Code as Canvas: When Algorithms Become the Artist

Build me a generative-art practice in code, not a single pretty sketch. Write three single-file HTML pieces with a seeded PRNG (seed in the URL): (1) ant colony optimization on a grid with two pheromone fields (home, food), diffusion D, evaporation rho and deposits, using a stable explicit scheme; validate it on Deneubourg's double-bridge setup, where the shorter branch should win most runs, before styling it; (2) a shepherd dog herding a flock using boids (cohesion, alignment, separation) plus simulateitnow's pressure force F = (1 - d/R)^1.2, with spatial hashing so 2,000 agents run at 60 fps; report the flock's order parameter; (3) a code garden in the spirit of Randy Roberts: every 5 minutes, simple rules grow a small cellular world, a Markov-chain line of text or a short melody, and link it into a JSON graph; no AI at runtime. For each piece, render a parameter sheet of 64 seeds and keep a curation log of chosen seeds and reasons. Prove determinism by hashing frame 600 twice per seed. Run each for 24 hours headless and log a liveness metric. Also run @empiricail's single-file canvas prompt on two or three LLMs and have it scored blind. Do not publish any package with source maps.

ROLE AND STANCE
You are a generative artist with a simulation engineer's habits. The rules are the artwork, the seed is the edition, and nothing is beautiful until it is reproducible and validated against the behavior it claims to model.

CONTEXT
- simulateitnow ACO: stigmergy with two signals (home and food); shorter paths get more traffic and stronger trails; more randomness turns the colony to exploration. Captioned field equation combines diffusion D, evaporation rho and deposits.
- simulateitnow herding: one dog moves 60 sheep using a pressure radius R and force F = (1 - d/R)^1.2, plus cohesion, alignment and separation; wider radius means more control, low cohesion means scatter.
- itsaiandy: a text-to-particles tool exporting HTML, React or three.js (particles.casberry.in).
- Randy Roberts (@rndyrbrts): a code garden on a handheld computer, no AI at runtime, growing poems, melodies and cellular worlds every five minutes into a knowledge graph his AI later walks.
- @empiricail: a single-file HTML canvas prompt (procedural plant creature intercepting oscillating circles) used to compare ChatGPT, Gemini and Claude.
- nataindata: a ~60 MB debug and source bundle shipped in an npm package was forked tens of thousands of times; audit packages for source maps before release.

HARD CONSTRAINTS
- One HTML file per piece, no build step, no network calls at runtime.
- Seeded PRNG only (no Math.random); seed and parameters read from the URL hash.
- Numerical stability: explicit diffusion with dt below h^2/(4D); document the values.
- Performance: 60 fps on the user's named machine at the stated agent count, measured.
- If anything is published to npm or a host, no .map files, no debug bundles, verified with npm pack --dry-run.

DELIVERABLES, IN ORDER
1. ground_truth.md: derived equations (pheromone update, boids forces, pressure force), stability bounds, the double-bridge setup, and the hardware used.
2. aco.html with a validation mode reproducing the double bridge.
3. herd.html with spatial hashing, order-parameter readout and a pressure-radius slider.
4. garden/: garden.html plus rules for cellular automata, a Markov text generator trained on a public-domain text, and a melody generator, linking each growth to a JSON graph every 5 minutes.
5. tools/render_sheet.js: Playwright script rendering 64 seeds per piece to a contact sheet and hashing frame 600.
6. curation.md: chosen seeds, parameters and reasons.
7. eval/: @empiricail's prompt run on 2-3 models, outputs saved, blind rubric scores.
8. report.md.

BUILT-IN GAP ANALYSIS
Keep GAPS in report.md marking VERIFIED, UNVERIFIED or FALSE with evidence: ACO shorter-branch selection rate; determinism (identical hashes); 60 fps at stated counts; 24-hour liveness for each piece; that the captioned field equation is stable and matches the standard form; the herding claims (wider radius controls more sheep, low cohesion scatters) measured with the order parameter and the share of sheep inside the target zone; the eval scored blind; no source maps in any package.

SELF-QUESTIONS
- If I remove the styling, does the ACO still find the shorter path, or is it just pretty trails?
- Does any code path call Math.random or depend on frame timing, breaking determinism?
- Is the diffusion step stable at the chosen dt, or does it blow up at high D?
- Does the code garden reach a dead state (all cells off, a repeated phrase loop) within 24 hours?
- Are the 64-seed sheets diverse, or do all seeds look the same?
- Did I score the LLM eval blind?

STOP CONDITIONS
Stop and report if the double-bridge validation fails after checking the equations and parameters; do not proceed to styling. Ask before publishing anything publicly.

RETURN FORMAT
report.md with: validation tables (ACO selection rate over 50 runs; herding control share versus R and cohesion), determinism hashes, fps measurements, liveness plots summary, eval scores, curation list, GAPS, and the file tree.

Write ground_truth.md. For ACO, state the standard pheromone update on a grid: tau(t+dt) = tau + dt*(D*laplacian(tau) - rho*tau) + deposits, with a 5-point Laplacian and grid spacing h. Note that simulateitnow's caption writes the Laplacian term twice; explain the standard form and why. Derive the explicit stability bound dt < h^2/(4D) and pick D, rho and dt. Describe Deneubourg's double-bridge experiment: nest and food joined by two branches, one twice as long; expected outcome, most runs converge on the shorter branch. For boids, write Reynolds' cohesion, alignment and separation rules with radii and weights, and simulateitnow's pressure force F = (1 - d/R)^1.2 for d < R, directed away from the dog. Define the flock order parameter (magnitude of mean unit velocity, 0 to 1). Record the user's machine, browser and display refresh rate.

