Computed, not sketched

Ask a language model to draw a supply-and-demand chart and it will draw something that looks right. Whether the equilibrium sits where the curves actually cross is another matter, and a student has no way to tell the difference. A diagram that is confidently wrong teaches the error more durably than text ever could, because the eye trusts a picture.

So Chameleon never lets the model draw. The model emits a small data specification, the numbers and labels taken from your material, and a deterministic renderer does the drawing.

Each figure type computes its own domain result on the server before a single line is drawn. The Riemann rectangles are actually summed for the given function and n, and the Σ printed on the figure is that sum. The market equilibrium is solved from the curve equations. The titration curve bends at the real pKa. The circuit's currents come from Ohm's law, the Punnett ratios from the crosses, the rejection regions from the exact critical values. If the numbers in the specification are inconsistent, the figure fails closed rather than drawing something false.

Getting the math right is half the job; a figure also has to survive its own labels. Chameleon measures every piece of figure text with the real glyph metrics of its typeface, the same way a typesetter would, so a long axis title or a crowded legend is fitted rather than guessed at. A layout pass then resolves the whole figure: labels are kept inside the frame, stacked annotations hold their line spacing, and free-floating labels are nudged apart when they would collide. Ticks and axis titles are pinned, because a tick that moves away from its gridline is a lie.

Figures are drawn in the app's own palette tokens, so the same figure is legible on the warm paper of the light theme and the lamplight of the dark one.

The catalog covers the figures each discipline actually leans on, and it grows by family rather than one-off. Process and relationship diagrams (flowcharts, sequence and state diagrams) render alongside them for material that is structural rather than quantitative.

Any figure can carry a one-line prediction prompt: commit to which way the curve shifts, or where the test statistic lands, before the answer opens. It is the cheapest form of retrieval practice there is, and it turns a figure from something you look at into something you answer to. The reader even reopens these prompts on concepts your review record marks as weak.