Harvey LAB: In-House Contracting (held-out)
Professional Work · 2026-06-12
Harvey's Legal Agent Benchmark: In-House Contracting extends LAB with 500 agentic contract drafting, review and negotiation tasks, each freezing a negotiation at one point with the contract, playbooks, email context and prior redlines. This row is Harvey's 50-task held-out set run with Harvey's upstream LAB harness and dual-judge profile; a task-attempt earns 1 when both judges find every rubric criterion passed, 0.5 when one does, and 0 otherwise, averaged over three attempts per task.
Top models (higher is better)
| Model | Score |
|---|---|
| Grok 4.6 | 10.0 |
| Kimi K3 | 8.3 |
| Gemini 3.8 Flash | 6.3 |
| GLM-5.3 | 6.0 |
| Ember-1 | 5.7 |
| GPT-6 Astra | 5.3 |
| Claude Opus 5.5 | 3.0 |
| GPT-5.6 Sol | 2.0 |
| Claude Opus 5 | 2.0 |
| GPT-6 Sol | 1.3 |
| DeepSeek V4 Pro 0813 | 1.0 |
| DeepSeek V4.1 Flash | 0.3 |
| GPT-6 Luna | 0.3 |