Questions down, tags across, both axes sorted by total attached weight, so the structure falls out of the sort rather than out of a layout algorithm. Cells are explicit links only — a source or annotation actually attached to that question. Nothing here is inferred, nothing floats, and there are no edges to untangle.
{{ m.label }}
{{ scaleNote }}
QUESTION ↓ · TAG →
{{ c.label }}
TOTAL
{{ r.glyph }}{{ r.label }}
{{ r.total }}
TOTAL
{{ c.total }}
attached material
none · 1 · 2–3 · 4–7 · 8+{{ matrixFoot }}
{{ fallbackLead }}
{{ l.title }}{{ l.n }}{{ l.note }}
{{ i.n }}{{ i.text }}{{ i.tail }}
FOUR READINGS OF ONE PANEL
{{ r.title }}{{ r.count }}
{{ r.what }}
{{ r.example }}
Clocked-but-unquestioned is the one worth reading twice: material you kept from Skim with no question behind it is usually an interest you have not articulated yet.
INFERRED LINKS — AN ACTION, NOT A FAINT EDGE
Shared tags and keyword overlap never appear in the matrix as ghost connections. The same computation is offered as a review: here are the papers that share tags with an unanchored question. Accepting writes a real edge and the cell fills in; skipping is remembered so the same pair is not offered again.
{{ unanchoredHead }}promoted 3 months ago · nothing attached
A faint edge for every shared tag would make the matrix look complete while asserting nothing. Suggestions are guesses; the panel is a record. Keeping them apart means a filled cell always means you decided something.
WHAT ACCEPTING DOES
Writes the source into the question's supporting column as unsorted — attached, but not yet claimed to support or oppose. The cell in Panel 1 fills on the next render.
BACKLOG
{{ backlogNote }}
PANEL 2 — ATTENTION SCATTER
Different data source, deliberately a different picture: external literature volume, synced periodically, against your own coverage. Growth rate is publication-volume acceleration rather than citation count, because citations lag by years and the question is where the field is going. Depth controls how deep the tag paths are rolled up — mixing a parent with its own children would double-count coverage and flatten the quadrants into noise.
TAG DEPTH
{{ d.label }}
{{ depthNote }}OpenAlex volume · as of 17 Sep 2026, 04:00 · next sync in 2d
HIGH COVERAGE · LOW ATTENTION ahead of the field, or a backwater
HIGH COVERAGE · HIGH ATTENTION crowded — expect competition
LOW COVERAGE · LOW ATTENTION obscure — opportunity or dead end
LOW COVERAGE · HIGH ATTENTION blind spot — the actionable quadrant
{{ p.label }}
field growth rate →
UNPLACED
{{ u.label }}
Too few papers in the external index to measure a growth rate. Not the same as low attention — these are unmeasured, and plotting them at zero would put them in a quadrant that means the opposite thing.
PANELS 3 & 4 — FLOW AND ORIGINS
What happens to questions after capture, and where they come from. Both are counts you can walk into: every bar and every row opens the set behind it.
QUESTION FLOW · LAST 12 MONTHS318 captured
{{ f.label }}{{ f.n }}{{ f.pct }}
{{ f.note }}
The wide top is the design working. A capture rate that fell to match the promotion rate would mean you had started editing yourself at the moment of wondering, which is the one thing this app is built to prevent.
WHAT GENERATES QUESTIONSsources and contexts, ranked
{{ o.rank }}
{{ o.label }}{{ o.kind }}
{{ o.n }}
Lecture and lab-meeting captures convert to Research Questions at nearly twice the rate of paper captures, on a tenth of the volume. Worth knowing before you decide where to spend a Tuesday.