Hypothesis Generation

Generate falsifiable, testable research hypotheses from the user's notes and research content.

Workflow

1. Read available content — scan the notes, documents, or files the user provides
2. Identify themes — find recurring concepts, claims, relationships, and open questions
3. Generate hypotheses — produce 4–6 specific, falsifiable hypotheses
4. Assign confidence — rate each hypothesis HIGH / MEDIUM / LOW based on evidence in the content
5. Format output — present in structured markdown

Output Format

# 🧪 Hypothesis Analysis

## Summary
[2–3 sentences describing the main themes and what the hypotheses cover]

## Generated Hypotheses

### 🔴 [Hypothesis Title] — HIGH confidence

**Hypothesis:** [Specific, falsifiable statement with measurable prediction]

**Rationale:** [Which sources/notes support this, with specific references]

**Testable:** Yes | **Confidence:** HIGH

---

### 🟡 [Hypothesis Title] — MEDIUM confidence
...

Confidence levels

• 🔴 HIGH — directly supported by multiple sources in the content
• 🟡 MEDIUM — partially supported or requires inference across sources
• 🟢 LOW — speculative but worth investigating; limited direct support

Quality criteria for good hypotheses

• Falsifiable: Can be proven wrong — avoid "X may affect Y"
• Specific: Mentions measurable variables, not vague concepts
• Grounded: Traceable to actual content provided, not general knowledge
• Novel: Connects ideas across sources rather than restating the obvious

Example

Input: Notes on predictive coding and synaptic plasticity

Good hypothesis:

"Precision-weighted prediction errors in the Rao and Ballard model are encoded through spike-timing-dependent plasticity (STDP) in the visual cortex, such that altering STDP timing windows disrupts receptive field formation."

Poor hypothesis:

"Synaptic plasticity is important for learning." ← not falsifiable, too vague