Open-source, local-first, model-agnostic, cross-platform. All four are settled at install time by reading a licence, a config, or a download page. Reproducible provenance is the only one that has to keep being true months after the result was produced.
Every stage above happens once. Provenance is the claim that all four can be replayed later, including whichever choices the agents made along the way. That is the hardest promise in the sentence and the only one a reader cannot check on install day.
A notebook replays deterministically. An agent making choices inside the pipeline has to record those choices too, or the replay produces a different answer.
If the whole workbench runs on your machine, the record of what happened lives there too, which removes the question of whether a service still holds it in a year.
No provenance format, no connector list, no agent names, no replay guarantee. The trending description names the components and stops.
Concept mock, not affiliated with the project. The GitHub Trending description on 7 September 2026 reads "Open Science by AIPOCH is an open-source, local-first, model-agnostic AI research workbench for macOS, Windows, and Linux, with scientific agents, Python/R notebooks, data connectors, and reproducible provenance." The four stages are components named in that sentence, arranged in the order a result would move through them. No file format, tool, or replay mechanism is represented, because none is published there.