Quorum — Multi-Agent Validation

Quorum validates AI agent outputs by spawning multiple independent critics that evaluate artifacts against rubrics. Every criticism must cite evidence. You get a structured verdict.

Quick Start

Clone the repository and install:

git clone https://github.com/SharedIntellect/quorum.git
cd quorum/reference-implementation
pip install -r requirements.txt

Run a quorum check on any file:

python -m quorum.cli run --target <path-to-artifact> --rubric <rubric-name>

Built-in Rubrics

• research-synthesis — Research reports, literature reviews, technical analyses
• agent-config — Agent configurations, YAML specs, system prompts
• python-code — Python source files (25 criteria, PC-001–PC-025; auto-detected on .py files)

Depth Profiles

• quick — 2 critics (correctness, completeness) + pre-screen, ~5-10 min
• standard — 6 critics (+ security, code_hygiene, cross_consistency†, tester) + pre-screen, ~15-30 min (default)
• thorough — all 6 shipped critics + pre-screen + fix loops, ~30-60 min

†Cross-Consistency requires --relationships flag with a relationships manifest.

All depth profiles include the deterministic pre-screen (10 checks: credentials, PII, syntax errors, broken links, TODOs, and more) before any LLM critic runs.

Examples

# Validate a research report
quorum run --target my-report.md --rubric research-synthesis

# Quick check (faster, fewer critics)
quorum run --target my-report.md --rubric research-synthesis --depth quick

# Batch: validate all markdown files in a directory
quorum run --target ./docs/ --pattern "*.md" --rubric research-synthesis

# Cross-artifact consistency check
quorum run --target ./src/ --relationships quorum-relationships.yaml --depth standard

# Use a custom rubric
quorum run --target my-spec.md --rubric ./my-rubric.json

# List available rubrics
quorum rubrics list

# Initialize config interactively
quorum config init

Configuration

On first run, Quorum prompts for your preferred models and writes quorum-config.yaml. You can also create it manually:

models:
  tier_1: anthropic/claude-sonnet-4-6    # Judgment roles
  tier_2: anthropic/claude-sonnet-4-6    # Evaluation roles
depth: standard

Set your API key:

export ANTHROPIC_API_KEY=sk-ant-...
# or
export OPENAI_API_KEY=sk-...

Output

Quorum produces a structured verdict:

• PASS — No significant issues found
• PASS_WITH_NOTES — Minor issues, artifact is usable
• REVISE — High/critical issues that need rework before proceeding
• REJECT — Unfixable problems; restart required

Exit codes: 0 = PASS/PASS_WITH_NOTES, 1 = error, 2 = REVISE/REJECT.

Each finding includes: severity (CRITICAL/HIGH/MEDIUM/LOW), evidence citations pointing to specific locations in the artifact, and remediation suggestions. The run directory contains prescreen.json, per-critic finding JSONs, verdict.json, and a human-readable report.md.

More Information

• SPEC.md — Full architectural specification
• MODEL_REQUIREMENTS.md — Supported models and tiers
• CONFIG_REFERENCE.md — All configuration options
• FOR_BEGINNERS.md — New to agent validation? Start here

⚖️ LICENSE — Not part of the operational specification above. This file is part of Quorum. Copyright 2026 SharedIntellect. MIT License. See LICENSE for full terms.