RCFL

RCFL (Recursive Chaotic Fractal Learner) is a harness that rewrites its own rules. It watches chaotic visual data, scores its rule set at three scales every generation (micro: each rule knocked out alone; meso: every pair; macro: a held-out batch it never trains on), and keeps a rewrite only if it beats the champion and then re-proves itself on fresh chaos. Nothing starts over: the all-time champion lives in MongoDB Atlas and every run begins from it. In one day it went from three crude rules at 0.135 held-out to 0.833, with 0.807 on an untouched test batch and 0.993 on rendered files the generator never made, across 140+ runs and 21 re-proven replacements. When it plateaued it gave itself bigger eyes (a perception pool of its own generated chaos in Atlas), kept a ledger of which kinds of rewrite paid off, and read Raindrop's flags back into that ledger before the next run. The same loop, pointed at five simulated worlds it was not built for (Turing patterns, cell division, chaotic lattices, fragmentation, percolation), learned every one from seed rules above the majority guess. 161 tests; every run replays exactly from its seed and its drand round. Dataset pre-generated before the event; all harness code built during the event. Live page: https://rcfl.brainmatch.app/demo/