_STATE.md`, `STATE.md` | its persisted state — this, not the chat history, was its memory | |
| | `supervisor.jsonl` | cycle boundaries, resumes, boots | |
| | `evals/` | eval logs for every panel it ran | |
|
|
| ## Reading the trajectory |
|
|
| The driver ran 100 hours under a supervisor that restarted it each cycle. Cycle boundaries are |
| in `supervisor.jsonl`; `resume:false` means that cycle began a fresh thread and the agent |
| re-derived its situation from `RUNLOG.md`. |
|
|
| ## Caveat on self-reported numbers |
|
|
| Scores in `RUNLOG.md` are the arm's own, measured with **its own harness on its own sample** — |
| sample sizes differ by an order of magnitude across cells, so they are not directly comparable. |
| Use them as the cell's account of itself, not as a cross-cell ranking. |
|
|
