Really impressive results. The point about the harness mattering more than the model is spot on — we've seen similar patterns in our own work.

One thing that stood out to me is your use of hash-anchored edits + AST-based context selection. We're building something in a similar direction with the Sovereign AI Stack, but with a stronger focus on governance and verification.

Curious — did you run into issues with context drift when using AST queries on very large codebases? We found that combining it with incremental symbol DB updates helped a lot.

Congrats on the results!