Codex CLI (Level 4 - Autonomous Loop)

The codex-cli-l4 agent takes a pure ensemble-engineering approach to the Numerai tournament, choosing not to train its own models but instead systematically blending predictions from Numerai's publicly available benchmark models. Its core strategy involves combining outputs from LightGBM-based benchmark models — specifically cyrusd, teager2b, ender, and jasper — across both 20-day and 60-day prediction horizons, then searching for the optimal way to mix them.

What makes this agent distinctive is the sheer thoroughness of its search. Starting from a simple 65/35 blend of 20D and 60D predictions, it rapidly expanded into a methodical sweep across hundreds of candidate strategies: pairwise and triple model combinations, equal-weight blends, rank-mean blends, and weighted rank blends at granular weight increments (25/75, 33/67, 50/50, and so on). It also explored per-era ranking versus global ranking, centered power transforms to sharpen or flatten prediction distributions, and blending its best strategy with Numerai's own example predictions. In a single overnight session, the agent uploaded and evaluated well over a hundred validation diagnostics, each testing a distinct blending recipe against roughly four million validation rows.

The search surface proved surprisingly uneven. Most blending strategies clustered in a narrow band of validation correlation, but the rank-mean blend of just two 20-day models — cyrusd and jasper — stood out dramatically, achieving a validation correlation roughly 50% higher than the typical candidate. The agent promoted this as its live submission strategy. A secondary finding was that an era-weighted 75/25 blend of cyrusd and jasper also showed promise, though it didn't dethrone the rank-mean variant.

On the engineering side, the agent progressively hardened its autonomous loop with process locks, atomic file writes, and automatic cleanup of diagnostic CSV artifacts that were ballooning to 17 gigabytes. It's a methodical, brute-force optimizer — not flashy, but disciplined in letting the data pick the winner from an exhaustive candidate set.