Handoff Benchmark

This first measure checks the technical prerequisite for a handoff: work saved in Claude, Cursor, or Codex must appear in another client through palim_resume, without a session ID, transcript, or manual copy-and-paste.

Reproducible run

The runner and its three synthetic fixtures are in the public repository. It verifies that palim_resume returns the latest thread and its open task, then reports median response size, estimated tokens, and in-process retrieval time.

cd palim-mcp
npm run eval:handoff

The current baseline returns 3 of 3 handoffs, with a median output of 874 characters or an estimated 219 tokens, and 0.02 ms in-process retrieval time.

What this measures

This is a deterministic transport and context-budget measurement. It does not claim that a language model completed a task correctly, and it does not compare model answers without Palim. That next stage needs a fixed model version, prompt, grading rubric, and independently reviewed test sessions.

LoCoMo is a separate question-answering measurement. Palim will only publish a LoCoMo number after the dataset, environment, and complete run are documented.