bidirectional-differential

You can audit EARS and code drift by running two parallel Claude sessions that reconstruct specifications from code and code from specifications, then classify mismatches to surface alignment gaps.

The bidirectional differential method checks specification coherence by spawning A- and B-direction sessions to reconstruct code from EARS and vice versa, classifying drift to resolve inputs and surface intent mismatches.

To run a differential audit, invoke the /differential-audit command with one or more EARS IDs, which triggers a six-step protocol to resolve scopes, spawn parallel sessions, classify drift, and write audit artifacts.

Yes, you can scope audits to a specific arrow and EARS by mapping user-described scopes to arrow overlays, EARS IDs, and segment metadata to drive targeted parallel session runs.

The bidirectional differential audit stores per-EARS audit records in the docs/arrows/_experiments/bidirectional-differential directory for traceability and review.

A limitation of using parallel sessions is that the audit requires resolving inputs and stripping identifiers before spawning sessions, meaning unresolved or poorly scoped arrow overlays may affect drift classification accuracy.