x = week · y = commits, split by how they landed: merged PR vs direct push. This is the thesis metric applied to ourselves — how much of the work passed review. The honest answer: 4% via PR overall, and the split differs sharply by repo.

x = messages per session (log) · y = output tokens (log) · color = model. The diagonals are tokens-per-message guides. Same effort can be terse or chatty — and whether that differs by model is now a lookable-up fact, not a vibe.

x = files touched (log) · y = lines changed (log) · one dot = one commit. Below the guides: wide-and-shallow (renames, sweeps, generated output). Above: narrow-and-deep (dense logic in few files — the commits that deserve review time).

x = words in a digest story · y = evidence citations it carries · one dot = one story. Our own honesty discipline, measured: longer claims should carry more receipts. A dot drifting right without rising is marketing creep — the chart exists so we'd see it.

Every "in flight" item ever published, arced to its fate: landed in a later issue, still moving, or went quiet. Quiet isn't dead — it's a debt the next digest owes. Matching is automatic (word overlap against later shipped stories), so the tracker can't be quietly edited.

Two radial 24-hour dials: when the AI generates (left) vs when code lands (right). The org has two circadian rhythms; whether they beat together or the AI runs ahead into the night is now a picture.

The busiest day in the window — 2026-11-11 — at message grain: x = time of day, one lane per session, one mark per active minute, sized by output. The far end of the granularity dial: 981 marks for a single day, and the day's shape (parallel morning sessions, the evening tail) is legible with zero reading.

x = cumulative AI tokens spent · y = sharp edges closed. The ultimate pair — expenditure against the honesty ledger actually shrinking. It has no marks because nothing has flipped to fixed yet. Publishing the empty chart is the commitment device: the first closure draws the first mark.