Reality Check

2026-05-14 · source · source ↗ · raw capture

A framework for rigorous, systematic analysis of claims, sources, predictions, and argument chains. Unified knowledge base with structured epistemic tooling.

v0.3.4, 458 tests. Python 91.4%, Jinja 5.3%, Shell 2.4%. Apache 2.0 license.

Overview

"With so many hot takes, plausible theories, misinformation, and AI-generated content, sometimes, you need a realitycheck."

Reality Check enables users to build a unified knowledge base with: - Claim registry tracking evidence levels and credence scores - Source analysis using a 3-stage methodology (descriptive → evaluative → dialectical) - Evidence linking connecting claims to sources with strength ratings - Reasoning trails documenting epistemic provenance - Prediction tracking with falsification criteria - Argument chain mapping for logical dependencies - Semantic search across the knowledge base

Technical Architecture

Installation

pip install realitycheck
# or
uv pip install realitycheck

From source:

git clone https://github.com/lhl/realitycheck.git
cd realitycheck
uv sync

Quick Start

  1. Initialize project: rc-db init-project
  2. Set environment: export REALITYCHECK_DATA="data/realitycheck.lance"
  3. Add claims and sources via CLI
  4. Search semantically across the database

Integrations

Taxonomy

Claim Types: Fact [F], Theory [T], Hypothesis [H], Prediction [P], Assumption [A], Counterfactual [C], Speculation [S], Contradiction [X]

Evidence Levels: E1 (Strong Empirical) through E6 (Unsupported)

Domain Codes: TECH, LABOR, ECON, GOV, SOC, RESOURCE, TRANS, GEO, INST, RISK, META

Design Philosophy

Recommends a single unified knowledge base rather than topic-specific databases: "claims build on each other across domains (AI claims inform economics claims)."

Links