The Archive

The Archive (/vault) is where a workspace’s knowledge lives next to everything that produced it. It opens on Knowledge — the findings and decisions of every research, as one index — and keeps two more views: the Corpus (chats, runs and logbooks, grouped into themes by similarity) and Runs (the execution list, with costs).

Knowledge

One outline per research, because a move’s number is an ordinal within its research: two researches both have an “m1”, and mixing them makes the numbering unreadable. Picking a move opens it in the inspector — the same inspector as in Research, so what you can do there you can do here: say something about it, hang it under another move, add a citation.

Following a citation leaves for the evidence. A cited graph version opens that research’s canvas at that version; an experiment opens its tab; a trial opens unfolded, which costs a round trip because a citation carries the trial’s id and the tab lists the trials of an experiment. In Research, following moves you within the research; here it takes you to it, because the Archive looks at the whole workspace.

Corpus, and the round trip

Selecting a chat, a run or a logbook shows what it is — the trace for a run, the messages for a chat, the preview for a logbook — and, above it, which findings cite it.

That link is not as direct as it looks, and the shape is worth knowing: a finding does not cite runs. It cites a trial, an experiment or a graph_version. A run in the corpus is the run of a trial, so the Archive answers the question through the trial the run belongs to.

Its experiment does not count

If a finding cites the experiment a run came from, the panel says so — but the run is still listed as not promoted. An experiment has many runs, and citing it speaks about the experiment, not about any one of them. Counting it would mark a hundred runs as promoted from a single citation, and “what nobody has promoted” would stop meaning anything.

Promoting: from material to knowledge

The round trip goes both ways. Promote to a finding, on any chat, run or logbook in the Corpus, writes a finding in the research you pick and cites the piece you were looking at — which is the moment you usually notice there is something there.

What it asks for is a research and the prose, and nothing else. The prose is required: a move’s text is what gets indexed, so a mute finding would exist, would read fine in the outline, and would never be found by a search of the archive. Promoting then lands you on the new finding in the inspector, where saying what it answers and hanging it under a question already live.

A piece can be promoted more than once

The button stays available on material that findings already cite: the same conversation can feed a finding in two different researches, and nothing about the archive says it should not.

Only unpromoted

The Corpus has a filter for material no finding cites yet. It is what the similarity themes are for once knowledge is the main view: grouping the corpus by resemblance does not tell you what was learned, but crossed with this filter it tells you something no other screen does — there is a whole theme here that nothing has come out of yet.

The filter applies before the theme filter so the two can be combined, and it does not filter at all until the workspace’s reasoning has loaded: not knowing is not the same as nobody having promoted anything.

Themes

Themes are built by clustering the corpus’s embeddings — no model writes them; the names come from the words that distinguish each cluster. Find themes recomputes them, and the Archive re-fits on its own when enough new material has arrived that a fit would change something. A theme you add by hand is never touched by a re-fit.

Themes need an embedding model; the knowledge view does not, which is why it is the one that opens.