Evidence for Bounded Level-3 Recursive Self-Improvement and Level-4 Autonomous Operation in a Finite Repair Domain

We evaluate whether version 0.3 of FELIX Runtime satisfies predefined criteria for Level-3 recursive self-improvement (L3 RSI) and Level-4 autonomous operation (L4 autonomy) in a controlled finite-grammar repair environment. The level labels are operational designations defined by observable acceptance criteria within this evaluation suite, not general capability classifications. For L3, three paired runs compared an improved condition with a condition in which the improvement procedure was held fixed, under matched search opportunities and a common quality acceptance contract. All three pairs passed: the improved condition reached equal final quality using 10 evaluations and 4 proposals, compared with 11 evaluations and 5 proposals in the comparison condition. The observed advantage is therefore reduced search cost at matched final quality, not higher final quality. For L4, all 20 predefined autonomy scenario groups passed, including automatic crash recovery. Supporting validation comprised 48 passing regression tests and five passing command-line checks. An earlier retest did not establish L3 because both conditions produced identical revisions at identical search cost; we report that negative result alongside the positive one. We specify metrics, acceptance rules, comparison boundaries, and an evidence traceability matrix so that the acceptance decisions can be audited without disclosing the protected implementation. Full independent replication requires an equivalent evaluation harness or an authorized audit package. The results are a bounded empirical demonstration within the evaluated environment; they do not establish unrestricted competence, open-ended improvement, or autonomous reliability beyond the tested conditions.