nx_final_rung_gate.nx

nx_final_rung_gate.nx -- THE FINAL RUNG (operator 2026-06-28): "have the system build something as good if not better than you [Claude] as the final rung." Grounded by the researcher survey: the systems that beat Eurisko (FunSearch/AlphaEvolve/CodeEvolve/LILO) all put a frontier LLM IN THE LOOP -- they are Claude-amplified, NOT self-reliant. The honest, harder Nishi test: on our KNOWN GOODS, does the system's NO-LLM synthesis MATCH a Claude-authored reference? This is a measured HEAD-TO-HEAD: - CLAUDE side: a reference implementation I hand-author (baked inline = "what Claude would write"). - SYSTEM side: ncf_synth synthesizes the program autonomously + GOD-BUILDS it (nx_cc->nxasm, NO Claude in the synthesis loop), runs on HELD-OUT inputs. - MATCH iff system_produced == claude_reference == ground_truth on held-out. T1-T3: on 3 known-good tasks (conditional/polynomial/recurrence) the system MATCHES Claude, no LLM in loop. T4: HONEST BOUNDARY -- an out-of-grammar task (a(n)=a(n-1)^2) -> system ESCALATES (Claude still needed). T5: FINAL-RUNG VERDICT -- as-good-as-Claude on the covered class AND more self-reliant (no LLM, deterministic, god-build-verified); honest scope: the grammar itself is still Claude-authored = the remaining gap. expect_exit: 0 license_tier: ORIGINAL