The viewer tells its treatment windows apart, and hands a visitor a recording

[syncytium2]

Three things Tony asked for after looking at the import page, plus one finding that fell out of building them.

Treatment regions blended together

Every window drew in the same grey wash, so baseline abutting drug abutting washout was one block with three names floating over it. Each label now takes a hue from a fixed categorical palette, assigned to the label rather than the position — so TTX is the same colour in every recording of the folder and comparing two of them is a comparison. The slot order is the one validated for colour-vision deficiency (adjacent pairs clear ΔE 8 under protanopia and deuteranopia in both light and dark); a ninth distinct label gets a neutral rather than a ninth hue that would collide with one of the eight.

No text is ever painted in the window colour. The wash stays pale enough that a mark over it is still a mark, a saturated rail along the top of the band carries identity, and the name sits in ink beside a colour chip — on the raster and in the legend.

Simulated data, for a visitor with no folder

The page will now invent one: recordings emitted as the CSV text the import contract describes and handed to the same loader, which is not told where they came from. An explorer drives the real reading path — same parser, same folder check, same raster — and the NA row for an ROI that fired nothing is exercised rather than described.

It opens at the measured settings, and that is the load-bearing part. Rate, ROI count, participation and jitter come off untreated slices: about six of thirty-three ROIs per event with a third of a second of spread. Shipping the old guesses as the public default would have published the arrangement that scored every detector at F1 0.9–1.0 and cost two weeks; they are on the switches instead, beside the flat background, because the difference is the thing worth looking at.

Three refusals, said on the page: it does not match Python's numbers (same model, different generator — a seed reproduces in this browser), it does not simulate a treatment (the windows are labels over identical statistics), and none of it is evidence — which the page says on the picture, not only beside the controls.

Accordion sidebar

Three sections — a folder you have, a folder that does not exist, and the recordings either produced — one open at a time, because the list wants the height and the controls want attention while they are being set. The folder check's verdict rides on the Recordings header, since that section can be shut and a check nobody sees is a check nobody ran.

One bug, one finding

all: unset on the list rows took box-sizing with it, so width: 100% went content-box and every row ran its event count past the panel edge, where the sidebar clipped it to a bare number.

And: the fitted background's quiet tail does not survive the events planted on top of it — 28.8% of ROIs with no event becomes 0.5% once the fifteen coordinated events go in, because participants are drawn uniformly. A third face of the mechanism PR #50 documents twice, and one neither section states, since it needs nothing switched on. Written up in docs/todo/, with tools/quiet_tail_vs_events.py to re-take the table rather than quote it. Two implementations agree.

Checks

pytest green (481 passed, 1 skipped), sapper clear, tools/build_site.py builds — including its refusal to publish a viewer that can reach the network, which tests/test_site_viewer.py also proves.