Hypothesis Generation

Hypothesis Generation

Generate evidence-bounded engineering hypotheses. This skill freezes the observation, frames the question, establishes evidence boundaries, generates rival candidates, and derives discriminating predictions. It never scores, ranks, or selects hypotheses — that is a human decision.

Invocation

/hypothesis-generation 

Include: the observed phenomenon, context, and any relevant engineering domain (materials, structures, thermal, etc.).

Methodology

1. Freeze the observation — write the exact observation before interpreting. Distinguish what was measured from what is inferred. Timestamp the observation date.

1. Frame the research question — restate as a falsifiable engineering question. Identify the claim type: descriptive, associational, predictive, causal, or mechanistic.

1. Establish evidence boundary — document: search date, databases consulted, query terms, inclusion/exclusion criteria, limitations. "Not located within documented search boundary" — never "no prior work exists".

1. Generate rival hypotheses — produce 3-5 candidates from DIFFERENT explanatory classes (e.g., material defect, design error, environmental, manufacturing process, maintenance). Each must be falsifiable. No two hypotheses from the same class.

1. Declare claim type and estimand — specify what would be measured, the expected direction/magnitude, and the population/context where it applies.

1. Derive discriminating predictions — for each hypothesis, state: condition, observable, expected pattern, falsifier (what result would be incompatible), and rival contrast (how does this prediction differ from the others?).

1. Prevent HARKing — timestamp the generation. Preserve this document as the pre-registration record. Any deviation in testing must be reported as a deviation, not hidden.

Output

Inline Summary (chat response)

• Observation (frozen, dated)
• Research question + claim type
• Evidence boundary (date, sources, queries, limitations)
• 3-5 rival hypotheses with class labels
• For each: discriminating prediction + falsifier

Full Record (saved to disk)

Save to outputs/hypothesis/.md:

# Hypothesis Generation: 

## Observation (Frozen)
- **Date observed:** 
- **Phenomenon:** 
- **Distinguish:** measured vs inferred
- **Context:** 

## Research Question
- **Question:** 
- **Claim type:** descriptive / associational / predictive / causal / mechanistic

## Evidence Boundary
- **Search date:** 
- **Databases:** 
- **Query terms:** 
- **Inclusion:** 
- **Exclusion:** 
- **Limitations:** 

## Rival Hypotheses

### H1:  ()
- **Mechanism:** 
- **Prediction:** 
- **Falsifier:** 
- **Rival contrast:** 

(repeat for H2, H3...)

## Discriminating Tests
- **Test that separates H1 from H2:** 
- **Test that separates H1 from H3:** 
- ...

## Pre-registration Record
- **Generated:** 
- **Status:** candidate (never scored/selected by tool)
- **Deviations:** none yet

Scope and Boundaries

• This skill generates hypothesis candidates — it does NOT score, rank, or select. Status must remain candidate.
• Research-only, not for final engineering sign-off. Hypothesis selection requires human engineering judgment.
• Never fabricate evidence to support a hypothesis. If a hypothesis lacks evidence, mark it unverified, not inferred.
• Evidence quality: see references/evidence-quality-tiers.md.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: adeerkhan
• Source: adeerkhan/vitruvius
• License: MIT

Install and usage instructions live in the source repository linked above.