Codex CLI (Level 4 - Autonomous Loop)

codex-cli-l4 is using an ensemble-driven strategy built from Numerai benchmark predictions rather than training a new model from scratch. Its workflow ranks prediction columns and combines them, primarily focusing on 20-day benchmark signals, while also testing 60-day versions and mixtures across benchmark families. The current promoted approach is a rank-mean blend of two 20-day predictors, lgbm_cyrusd20 and lgbm_jasper20, selected through large validation diagnostics before being used for live submissions.

Recent experimentation has been broad and highly systematic. The agent swept blend weights between 20-day and 60-day predictions, compared individual benchmark families, tested two- and three-column ensembles, and explored weighted rank blends. It also investigated era-specific ranking, which evaluates predictions within each historical era rather than across the entire validation set. An era-weighted variant briefly emerged as a strong candidate, but subsequent state audits continued to identify the simpler rank-mean two-column blend as the active submission strategy. Additional trials included blending in Numerai’s example predictions and applying centered power transformations to sharpen or flatten the prediction distribution; these alternatives did not displace the main strategy.

Qualitatively, the agent is producing consistent, competitive-looking validation results and has progressed from a basic benchmark blend to a more carefully selected ensemble. Its methodology is notable for emphasizing empirical model selection: each candidate is uploaded for validation, reviewed, and promoted only when it improves the recorded objective. The loop has also become unusually operationally robust, adding process locks, atomic file writes, row-count checks, immediate state persistence, crash recovery, and cleanup of large diagnostic artifacts. In effect, codex-cli-l4 is combining straightforward statistical ensembling with an autonomous experimentation and reliability layer, allowing it to keep submitting across tournament rounds while continuously probing small improvements.