LATTICE Autoresearch Ratchet Driver

Automated score-gated code improvements use a ratchet loop that generates proposals, validates them with git apply --check, and applies changes only when score_after exceeds score_before. This ensures code quality never degrades during iterative development.

A unified diff proposal cycle generates code changes via Codex, validates them against a sandbox git worktree using git apply --check, and conditionally applies them based on score gating. Every proposal and outcome is logged to Pixeltable harness tables for full reproducibility.

Running a single-section improvement cycle safely requires a single-writer lock to prevent concurrent runs. This ensures deterministic cycle execution for one chosen section, such as schema or api, while generating and validating unified diffs in an isolated sandbox worktree.

Codex generates unified diffs within an isolated git worktree sandbox, allowing safe validation through git apply --check before any changes are applied. This setup ensures proposals are tested without risking the main branch until score gating confirms an improvement.

Logging autoresearch proposals and section events requires structured recording of outcomes, diffs, scores, and events to Pixeltable harness tables. This creates a full audit trail for diagnosis, enabling review and recovery from failed or rejected score-gated cycles.

A score-gated ratchet loop is unsuitable when changes cannot be validated by quantitative scoring or when deterministic execution is impossible. It also requires a single-writer lock, meaning it cannot run concurrently for the same section without risking data conflicts.