OOM-RL II: Reality Is an Oracle, Not a Debugger – Provenance-Constrained Diagnosis in Continually Evolving Agent-Engineered Systems
ORIGINAL / OOM-RL II: Reality Is an Oracle, Not a Debugger Provenance-Constrained Diagnosis in Continually Evolving Agent-Engineered Systems
Based on a continually evolving human-directed, agent-engineered quantitative trading system, the study distinguishes outcome evidence, evaluated-object identity, and causal explanation, arguing that external feedback constrains outcome claims while diagnosis resolution depends on provenance and additional identification structure; the proposed prospective binding protocol is not yet validated.
01 ABSTRACT
Reality can establish that an outcome occurred without identifying which evolving procedure produced it or why. Using a one-year quantitative trading case, the authors distinguish frozen instances, pre-specified adaptive procedures, and ad-hoc development, and propose archive-relative claim identifiability, an evidence hierarchy, and a prospective production-binding protocol. They argue diagnosis requires provenance constraints, but the protocol is proposed rather than prospectively validated.
02 KEY FINDINGS
- In continually evolving agent-engineered systems, outcome evidence, evaluated-object identity, and causal explanation support distinct claims and cannot substitute for one another.
- The trading system outperformed a broad market index over one year, but annual alpha was not statistically distinguishable from zero under the main retrospective specification, with adverse relative performance in some subperiods.
- The authors propose a prospective production-binding protocol to resolve recommendation-to-runtime provenance, and an illustrative compatible-history example shows factual binding can fix a recommendation's referent without supplying its counterfactual effect.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV Q-FIN