# agentico-glm53-mtp-20260903

Public raw data package for this measurement. Every number on the measurements page citing this window comes from these files.

Two arms of 8 DeepSWE tasks with 8 workers: without speculation (2026-09-03) and with native MTP k=3 (2026-09-04). One run per arm, on different days.

`tarefas.json` is derived from the runner log result lines (agent seconds and turns per task) and runner JSON (status). Per-task token counts were removed: combined with wall-clock time, they would disclose aggregate concurrent throughput, which we do not publish. Full-file sha256 in `tarefas.json`.

`resumo.json` provides the paired task duration ratio and shows why the tasks-per-hour ratio measured by wall-clock time does not hold: wall-clock time is governed by the slowest task, which is the same task in both arms.

| file | what it is | sha256 (first 16) |
|---|---|---|
| `mtp_k3_parede_s.txt` | arm wall-clock time, in seconds | `898a82b86f3661db` |
| `resumo.json` | means, standard deviations, and window result (K repetitions per arm) | `057e2c1954a719f5` |
| `sem_mtp_parede_s.txt` | arm wall-clock time, in seconds | `6306dc76655e87cc` |
| `tarefas.json` | per task: status, turns, and agent seconds for both arms | `aefda0ce149a04b0` |