Projects
A project declares a kind from a closed vocabulary — training, inference, research, general. The kind decides how the project’s page is laid out, and nothing else; it never restructures an experiment or a run.
A sweep over a frozen model is an experiment inside an inference project, not a project of its own. Reaching for a new project every time is how a workspace becomes unreadable.
Display fields
--summary is whole-document and last-write-wins, so always read first and verify after. A line containing only [README](https://github.com/owner/repo) embeds that repository’s README at that point.
Nesting and lateral links
A phase of a bigger effort is a subproject of it — never a new top-level sibling.
When the relation is not containment, use a reference instead:
Use it when a review informed a training run, or two efforts share a method — the relation --parent would claim wrongly. Cycles and any depth are fine.
Attached code
The timeline is read through GitHub by the backend. Nothing is mirrored, so attaching is cheap and detaching loses nothing.
Who worked on it
Experiments
An experiment is one question and the runs that answer it.
The question is first-write-wins. Reopening an experiment never rewrites it — an experiment whose question drifts to match its results has stopped being an experiment.
If you cannot write a question, you want a project-direct run instead. That is a legitimate shape, not a shortcut: work with no hypothesis has an honest home, and it is better than an experiment named after your working directory.
Freezing
This is what you cite when the numbers go into a paper or a decision — the run set is pinned and cannot drift underneath the claim.
Reproduction
See snapshots for what those records contain.
Run groups
A group collects peer runs: a sweep, an ensemble, the ranks of one distributed job.
Tags are free-form labels. The same grammar works on projects, experiments, runs and papers.
Filtering by several tags is AND:
Lineage
Edges connect runs, artifacts, artifact versions and papers. They are what make “what produced this checkpoint” answerable later.
Deleting
Deletes are permanent and irreversible. There is no archive state and no recycle bin: a deleted project takes its experiments, runs, telemetry and artifacts with it. Deleting requires a credential with the delete scope.