CLAUDE.md — jiantastic.github.io

Jekyll static site hosted on GitHub Pages. Dark-themed personal site with interactive prototypes.

Stack

Jekyll (GitHub Pages) — no local build needed; push to master deploys
Plain HTML/CSS/JS for prototypes (ES modules, no bundler)
Vitest for unit tests (root vitest.config.js, env: jsdom)
Run tests: npm test or ./node_modules/.bin/vitest run

Key directories

matrix-evolution/   BFF primordial soup simulation (main prototype)
artificial-life/    Python/Numba reference implementation (main.py)
_includes/          Jekyll partials (nav.html etc.)
_layouts/           Jekyll layouts

matrix-evolution — BFF Primordial Soup

Based on Computational Life (Agüera y Arcas et al., 2024).

Files

File	Role
`bff-core.js`	Pure simulation engine — no DOM, fully testable
`sim-worker.js`	Worker loop for paper-scale browser execution
`app.js`	Canvas renderer + worker/control wiring
`index.html`	Page markup (Jekyll front matter, controls, canvas)
`style.css`	Dark theme; `#sim-canvas { image-rendering: pixelated }`
`sim-core.js`	Kept on disk — original neural-agent sim (untouched)
`tests/bff-core.test.js`	16 BFF tests
`tests/*.test.js`	6 original neural-agent tests

Grid & constants (bff-core.js)

GRID_W = 240, GRID_H = 135   → 32,400 programs
TAPE_SIZE = 64               → ~2.0 MB total
MAX_ITERS = 8192             → BFF steps per interaction
DEFAULT_MUTATION_RATE = 0.00024

BFF opcodes

< 60  > 62  { 123  } 125  - 45  + 43  . 46  , 44  [ 91  ] 93
<>  → head0   {}  → head1   +-  → scratch[head0]
.   → scratch[head1] = scratch[head0]
,   → scratch[head0] = scratch[head1]
[]  → loop brackets (check scratch[head0])

Semantics

Two programs (A, B) concatenated into scratch[0..tapeSize*2]
head0 and head1 wrap around tapeSize * 2
PC terminates when out of [0, tapeSize*2) OR after maxIters steps
seekMatch is iterative (not recursive) — avoids stack overflow
Mutation uses geometric skip (O(expected mutations), not O(total bytes))

Indexing

Programs: column-major — cellIdx = cellX * GRID_H + cellY (matches Python)
ImageData: row-major — pixBase = (cellY * GRID_W + cellX) * 4
Neighbors: Chebyshev radius 2 → up to 24 per cell, stored flat in Int32Array[numPrograms * 24]

Rendering

Canvas intrinsic = GRID_W × GRID_H (240×135 pixels)
CSS width:100%; height:100% + image-rendering: pixelated scales it up crisply
One pixel per cell; color = dominant opcode in that cell’s tape
Worker sends compressed dominant-opcode snapshots to the main thread
getDominantOpcode/computeDominantOpcodes use a module-level Uint16Array(256) scratch

Color palette (matches main.py)

< [239,71,111]   > [255,209,102]  { [6,214,160]    } [17,138,178]
- [255,127,80]   + [131,56,236]   . [58,134,255]    , [255,190,11]
[ [139,201,38]   ] [255,89,94]    noise → [20,20,20]

Epoch loop (runEpoch)

Fisher-Yates shuffle of all cell indices
Each cell picks a random neighbor → proposals[]
Greedy pairing: both must be untaken → run BFF pair
Background mutation via geometric skip

Performance notes

xorshift32 used in hot path (not Math.random())
scratch pre-allocated in SoupState, reused per pair
seekMatch iterative — safe for deep nesting
Browser page runs simulation inside sim-worker.js so paper defaults do not block the UI

Python reference

artificial-life/main.py — Numba-accelerated version. Semantic baseline for the browser implementation. Key difference: Python uses prange parallelism; JS is single-threaded.