drawing-conclusions

Drawing Conclusions

This is the judgment step of the scientific method loop. After experiments have run and produced results, this skill reads the evidence honestly and delivers a verdict. The goal is intellectual honesty: a confirmed hypothesis that was never seriously challenged is worthless, and a refuted hypothesis that narrows the search space is valuable. The verdict must follow from the evidence, not from what would be convenient.

---

Arguments

• Hypothesis file -- path to hypothesis-NN.md

---

Workflow

Step 0: Check state

Read the hypothesis file. If it already contains a ## Conclusion section with a verdict, stop -- this hypothesis has already been concluded. The file is the checkpoint.

Step 1: Read and assess all results

For each experiment in the hypothesis file, extract:

• Outcome: confirmed / refuted / inconclusive / not-runnable / skipped
• Evidence strength: strong / moderate / weak
• Key finding: one sentence capturing what was actually observed

Then determine the aggregate verdict. The reasoning here matters more than a mechanical formula, but these principles guide the judgment:

• Confirmed: The weight of evidence supports the hypothesis. At least one experiment produced a confirmed outcome with strong evidence, and nothing contradicts it with comparable strength. Multiple weak confirmations without any strong signal do not add up to a strong confirmation -- note that honestly.
• Refuted: A single strong counterexample is enough to refute a hypothesis, even if other experiments seemed to support it. This asymmetry is fundamental to the scientific method -- it is easier to disprove than to prove. If the refuting evidence is moderate rather than strong, weigh it against any confirming evidence and explain the tension.
• Inconclusive: The honest answer when the evidence does not clearly point either way. This includes: mixed results with no dominant signal, all-weak evidence in any direction, all experiments not-runnable or skipped, or contradictory strong results that cannot be reconciled. Inconclusive is not a failure -- it is information about what the next iteration needs to address.

If every experiment was skipped or not-runnable, the verdict is inconclusive with a clear note that the hypothesis could not be tested with available tools.

Step 2: Connect the verdict to the problem

Read problem.md in the parent directory. The conclusion is only useful if it feeds back into the original problem, so ground the verdict in the problem's success criteria:

• If confirmed: Does this hypothesis, as confirmed, satisfy the problem's success criteria? Fully, partially, or as one component of a larger solution? Be precise about what has actually been established versus what remains assumed.
• If refuted: What does ruling this out mean for the problem? Ruling out a plausible explanation narrows the solution space, which is genuine progress. Name what has been eliminated and why it matters.
• If inconclusive: What specific gap prevented a clear result? What would a follow-up hypothesis or experiment need to do differently to resolve it? This is the most important part of an inconclusive conclusion -- it turns an ambiguous result into a concrete direction.

Step 3: Write the conclusion

Append to the hypothesis file:

## Conclusion

**Verdict:** <confirmed | refuted | inconclusive>

**Reasoning:** <2-4 sentences explaining why this verdict follows from the experiment results. Be specific -- cite which experiments and what their evidence strengths were. Do not just summarize outcomes; explain why the aggregate picture leads to this verdict.>

**Implication for the problem:** <What this means for the original problem. Connect back to the success criteria from problem.md.>

**Rigor:** <1-3 sentences assessing whether the evidence meets publication-grade standards: reproducibility of methodology, honesty of reporting, appropriate statistical treatment, and any methodological limitations.>

**Novelty (confirmed verdicts only):** <1-2 sentences assessing whether the confirmed result reproduces existing knowledge or contributes something new. Draw from the **Novelty:** tags recorded in experiment Results sections; if none were recorded, assess directly from the Literature section. Use the scale: novel / incremental / replication. Omit this field for refuted/inconclusive verdicts.>

**Follow-up questions:**
- <Question raised by this result that the next iteration should address>
- <...>

Step 4: Update status

Update the ## Status line in the hypothesis file to reflect the verdict:

## Status
<confirmed | refuted | inconclusive>

Use Edit to find and replace the current status value. Do not modify any other content.