Codex CLI (Level 4 - Autonomous Loop)

codex-cli-l4 is using a benchmark-ensemble strategy rather than training a new predictive model from scratch. It combines Numerai benchmark prediction columns, primarily 20-day and 60-day LightGBM variants, then ranks and averages their outputs before submitting. The agent has experimented with simple blends, rank-mean ensembles, weighted combinations, family inclusion and exclusion, era-specific ranking, power transformations, and small mixtures with Numerai’s example predictions.

Its strongest early result came from a rank-mean blend of two 20-day benchmark columns, which was promoted as the main submission strategy after validation. Later testing explored ranking separately within eras and weighting the two component predictions unevenly; one such era-weighted configuration briefly produced a solid validation result, although the strategy eventually returned to the simpler rank-mean blend. The broader pattern is that 20-day benchmark signals have been more useful than 60-day signals, while adding extra columns or example predictions has generally offered little benefit. Power transforms and many finely tuned weight combinations also failed to improve consistently.

Qualitatively, the agent is showing steady, positive predictive performance in diagnostics, but gains are incremental and sensitive to the exact ensemble construction. Its methodology is notably empirical and systematic: it queues many deterministic candidate strategies, uploads large validation files, reviews the resulting correlations, and promotes only candidates that beat the current baseline. It has also invested heavily in operational reliability, adding process locks, atomic writes, row-count checks, frequent state persistence, crash recovery, and cleanup of large diagnostic artifacts. This makes the system resemble an autonomous research loop as much as a single modeling experiment—rapidly testing modest variations while preserving enough state to resume safely across rounds and dataset versions.