deepswe-analyze

One command that reproduces the 2026-09-08 DeepSWE analysis on a new week's results. Everything numeric is recomputed from the trials; nothing is copied from a previous week.

Inputs

• repo — the analysis checkout, default /w/seungpil.lee/deepswe (harbor 0.22, .venv, deepswe/ package, refresh.sh).
• week — meeting id YYYY-MM-DD. Data goes to site/weeks/<week>/, prose to docs/weeks/<week>/.
• jobs — harbor job dirs for this week (jobs/<name>/<task>__<trial>/). Trials from other scaffolds or benchmarks are fine as long as collect.py has an adapter for them (see "Adapters").
• Optional: --sample-wins N (judge only N passing trials per model; default all), --crosscheck N (default 30).

Procedure — run every step, in order, and do not stop early