FRSIS: A Falsifiable Recursive Self Improving System — An FSD-Scale Software Stack

This deposit is a follow-up to Empirical Recursive Self-Improvement on Coding Agents: L1/L2 Confirmation with Negative L3 Results (Zenodo DOI 10.5281/zenodo.22318825). Where that work established a controlled empirical RSI ladder (L1 Weak RSI / L2 improver RSI / L3 meta-RSI) on coding agents, the present manuscript treats the RSI factor as a singular software system: FRSIS (Falsifiable Recursive Self Improving System) — an FSD-scale, software-only vehicle embedding that ladder as substrate, plus SAE/Tesla-style lab conditional automation (goals, continuous drive, PersistenceGate), under a black-box research stance. On the 2026-09-07 coverage battery (Asia/Taipei; no LLM; dry_run default except one explicit persist proof), we report: Felix L2 on CodeHost at 6/6 seeds; CodingAgentHost L2 at 3/6 under a harder discrete basin; dual-pass bitwise-identical arm means; SAE/Tesla-style L3 autonomy on CodeHost with hit_l2 and cycles goals met, plus real --persist commits that then disarmed; Felix RSI L3 meta not demonstrated (L3_RSI_MISS; MetaHitRules unchanged). We do not claim L3_RSI_HIT, AGI, or production vehicle autonomy. Architecture internals are not revealed; capabilities, HitRules, scores, seeds, and pass/fail are.