Skip to content

CLAUDE.md

Latest commit

 

History

History
56 lines (43 loc) · 2.7 KB

CLAUDE.md

File metadata and controls

56 lines (43 loc) · 2.7 KB
Raw

Brain in the Fish

Score any document. Prove every claim. The LLM decomposes documents into OWL knowledge graphs, SPARQL rules mine facts from the graph, and a gate verifies the score is consistent with the evidence.

Relationship to open-ontologies

This project depends on open-ontologies as a library (../open-ontologies). It uses the ontology engine (GraphStore, Reasoner, SPARQL) to load and query document argument graphs. It is NOT a fork — it consumes open-ontologies as a dependency.

Module Overview

Module Purpose
types Core evaluation domain types
ingest PDF ingestion and document ontology generation
criteria Evaluation criteria ontology loading/generation
agent Agent cognitive model and spawning
scoring Independent scoring engine with ReACT loop
debate Multi-round structured debate orchestrator
moderation Consensus moderation and convergence detection
report Evaluation report generation
gate Verdict system — CONFIRMED/FLAGGED/REJECTED based on evidence
rules SPARQL rule mining over argument graphs (8 rules)
argument_graph OWL Turtle → typed argument graph with structural metrics
server MCP server exposing eval_* and eds_* tools
snn Deprecated — legacy scoring, use gate module instead

Pipeline

  1. LLM reads document → produces OWL Turtle (typed nodes with source quotes)
  2. Turtle loads into GraphStore → structural metrics extracted via SPARQL
  3. SPARQL INSERT rules derive facts (StrongClaim, UnsupportedClaim, etc.)
  4. Gate compares LLM holistic score against structural + quality evidence
  5. Verdict: CONFIRMED / FLAGGED / REJECTED

MCP Tools

Tool What it does
eds_feed Push structured evidence into the scorer
eds_score Get score, confidence, low-confidence criteria
eds_challenge Apply lateral inhibition during debate
eds_consensus Check if agents have converged

Build

cargo build

CLI

brain-in-the-fish demo                                    # 3 examples with verdicts
brain-in-the-fish evaluate <document> --intent "..."      # evaluate a document
brain-in-the-fish serve                                   # MCP server for Claude