burger-generator

I got tired of chatbots guessing at my RTL. So I built an agent that has to prove it: the sim goes RED, the agent reads the failing trace, Muse Spark writes the diagnosis, a diagram explains the failure for a human, a ≤40-line patch lands, the sim re-runs — and the fix only counts if the checksum flips to MATCH. Every run leaves a proof bundle: the trace, the raw model requests, the diff, before/after checksums. Nothing hand-waved. On a pinned PicoRV32 + UART rig it holds locked 3/3 and full 7/8 with zero human touches, fixes 72 fresh injected faults (each RED-proven, quarantined from older families), and a preregistered ablation shows waveform images on top of text turn a real compositional miss green where text alone fails. One command reproduces it: ./afterglow run sim/targets/picosoc-uart-div.yaml. Open silicon, open loop — if we can prove a fix here, the same loop scales to any RTL you bring it.