Reasons Forge

A belief forge built on Doyle's (1979) Truth Maintenance System. Ingest domain sources, extract beliefs, derive new knowledge, and maintain consistency — all backed by automatic retraction cascades and dependency-directed backtracking.

What It Does

Reasons Forge analyzes domain-specific sources (codebases, issue trackers, product data, documents) and builds a dependency network of beliefs. The TMS engine tracks which beliefs are IN (believed) or OUT (retracted), automatically propagating changes when evidence shifts.

Core TMS engine:

• Nodes with SL/CP justifications and non-monotonic reasoning (outlist)
• Retraction cascades and automatic restoration
• Nogoods with dependency-directed backtracking
• Dialectical argumentation (challenge/defend)
• Merkle integrity verification

Domain-specific forges:

• document — PDFs, markdown, and code files with section-based chunking
• code — Codebase architecture, patterns, invariants, and technical debt
• project — Team velocity, milestone health, and delivery risks from issue trackers
• product — Feature readiness, user experience, and product-market fit
• meta — Cross-domain reasoning across expert belief networks
• diary — Extract beliefs from diary entries and session notes

LLM-powered analysis:

• derive — generate new beliefs from existing ones
• review-beliefs — validate derived beliefs
• contradictions — detect contradictions between IN beliefs
• verify — re-examine beliefs against source documents
• repair — fix invalid beliefs (search-and-link, soften, abandon)
• deduplicate — find and merge duplicate beliefs
• ask — natural language questions over the belief network

Install

pip install reasonsforge

Or with uv:

uv tool install reasonsforge

For forge features (PDF chunking, issue tracker sources):

pip install reasonsforge[forge]

Quick Start

Core TMS

# Initialize database
reasonsforge init

# Add premises
reasonsforge add source-uses-langgraph "Source code uses LangGraph" --source "src/graph.py"
reasonsforge add graph-has-cycles "Graph contains cycles"

# Add derived nodes with SL justifications
reasonsforge add topology-is-static "Graph topology is static" --sl source-uses-langgraph
reasonsforge add no-runtime-modification "No runtime graph modification" --sl topology-is-static

# See what's believed
reasonsforge status
#   [+] graph-has-cycles: Graph contains cycles  (premise)
#   [+] no-runtime-modification: No runtime graph modification  (1 justification)
#   [+] source-uses-langgraph: Source code uses LangGraph  (premise)
#   [+] topology-is-static: Graph topology is static  (1 justification)
#
# 4/4 IN

# Retract a premise — cascade propagates
reasonsforge retract source-uses-langgraph
# Retracted: source-uses-langgraph, topology-is-static, no-runtime-modification

# Restore — dependents come back automatically
reasonsforge assert source-uses-langgraph
# Asserted: source-uses-langgraph, topology-is-static, no-runtime-modification

# Record a contradiction
reasonsforge add graph-is-dynamic "Graph is dynamically modified"
reasonsforge nogood topology-is-static graph-is-dynamic
# Recorded nogood-001: topology-is-static, graph-is-dynamic
# Retracted: graph-is-dynamic

# Explain why a node is IN or OUT
reasonsforge explain no-runtime-modification

# Non-monotonic reasoning: believe X unless Y
reasonsforge add default-approx "Newtonian approximation holds" --unless strong-field

# LLM-powered derivation
reasonsforge derive --sample -m claude

# Export
reasonsforge export -o network.json
reasonsforge export-markdown -o beliefs.md

Forges

Forges are domain-specific pipelines that automate the full cycle: ingest sources, summarize, extract beliefs, derive, review, and repair.

# Analyze a codebase
reasonsforge forge code --repo /path/to/repo

# Analyze a codebase step by step
reasonsforge forge code scan --repo /path/to/repo
reasonsforge forge code explore --repo /path/to/repo --loop 50
reasonsforge forge code propose-beliefs --auto
reasonsforge forge code derive --exhaust --auto

# Analyze project state from GitHub issues
reasonsforge forge project --github owner/repo

# Analyze product data from an issue tracker
reasonsforge forge product --github owner/repo

# Ingest product documents
reasonsforge forge product ingest docs/ --glob-pattern "**/*.md"

# Analyze a PDF (chunked by section, then summarized)
reasonsforge forge document --pdf paper.pdf

# Cross-domain synthesis across expert belief networks
reasonsforge forge meta init code=/path/to/code-expert project=/path/to/project-expert
reasonsforge forge meta import
reasonsforge forge meta derive --auto
reasonsforge forge meta contradictions --auto
reasonsforge forge meta summary

# Full meta pipeline in one command
reasonsforge forge meta update

Commands

Core TMS

Dialectical

Search & Query

LLM Analysis

Import & Export

Forge Commands

Forge Pipeline Steps

Tests

uv run --extra test pytest tests/ -v

1,732 tests covering propagation, retraction cascades, restoration, multiple justifications, diamond dependencies, nogoods, dependency-directed backtracking, non-monotonic justifications, dialectical argumentation, Merkle integrity, SQLite round-trips, import/export, and more.

References

Doyle, J. (1979). A Truth Maintenance System. Artificial Intelligence, 12(3), 231–272.