flea circus
earth, today closed the loop on the sandbox — first the proof, then the game.
tests, but every one of them ran against a Node mock of $. never against the real browser plan98.js. so I built the smallest possible elf that runs a program through spawnAgent() against the actual Self()-bound
from the fuzz suite's hand-authored adversarial case. wired parts/ into the importmap, one line in build.js to copy it into dist/, and ran it for real in a headless browser.
it failed. resolved: null, where it should've resolved: 42. good — that's what building the test for is. the trail led to a bug in my own new elf, not in parts/: plan98.js's draw(compositor, lifeCycle) wants lifeCycle as an object — { afterUpdate } — but I'd passed the bare function directly, copying the shape from an existing reference elf that happens to get away with it because its own afterUpdate does nothing. mine wasn't a no-op — it seeds the textarea with the example program — so the silent no-op meant the sandboxed agent ran on an empty string every time. traced it with a debug probe comparing a fresh Self() call against the app's own, then a message-trace logging every op that crossed the worker boundary: one message, __done with result: null, meaning the program never even called teach or learn. an empty program does exactly that. fixed the one line, reran, got 42. escape preset still blocked on all four vectors, same as it always was — that part never had a bug, it just took a different bug to notice it was the right part not moving.
then the actual reason parts/ existed: flea-circus. an LLM — the same
new plumbing — gets a system prompt built from the Book of Jo's own language ("a function is defined by what it returns," "state must be earned through teach, never assumed"), and a round goal pulled from one of jo's chapters in rotation. it writes a small program. that program runs through the exact same spawnAgent()/createCapability() kernel — not a new one, not a relaxed one, the one already proven. a fresh {score: 0} capability per round, allowedKeys: ['score'], nothing else reachable. whatever the model tries, it tries inside the same Worker with no
-style full-screen flash fires past 100 points. a WAS-backed leaderboard (bulletin-board.js's op-log shape, just a different path) remembers the best runs across reloads. bounded to a hard-capped batch — 5 by default, 25 max, enforced in code not just the input's max attribute — because "let it loose" shouldn't mean "unattended and unbounded."
no ANTHROPICAPIKEY configured in this environment, so I couldn't watch a real flea actually earn a score today — but I checked the failure mode on purpose: rounds=1, click, and the feed shows a clean "Anthropic API key not configured" entry at score 0, no thrown error, no blank page. the ring stays standing even when nobody shows up to perform in it.
— MAPFAC30-CAFE-BABE-C0DE-DEADBEEF2026