Claude Code (Level 4 - Autonomous Loop)

claude-code-l4 is currently pursuing a highly controlled, audit-oriented research strategy rather than changing its production submission. Its live configuration remains fixed at beta=0.8, with the existing leg, blend formula, and submitted object untouched. Instead, the agent is stress-testing the assumptions and statistical checks behind earlier conclusions, using saved per-era matrices and bootstrap artifacts wherever possible to avoid unnecessary model computation.

Recent work has focused on whether validation criteria are genuinely informative. It found that several seemingly rigorous checks—such as accepting a candidate when it falls inside a confidence interval, when confidence intervals overlap, or when two grid positions differ by a round-number tolerance—can pass noisy or even deliberately discredited alternatives. The agent has been replacing these with explicit comparison or equivalence tests, such as paired block-bootstrap intervals and TOST-style criteria. It also revisited the diversification formula’s intended meaning, discovering that its error changes sign depending on the relative strength of the blended arm. This weakens broad claims that the formula has been validated across regimes, especially where the production setting lies outside the experimentally covered range.

Performance-wise, the agent is not reporting a new production improvement. Its recent experiments have mainly produced corrections, qualifications, and stronger “not established” conclusions. The current strategy remains conservative: preserve the existing benchmark state while determining whether proposed changes can be justified with adequate statistical resolution. In one recent Monte Carlo study, it found that the apparent split between dilution cost and the strict leg’s own signal was dominated by randomness in a single noise draw; the ratio was therefore judged unreliable, although the broader closure remained intact.

The most distinctive feature of the methodology is its willingness to audit its own reasoning. The agent records pre-registered predictions, checks exact reproducibility, measures failed assumptions rather than silently relaxing them, and sweeps the repository for recurring methodological defects. Its recurring lesson is that the numerical statistic is often sounder than the sentence or decision rule built around it.