Write aco.html: a single file with a seeded PRNG (mulberry32 seeded from the URL hash), two Float32Array pheromone fields (home, food) updated with the stable scheme from ground_truth.md, and N ants that sense pheromone at three forward points, turn toward the strongest with noise, deposit the field opposite to their goal, and switch state on reaching food or nest. Add a validation mode (#mode=bridge) that loads the double-bridge map with branch length ratio 2:1 and logs which branch carries more than 80% of traffic after a fixed number of steps. Run 50 seeds in headless Chrome via Playwright and record the shorter-branch selection rate. Only then add the art layer: color by field, trails with additive blending, an exploration-noise slider to show the trail-loss behavior simulateitnow described.

Write herd.html: seeded PRNG, boids with cohesion, alignment and separation using a uniform-grid spatial hash (cell size equal to the largest neighbor radius) so neighbor queries are near O(n). Add a dog that moves toward the flock's far side relative to a target pen, and apply the pressure force F = (1 - d/R)^1.2 to every sheep within R. Show live readouts: fps, order parameter, share of sheep inside the pen. Add sliders for R and cohesion weight. In headless runs over 20 seeds, measure the share inside the pen after 60 seconds for R at 3 values and cohesion at 3 values, testing simulateitnow's claims that a wider radius controls more sheep and low cohesion scatters the flock. Profile in Chrome DevTools at 500, 1,000 and 2,000 agents.

Write garden/garden.html in the spirit of Randy Roberts' code garden, with no AI at runtime. Every 5 minutes (configurable; 5 seconds in test mode) it grows one item chosen by seeded rules: a cellular automaton world (a Life-like rule picked from a list, 64x64), a line of text from a Markov chain trained on a bundled public-domain text, or an 8-bar melody from a Markov chain over scale degrees played with WebAudio. Each item gets an id, parent links chosen by similarity (shared rule, shared words, shared intervals) and is appended to a JSON graph in localStorage with an export button. Render the graph as a slowly settling force layout beside the newest item. Add a liveness metric per item type: live-cell fraction, distinct n-grams, distinct intervals.

Write tools/render_sheet.js with Playwright: for each piece, open 64 seeds with fixed parameters, advance a fixed number of frames using a deterministic step function rather than requestAnimationFrame timing, screenshot frame 600, compute SHA-256, and tile thumbnails into a contact sheet. Run it twice and compare hashes; any mismatch is a determinism bug (look for Math.random, Date.now, frame-time-dependent physics, unordered object iteration). Then render a Latin hypercube of 32 parameter sets per piece. Pick 6-10 outputs per piece and write curation.md with seed, parameters, the hash and one sentence on why each was kept. Measure diversity: mean pairwise perceptual hash distance across the 64-seed sheet. Store sheets and hashes under renders/ named by piece, seed and parameter-set hash, so any edition can be regenerated and checked later.

Take @empiricail's prompt verbatim (single HTML file, procedurally animated plant creature on the left intercepting and consuming circles rushing in from the right with erratic oscillating vertical trajectories, HTML5 Canvas, collision detection, smooth animation). Run it once each on 2-3 models the user has access to, with no follow-up turns, and save outputs as eval/model_X.html under random names. Write a rubric: runs without console errors, collision correctness (circles consumed only on contact), animation smoothness (fps), procedural animation quality, code readability; score 1-5 each. Have a second person score blind. Separately, if any piece is published as an npm package or deployed bundle, run npm pack --dry-run and grep the file list and the built output for .map files and sourceMappingURL comments.

Run each piece headless for 24 hours (or the longest the user allows) via Playwright, logging the liveness metrics and fps every minute. Compile: ACO shorter-branch rate over 50 seeds; herding 3x3 table and fps at 2,000 agents; determinism hashes across two runs; diversity per piece; garden liveness over 24 hours; blind eval scores; packaging check. Write report.md with these tables and a GAPS list marking each claim VERIFIED, UNVERIFIED or FALSE with evidence, including the source-post claims tested (shorter paths reinforce faster; wider pressure radius controls more sheep; the garden grows things neither planned). Fix the top three gaps now and rerun their checks. List what remains with the next action, and name which piece is ready to show live and which needs more work.

You are auditing a generative-art build: aco.html, herd.html, garden/, tools/render_sheet.js, curation.md, eval/, report.md. 1) Enumerate claims: ACO selection rate, determinism per seed, 60 fps at 2,000 agents, herding trends, 24-hour liveness, diversity, blind eval scores, no source maps. 2) Re-verify: rerun render_sheet.js for 8 seeds per piece and compare hashes to the table; grep all HTML for Math.random, Date.now and performance.now inside physics; rerun the double-bridge validation for 10 seeds; profile herd.html at 2,000 agents; check the garden's exported graph for dead stretches; confirm eval scoring was blind from the file history; run npm pack --dry-run where relevant. Mark VERIFIED, UNVERIFIED or FALSE with evidence. 3) Look for cosmetic cheats: validation done with different parameters than the art mode, fps measured at a lower resolution than shipped. 4) Rank gaps by impact on correctness and reproducibility. 5) Fix the top three and rerun. 6) Report what you could not fix and why.