dumb-n-honest
v0.2.11
Published
Private local benchmark of explicit correction acknowledgments in your Claude Code and Codex history. Downloads and SHA256-verifies the official dumb-n-honest release, then runs it with your local Python 3.10+. No transcript content leaves your machine.
Maintainers
Readme
dumb-n-honest (npx)
Run the dumb-n-honest private local benchmark with one command:
npx dumb-n-honest
npx dumb-n-honest doctor
npx dumb-n-honest run --output-dir ./my-audit --no-pngWith no arguments, the launcher runs the full audit with both providers and both language packs. It
writes to a new timestamped dumb-n-honest-output-* directory in the current working directory, so
existing results are never overwritten.
Open report.html for the chart, editable draft, provider/family/model filters and PNG/SVG downloads.
Older generations are hidden by default. The large number is correction acknowledgments per 100
answered turns; usage is labelled separately. A ready-to-share chart.png is exported when
rsvg-convert is installed, or can be downloaded from the report in your browser. Pass
--export-share-pack for separate poster, draft and alt-text files.
This package is a thin launcher with no npm dependencies. It:
- finds a local Python 3.10+ interpreter,
- downloads the official
dumb-n-honestrelease from GitHub Releases, - verifies its SHA256 against a hash pinned in this package before every run,
- extracts the verified archive to a fresh temporary directory,
- runs the original Python scripts and removes the temporary directory afterward.
The launcher uses the network only to fetch a missing release archive. It never reads transcripts; all parsing stays in the verified Python release. The audit itself runs offline, writes only aggregates, and never publishes anything automatically. Requires Node 18+ (for the launcher only) and Python 3.10+.
DUMB_N_HONEST_VERSION can select only a release whose SHA256 is pinned by the installed launcher.
