Claude Code (Opus 5.5) implemented ten published models of how animal forms arise and rendered each animal from its model's output. Each animal is one HTML page of Canvas 2D code with no image files, 15,227 lines in all. Each page draws its construction next to the form, and its NOTES.md cites every parameter.

Site: each plate next to its source, finding, model and implementation notes. Plates compute in the browser; click one to run its model, then scrub time. Film (64 s, silent).

Fig. 1 Honeybee	Fig. 2 Sea urchin	Fig. 3 Dragonfly	Fig. 4 Orb-weaver	Fig. 5 Nautilus
Fig. 6 Camp pitar	Fig. 7 Sea fan	Fig. 8 Cone shell	Fig. 9 Angelfish	Fig. 10 Tortoise

Each builder session started from the source on its animal, implemented the model it describes, and rendered the animal from the model's output. The construction is drawn on the plate, so the picture can be checked against the model.

Each page's NOTES.md gives the mathematics, the sources for every number and the trials behind it. Fact-checker sessions verified every caption in the film against a primary source.

Claude Code (Opus 5.5) wrote every line. I directed and gave notes.

• A lead session wrote the concept, one shared builder brief (BUILDERS.md) and a card per animal: its rule, its sources, its moment.
• Ten builder sessions worked in parallel, one per animal, each reading its research and implementing the model.
• After every round, a new reviewer session that had never seen the work judged the rendered frame alone and returned PASS or REVISE with ranked changes.
• Fact-checker sessions verified every caption against a primary source.

154 Claude sessions in all (3 leads, 46 builders, 97 blind reviewers, 8 fact-checkers): about 92 agent-hours and 9,900 tool calls, with runs of up to 8.5 hours unattended on a remote desktop. The ten pages took about six and a half hours of wall-clock time, and the film went through 20 drafts on my notes.

1. The rule generates the image. Choose a peer-reviewed model of the form (a growth law, a pattern-forming equation, a tiling). Implement it, and render its output directly.
2. The construction is part of the image. Draw the computation (the field, the lattice, the seed points) in register with the form it produces, so a viewer can check the picture against the rule.
3. The render is a pure, seeded function of time. render(t) draws one frame from a hashed seed and nothing else. This makes animation, frame-exact film, regression checks and reproducible output possible without extra work.
4. One shared material and light kit. A single shading library (soft area light, occlusion, scattering, reflections blurred by roughness) is inlined into every piece, so the set reads as one body of work.
5. Every number has a source. Each parameter and caption comes from a primary source and is cited in the frame. A value without a source is cut.
6. The maker does not judge its own work. Each round, a new reviewer sees only fixed inputs (the output at feed size, one crop, one reference) and returns a verdict with up to three ranked changes.

flowchart LR
  P[Primary literature] --> M[Model implemented in code]
  M --> R["render(t): seeded, pure"]
  K[Shared light and material kit] --> R
  R --> V[New reviewer: output, crop, reference]
  V -- REVISE + ranked changes --> M
  V -- PASS --> O[Stills, build strips, film]

• Contract: all ten pages are blank at t = 0, built by 4.5 s and deterministic, and none uses an image file.
• Form: each animal is drawn by its cited rule, and the construction stays on the page.
• Material: nine of the ten were reviewed in the first series. One passed (the dragonfly), and the other eight ended on REVISE after four to six rounds, mostly because the pencil was too faint at feed size and the surfaces looked computed. The same nine then went through a round on material realism, and all 45 reviews in that round returned REVISE.

Any form with a published generative model: shells and horns, phyllotaxis, crystals, river and vein networks, cracks, coat and pigment patterns. Each result can be checked against its source, line by line, which images from generative models cannot.

• Material realism: computed shell, pigment and membrane are still recognised as computed. The next step is to sample material statistics from reference photographs, beyond their palettes.
• Review signal: reviewers who see only the output rarely return PASS. A score per criterion would show progress between rounds more clearly than PASS or REVISE.

• NN-name/index.html: one page per animal, self-contained. Open it in a browser with ?play to watch it build, or ?t=2.5 for one moment. NN-name/NOTES.md gives the mathematics, its sources and the trials behind each page.
• out/: each page's finished frame at 5.2 s, 1080 × 1350.
• kit/: the shared light and material kit, inlined into every page by node kit/sync.mjs <page.html>.
• CONCEPT.md: the concept card and the concept frames considered (concept-frames/). It describes the first series of ten, in which a glass sponge held the place the camp pitar holds in the film.
• BUILDERS.md: the brief each builder worked from.
• index.html: the site served at r3al1tym.github.io/math-animals, set in Libre Caslon Text (fonts/, OFL).
• film/math-animals.mp4: the film, 1080 × 1350, 30 fps, silent.
• tools/render.mjs: renders a page at given times and checks it is deterministic.

cd tools && npm install
node render.mjs ../05-nautilus/index.html /tmp/nautilus 0,1.5,3,4.5,5.2

Made by Sanju Sunny (@sanjusunnytwt) with Claude Opus 5.5. Math-Animals extends tenfold, and its bar for craft is tenfold's 16-specimen page. Every page and every line of code was written by Claude Opus 5.5. MIT licence.