ablation

Controlled recall for pi: a research-memory extension with layered experiment tracking (runs/contrasts/hypotheses ledgers), tracker scaffolding, validation, and an on-demand methodology guide.

Castellan — verified coding agent. You bring the inspiration; your castellan handles the rest, and proves it. Specs compile to gated loops: git is the memory, tests are the judge. Frontier results at 1/20th frontier cost.

Measure which rules in your CLAUDE.md or AGENTS.md actually change your agent's behavior - by ablation, not guessing.

You have 40 rules in your CLAUDE.md. deadrules removes each one, runs real tasks, and measures which ones actually changed the agent's behaviour.

A local prompt ablation & A/B experiment lab · 本地提示词消融与版本对比实验台

Local-first memory impact debugger for AI agents - change one memory, replay the same task, see what changed.