Imported from Shiien/research-closure-harness (SKILL.md). Install upstream with npx skills add Shiien/research-closure-harness. Copyright stays with the author.

Research Closure Skill

Objective

Turn open-ended research into small, inspectable, closed units.

The governing loop is:

Frozen claim
→ falsifiable hypothesis
→ bounded test
→ evidence
→ explicit decision
→ written artifact

Start-of-task protocol

Read .research/state.json and .research/claim_graph.json (the claim graph is the engine), then run:

python tools/research_closure.py guard
python tools/claim_graph.py frontier
python tools/research_closure.py next

Report:

Frozen claim:
Session deliverable:
Out of scope:
Ready frontier:      (the probes whose guards are satisfied)
Resolution map:      determined (<verdict>) | open

If there is no frozen claim, do not create a large implementation. Help define a 7–14 day claim first, then author the claim graph (claim_graph.py init --claim "...", then add-variable / add-edge / add-absent / add-probe / add-resolution, then validate) before start-sprint.

New-experiment gate

Before coding, require all fields:

Question:
Hypothesis:
Intervention:
Measurement:
Expected artifact:
Kill criterion:
Time budget:
Claim-graph probe:   (the P-node this experiment runs; must be on the ready frontier)

Reject vague tasks such as:

• “explore whether this works”;
• “try several alternatives”;
• “improve the method”;
• “find a better objective”.

Rewrite them as a single testable claim.

Result gate

Every experiment ends in exactly one decision:

• supported
• falsified
• inconclusive
• terminated

“Inconclusive” must name a specific defect, such as estimator variance, implementation uncertainty, insufficient intervention range, or metric saturation.

Never accept “needs more experiments” as a complete decision.

Scope gate

A new idea is in scope only if success or failure directly changes confidence in the frozen claim.

Otherwise:

python tools/research_closure.py add-idea \
  --idea "..." \
  --reason "Does not test the current frozen claim."

Failure handling

Use this order:

1. implementation correctness;
2. measurement validity;
3. hypothesis falsification;
4. claim narrowing;
5. termination;
6. project pivot.

Do not jump from step 1 to step 6.

Graduation-mode optimization

Prefer:

• existing code over a rewrite;
• a minimal environment over a new benchmark;
• a correct limited theorem over a universal informal claim;
• a complete note over an additional method;
• a decision over continued ambiguity.

End-of-task protocol

Report:

Artifact produced:
Evidence:
Decision:
Next smallest action:

Then update the CLI state or write the relevant decision log.