# Run Hypothesis

> Run a 6-phase scientific hypothesis trajectory on any open problem — from rough intuition to a formally structured, mathematically testable framework. Use when the user wants to formalize a hypothesis, prove a theory works, improve an existing model (e.g. Black-Scholes, CAPM, relativity, evolution), or explore any open research question across any domain.

- Skill: `justrach/run-hypothesis` (Agent Skill)
- Install (CLI): `npx skillmds@latest add justrach/run-hypothesis`
- Raw SKILL.md: https://api.skillmd.com/api/skills/justrach/run-hypothesis/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: justrach (https://skillmd.com/u/justrach)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/justrach/run-hypothesis

---


# Hypothesis Trajectory Engine

This skill invokes the `run_hypothesis` tool from the AgentLab MCP server (`mcp__agentlab__run_hypothesis`), which runs a 6-phase multi-agent reasoning pipeline that turns rough intuitions into formal, testable hypotheses.

## Pipeline

| Phase | Agent | Mode | Runs |
|-------|-------|------|------|
| 0 | `formalizer` — maps every informal term to a formal mathematical object | deep (Opus) | solo |
| 1 | `existence_prover` — proves the critical condition/moment exists (EVT, IVT) | smart | parallel |
| 2 | `constraint_closer` — writes the full conservation/constraint balance sheet | smart | parallel |
| 3 | `gap_finder` — quantifies the gap between current theory and observations | smart | parallel |
| 4 | `mechanism_designer` — designs a mechanism that closes the gap | deep (Opus) | sequential |
| 5 | `deeper_connector` — connects mechanism to deeper principles, lists predictions | smart | sequential |
| ∞ | `synthesizer` — structured final report written to `workspace_dir/final_report.md` | smart | sequential |

Phases 1–3 run in parallel threads. Each phase reads prior phase outputs. If a phase writes a `compute.py`, the runner executes it with `python3` (numpy, scipy, sympy available) and injects results into the next phase.

## How to invoke

**Directly:**
```
/run-hypothesis prove that special relativity works
/run-hypothesis make Black-Scholes better
/run-hypothesis why do L-amino acids dominate in biology
/run-hypothesis improve the SEIR epidemic model
```

**Claude auto-invokes this skill** when the user asks to:
- Prove or verify a theory
- Formalize a hypothesis or rough idea
- Improve an existing model or framework
- Explore an open research question
- Find the mechanism behind an observed phenomenon

## Instructions

1. Parse `$ARGUMENTS` to extract the problem statement and any domain hint (e.g. "quantitative finance", "theoretical physics", "biology", "economics")
2. If no domain is specified, infer it from the problem content
3. Call `mcp__agentlab__run_hypothesis` with:
   - `problem`: the full problem statement from `$ARGUMENTS`
   - `domain`: inferred or specified domain
   - `workspace_dir`: `hypothesis_workspace` (relative, in current repo)
4. The tool returns a JSON object with phase outputs and a synthesis section
5. Present the synthesis report clearly to the user, noting which workspace files were written

## Output

The tool writes phase findings to `hypothesis_workspace/phaseN_<name>/findings.md` and a final report to `hypothesis_workspace/final_report.md`. Reference these files if the user wants to dig into a specific phase.

## Domain examples

| Problem type | Domain hint |
|---|---|
| Options pricing, volatility surface | `quantitative finance` |
| Particle physics, cosmology | `theoretical physics` |
| Protein folding, amino acids | `biochemistry` |
| Epidemic modeling | `epidemiology` |
| Market microstructure | `financial economics` |
| Neural synchrony, consciousness | `computational neuroscience` |
| Economic tipping points | `complexity economics` |