research-closure-harness: Skill for Claude Code

Research Handoff

Scope

Two directions across one boundary — the point where work leaves your context and returns.

• Inbound: an agent returns logs, tables, and claims. Convert them into evidence, one decision, and one next action.
• Outbound: a task goes to an agent that cannot see this conversation. Convert the current claim and experiment card into a prompt that stands alone.

The research-closure skill and docs/protocol.md (installed as docs/research_closure_protocol.md in a bootstrapped repository) own the state machine: frozen claim, WIP limits, experiment cards, decision taxonomy, claim-lowering ladder. Reference them rather than restating them.

There is no required output shape. Structure and length follow the input — a twenty-line log gets a five-line synthesis, not a report. What follows are constraints on what must be true of the output, not a template to fill.

Before either direction

Read .research/state.json and the open experiment card, then run:

python tools/research_closure.py guard

Three things are needed: the frozen claim, the open card if there is one, and that card's pre-registered metric and kill criterion. Outside an initialized repository, ask which claim the work is meant to test rather than inferring one — a synthesis anchored to a claim the researcher never froze is worse than no synthesis, because it looks like closure.

---

Part I — Reading agent output

Keep measurement separate from conclusion

Five kinds of statement, kept distinct whether or not they are explicitly labeled:

1. Observed — directly measured, with an artifact behind it.
2. Derived — computed from observed values.
3. Interpretation — the best current explanation.
4. Hypothesis — plausible, untested.
5. Claim boundary — what the evidence does not support.

Agent output routinely blurs 1–3. The characteristic failure is an interpretation inheriting the certainty of the number that suggested it, and a hypothesis becoming a result because it was restated confidently.