PytxoEvidence
What Pytxo has been measured or observed to do, how, and what each record does not show.
Every number Pytxo publishes comes from a record you can download, with the build it ran on and its limits. A benchmark, a recorded run and a byte measurement answer different questions, so they are kept apart.
| Record | Answers | Does not answer |
|---|---|---|
| First use | How long a new user takes from npm i -g pytxo to a landing decision on a real repository with two live agents' branches, and where they go wrong | How long it takes on other repositories, languages or machines |
| Pybench | What per-branch CI would put on main compared with what pytxo land decides, in known situations; what the gate costs | How often those situations happen in real projects |
| Live runs | Whether a released build completes a real multi-agent job with live vendor CLIs, Review, a stale refusal and Apply | Speed, reliability rate, or whether the resulting code is correct |
| Signal Core | How much smaller the structural scaffold is than the source, in bytes | Model tokens, cost, or task success |
Rules we hold ourselves to
- A number is quoted only from its record file, with the date and build.
- Failed and refused attempts stay published next to the successful ones.
- Passing checks are reported as "these commands passed on these exact files", never as "the code is correct".
- Synthetic benchmarks say they are synthetic.
- Agents that have not completed a live end-to-end run on a release are labelled beta.
Run Pybench yourself with pytxo bench, and compare with the published board.