# deepswe-glm53-20260908

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

DeepSWE (deep-swe v1.1), 111 tasks, custom agent: baseline K=3 (2026-09-08–09) and scaffold variant v1lite K=6 (2026-09-17–21). `deepswe_sanitizado.json` contains, per run and task, status, infra error, and test counts; tokens, timings, and internal paths were removed (full-file sha256 in `corridas[]`).

The K runs are independent repetitions at T=1.0: the agent discards the seed. `resumo.json` is derived from this package.

Sensitivity: 8 of the 13 baseline exclusions (3, 2 and 3 per repetition) are tasks whose agent history did not fit the 262,144-token context without cutting the instruction. Counting them as failures gives 63.72 ± 2.65% (61.47 · 63.64 · 66.06); for v1lite K=6, 62.80 ± 5.41%. The first baseline repetition mixes two serving setups: 29 tasks at 20 workers without MTP (1,800 s timeout) and 82 at 8 workers with MTP k=3 (3,600 s).

| file | what it is | sha256 (first 16) |
|---|---|---|
| `deepswe_sanitizado.json` | items per question or task, without question or response text; file lists what was removed | `3b0a0b9cceb86cbb` |
| `resumo.json` | means, standard deviations, and window result (K repetitions per arm) | `7842c580bd591552` |