nx_artloop.nx

nx_artloop.nx -- AL1: THE RECURSIVE ART-LOOP BEAT (operator 2026-08-09: "i want nishi ecosystem to gain these functionalities as a recursive learning setup so that it can keep growing" + "you need to have benchmarks and oracles so you arent just generating junk"). One beat = one unattended cycle of the emit-score-look loop, clock-plane scheduled: fork the DEPLOYED engine `sweep` (6 procedurally-derived material families rendered, scored, contact-sheeted to the served site by the engine itself), then bank the cycle APPEND-ONLY. ORACLES -- junk cannot pass, and a drifting instrument cannot silently bless anything: O1 INSTRUMENT SANITY: the caricature controls must hold their banked bands EVERY cycle -- flat_control < 300 (TOY) and squares_control in [598,658] (banked 628); plus a LIVE benchmark re-score: breeders1 must hold [779,799] (banked 789). Any drift => the CRITIC moved => RED, and no other number this cycle is believable (stale-instrument law). O2 JUNK FLOOR: the shipped-default variant must clear squares_control + 120. A frame that cannot beat a literal random-squares mosaic by a real margin is junk => RED. O3 RATCHET: best= is the historical max of the default variant's STRUCTURE score (the composite is a structure floor, NEVER quality -- caricature conviction 2026-08-09); dropping below the banked best is a regression => RED. The record still lands FIRST: report, never hide. JOURNAL: knowledge/status/artloop.jrnl, append-only via full-rewrite to .tmp + sys_renameat (atomic -- a crash never truncates history). One line per cycle, RAW sweep receipt included (bank the raw record; derived analysis is downstream's job). PICKS stay judged by eyes (AL2). license_tier: ORIGINAL expect_exit: 0