Skip to main content

​
Series identity

Every metric series has a stable drill-down identity: (kind, key, dimensions). Chart links preserve all three, so coordinate variants can never accidentally share an underlying-data table.

​
Three kinds of card

Logged series

What the training loop wrote. Full chart, table and export.

Derived series

Computed after the fact, with stored rows. Carries a derived chip — producer and computed-at on hover.

Expression panels

A read-time formula over logged series. Carries an ƒ chip and shows the rendered formula as a subtitle.
The distinction is load-bearing. A post-hoc number must never look like training-loop capture, so the chips are not decoration — they are how you know which you are reading.

​
Expression panels

The stored object is a formula, never data. Panels are agent-authored: the expression comes from a researcher’s agent session or their own script, through the SDK or CLI. The dashboard renders, renames and deletes a view but never composes one — there is no expression editor and no template picker, because a fixed form cannot express what an arbitrary formula needs. What the dashboard owes instead is legibility. Opening a panel gives you the formula, its inputs, the raw spec, and who or what generated it. The name is the only editable field. A panel charts its evaluated result and exposes no captured-data drill-down — there are no underlying rows to open. Evaluation trouble surfaces on the card (named missing inputs, a dropped non-finite count) rather than silently rendering a partial curve, and the detail page’s value table is labeled computed.

​
Chart settings

Available on logged-series and expression cards alike, immediately before Delete. They are presentation-only: title, native X axis, ranges, labels, scale and smoothing — never the stored expression or its inputs. Settings anchor to the immutable view id, not its editable name, and inherit run-wide defaults. Deleting a view removes its settings row in the same transaction, and a stale save serialises behind that deletion.

​
Zoom

Drag a box on a Cartesian chart to narrow both axes to it. Hold Shift to lock the zoom to whichever direction the drag travelled farther in, leaving the other axis at the window it already had — the selection band stretches across the full plot on the free axis and the cursor turns directional to say so. The resolved axis is fixed at the last painted frame, so the window that commits is always the band that was on screen. Reset restores both axes. The gesture hint rides the chart’s accessible description as well as its hover tooltip, so reading it never depends on having a pointer that can hover.

​
Seeing the actual rows

Every chart carries a data action — on run overview charts, stat tiles, step charts, sandbox-diff charts and AI-summary charts alike.
The chart may use a bounded plotting read. The table does not. It uses lossless, cursor-paginated export rows rather than downsampled chart data, so what you inspect is what was captured.Transformations — grouping, smoothing, pivoting — are never presented as captured source rows.
Eligible cards also expose Raw samples as a direct action rather than an overflow item.

​
Stat tiles

A series with one point has nothing to plot, so it renders as a scalar tile. That is the intended shape for a headline number — and the unintended shape for a metric logged with a high-cardinality dimension, which produces a wall of them. See metrics for how to avoid that at the write.

​
In the terminal

The same read, drawn instead of dumped. See metrics.