One generation through the Darwin Gödel Machine

    architecture node:

    Representation at this boundary

    Source modules

      Loop map

      Outer driver — choose_selfimproves → parallel self_improve futures → filter_compiled → update_archive → dgm_metadata.jsonl checkpoint. Two children per generation, 1.5-hour timeout each: a bad child is logged, never fatal to the generation.
      Inner loop is the evolving thing — coding_agent.py is one chat_with_agent call over a bash+editor tool loop. Only the coding-agent subtree is rewritten, and only that subtree is copied into eval containers — edits to the outer loop are silently dropped at eval time.
      The archive is the Darwin — keep_all admits every compiled child regardless of score, and the parent lottery weight is sigmoid(10·(score−0.5)) × 1/(1+children): winners preferred, under-exploited branches discounted. True greedy hill-climbing does not exist as an option.