edge-or-memory

LLM agents propose volatility, skew and term-structure signals as executable code on a frozen hourly SPY option-surface panel. The harness measures how much of the backtest survives three things: a training-cutoff split, blind versus feedback iteration, and a matched random-search null.

Nothing in the repository is a result. Every table is empty until the runs exist. The numbers in docs/ are pipeline checks on controls. The harness is published first so the bar is visible before anything is measured against it.

What the seal means

Submitted code runs in an isolated subprocess: AST screen, import allowlist, a deny-default sandbox-exec profile, and a whole-prefix look-ahead probe. Evaluation is streamed and expanding: the value traded at the 15:30 entry is the signal at the preceding 14:30 snap, computed by a worker that has never received a later row. A spot check re-computes sampled taped values in fresh processes.

The panel itself is built from private CBOE-derived data and isn’t shipped. A structurally identical synthetic panel is, so everything runs without it. The verification battery checks exact-lag joins, no intraday leak, no daily broadcast, uncrossed legs and stale columns.

Scoring uses three fixed structures and four accounting lines, gross-mid through full-cross, including signed SPY hedge carry.