Every figure a recipe.
Every result reproducible.

01 The Monday 8:40 a.m.

The matrix arrives. The dread arrives with it.

Dr. Nadia Okafor has a fresh RNA-seq run — forty-one tumor samples, clinical annotations, a mutation table her collaborator swears is "basically clean." What she also has is the memory of last spring: three weeks of R scripts, a heatmap regenerated eleven times for the reviewers, and a final panel nobody could rebuild when the raw counts were re-normalized.

The biology is the easy part. The provenance is what eats the month — which script made which figure, from which version of the data, with which clustering seed. By the time the paper is accepted, the answer is a folder named final_v7_REALfinal.

02 The ask plain language

She doesn't open an IDE. She just says what she wants.

She drops the matrix into the bench and types the sentence she'd say to a rotation student:

Under the sentence, the bench doesn't hand her a black-box PNG. It writes a recipe — a small, legible spec of every step — and renders it live with real, native widgets: a clustered heatmap with its dendrogram, an oncoprint no other browser library ships whole, a Kaplan–Meier below it.

03 The turn it talks back

She argues with the figure, and the figure keeps its history.

"Reorder the oncoprint by mutation frequency." Done. "That teal is too close to the stage track — make the subtype magenta." Done. Every nudge is another line on the recipe, every state a point she can step back to. When her collaborator asks why cluster 2 looks that way, she doesn't reconstruct her reasoning — she scrolls it.

The undo stack isn't a convenience. It's the lab notebook the field never had.

Because the whole thing is one portable {data, config} object, she pastes it into the manuscript's supplement as-is. Not a screenshot — the living recipe. A reviewer with the raw counts gets byte-for-byte the same panel, six months from now, on their own machine.

04 Why it can only be this library

Four things that almost never live together.

This isn't a mock-up. Here it is, running.

The story above ships in this repository. The cohort below was distilled from a simulated 20,000-gene run, packaged by a recipe, and rendered live — clustered heatmap, oncoprint, and Kaplan–Meier, each computed by CanvasXpress in the browser from figure-scale tables.

And the "she just says what she wants" step is real too. Below, the Bench: survival dataset is ticked and the assistant is asked for a Kaplan–Meier plot — time, event, split by the derived cluster — and the figure lands on the canvas, ready to refine by hand or by chat.