Quorum's critics need models that can reason about complex artifacts, follow structured evaluation rubrics, use tools (file reading, web search, shell commands), and maintain state across multi-step analysis. Not all models can do this reliably.
Quorum uses a two-tier architecture. Tier 1 handles judgment-heavy roles (supervisor, aggregator, security critic). Tier 2 handles pattern-matching roles (correctness, completeness, architecture critics). See CONFIG_REFERENCE.md for which roles map to which tier.
These models have the reasoning depth and instruction-following precision required for orchestration, synthesis, and high-stakes evaluation.
| Model | Provider | Notes |
|---|---|---|
| Claude Opus 4.6+ | Anthropic | Best overall for Quorum's judgment roles |
| Claude Sonnet 4.6+ | Anthropic | Strong alternative, significantly cheaper |
| GPT-5.2+ | OpenAI | Capable; tool-use patterns differ slightly |
| Gemini 2.0 Ultra+ | Capable; long-context advantage on large artifacts |
These models handle structured evaluation well but lack the depth for synthesis and conflict resolution.
| Model | Provider | Notes |
|---|---|---|
| Claude Sonnet 4.6+ | Anthropic | Recommended Tier 2 default |
| Claude Haiku 4.5+ | Anthropic | Usable for simple rubrics; degrades on nuanced evaluation |
| GPT-4o | OpenAI | Adequate for pattern-matching critics |
| Gemini 2.0 Flash | Fast, cost-effective for coverage checks |
These models consistently fail at one or more critical requirements (tool use, rubric adherence, evidence grounding, state tracking):
- Claude Haiku (pre-4.5) — Insufficient reasoning depth; produces false-precision scores and contradictory findings
- GPT-4 (original) — Tool-use reliability too low for critic pipelines
- Llama 70B and below — Open models at this scale lack the instruction-following precision Quorum requires
- Mixtral, Command R — Similar limitations to Llama at comparable scales
Why not smaller models? Quorum critics must: (1) follow a structured rubric item-by-item, (2) cite specific evidence for every finding, (3) assign calibrated severity ratings, and (4) avoid hallucinating findings that aren't grounded in the artifact. Models below ~100B parameters reliably fail at #2 and #4.
Set your model tiers in quorum-config.yaml:
models:
tier_1: anthropic/claude-sonnet-4-6 # Judgment roles (supervisor, aggregator, security)
tier_2: anthropic/claude-sonnet-4-6 # Evaluation roles (correctness, completeness, architecture)For per-critic overrides, see CONFIG_REFERENCE.md.
First-run behavior: If no config exists, Quorum auto-detects your available model, asks two setup questions (preferred tier 1 and tier 2 models), writes the config, and proceeds immediately. Zero-config users get sensible defaults.
Rough per-run costs at default (standard) depth with 6 critics:
| Configuration | Tier 1 | Tier 2 | Estimated Cost |
|---|---|---|---|
| Budget | Sonnet | Haiku | ~$0.15–0.30 |
| Balanced | Sonnet | Sonnet | ~$0.30–0.60 |
| Premium | Opus | Sonnet | ~$1.00–2.50 |
Costs scale roughly linearly with critic count and artifact size. The quick depth profile (2 critics, no fix rounds) is the cheapest option for iteration.
Quorum is designed for OpenClaw and tested primarily on that platform. The architecture is intentionally model-agnostic — the spec and rubric format don't assume a specific provider. Cross-platform compatibility with other agent frameworks is under active research.
If you're running Quorum on a non-OpenClaw platform and want to share results, open an issue.
⚖️ LICENSE — Not part of the operational specification above. This file is part of Quorum. Copyright 2026 SharedIntellect. MIT License. See LICENSE for full terms.