--- pretty_name: ContractBench license: cc-by-4.0 task_categories: - text-generation language: - en tags: - legal - contract-redlining - docx - tracked-changes - llm-judge - agents - harbor - benchmark size_categories: - n<1K configs: - config_name: default data_files: - split: test path: data/tasks.parquet --- # ContractBench ContractBench measures how well an AI agent reviews and redlines a commercial contract the way an in-house attorney would: deciding where to intervene, what to change, what to leave alone, when to ask the counterparty a question, and when to escalate a point to its own side instead of deciding it. Every task is a real Word document, a set of grounding documents (the represented side's commercial context and, in most scenarios, its playbook), and a brief. The agent works directly in the `.docx` with tracked changes and comments, and is graded against pass-or-fail criteria written by the attorney who reviewed the same contract. **Benchmark Report**: https://www.ivo.ai/research/ivo-micro1-contractbench and https://www.micro1.ai/benchmark/ivo-micro1-contract-bench **Methodology and Runner**: https://github.com/micro1-research/contract-bench This repository holds the **public set** of the benchmark: 210 tasks from 21 scenarios, each a complete, runnable [Harbor](https://github.com/laude-institute/harbor) task. A private set of 140 tasks from 14 further scenarios is held out; the leaderboard below reports model scores on both sets. ## Leaderboard | Model | Private score (140 held-out tasks) | Public score (210 tasks) | Full benchmark (350 tasks) | Completion rate (all 350 tasks) | | --- | ---: | ---: | ---: | ---: | | Grok 4.7 | 58.3% | 58.3% | 58.3% | 99.5% | | Opus 5.5 | 58.3% | 58.0% | 58.1% | 80.8% | | Kimi K3 | 56.7% | 57.4% | 57.1% | 100.0% | | Gemini 3.8 Flash | 55.4% | 54.8% | 55.0% | 99.8% | | GLM-5.3 | 55.4% | 54.8% | 55.0% | 95.2% | | DeepSeek V4 Pro | 49.1% | 47.9% | 48.4% | 99.5% | | GPT-6 Astra | 45.3% | 45.8% | 45.6% | 100.0% | ## Design A scenario is one contract negotiation: a contract, the two parties, the represented side's commercial context and, in most scenarios, its playbook. Each scenario yields ten tasks. - **Turn 1 (opening redline).** Five attorneys independently reviewed the same clean contract for the same side, each producing a redline and a rubric set. The five tasks share one input; a model's Turn 1 redline is graded against all five rubric sets, so the benchmark scores it against the range of defensible attorney approaches rather than one reference. - **Turn 2 (response).** Five more tasks start from one of those Turn 1 redlines, now seen from the other side: the model represents the counterparty and responds to an existing markup, accepting, rejecting, countering or introducing new edits. | Dimension | Coverage in this set | | --- | --- | | Agreement types | MSA: 4 scenarios; IP Licensing: 5 scenarios; Research Partnership: 4 scenarios; Marketing Agreement: 4 scenarios; Commercial Lease: 4 scenarios | | Playbook | none: 3; single-agreement: 11; multi-contract: 7 | | Jurisdiction | US: 18; UK: 3 | | Tasks | 210 (105 Turn 1, 105 Turn 2) | | Rubric items | 3,521 | ## What a task directory contains ``` tasks// task.toml Harbor task definition: name, description, metadata, container and verifier settings instruction.md the brief the agent receives: who it represents, the redlining toolkit, the turn, the grounding files environment/ Dockerfile python:3.12-slim with the redlining toolkit; copies app/ and skills/ into the image app/contract.docx the Word document to redline (Turn 2: the counterparty's tracked-change markup) app/grounding/ the represented side's commercial context and playbook, as PDF or Word files skills/contract-redliner/ the scripts the agent uses to read the contract and write tracked changes, comments and escalations tests/ test.sh, judge.py, ... the grader (validity gate, change list, judge panel) rubrics.json the attorney's rubric items for this task baseline/contract.docx the input, kept for the grader's diff references/.docx the attorney's own redline for the task (see references/README.md) data/tasks.parquet one row per task (the dataset viewer) registry.json the Harbor dataset manifest leaderboard.md, leaderboard.csv ``` ### Task names `---t-`, for example `ip-whithill-hartsong-t1-a` and `ip-whithill-hartsong-t2-c`. - `agreement`: `msa`, `ip`, `rpa` (research partnership), `mktg`, `lease`. - The two party words name the scenario; the Turn 1 side comes first, so both turns of a scenario share the prefix. - `t1-a` to `t1-e` are the five attorneys' Turn 1 tasks. A Turn 2 task carries the letter of the Turn 1 redline it answers: `t2-c` starts from the redline of attorney C. `task.toml` records this as `responds_to`. ### Metadata in task.toml `[metadata]` holds the scenario, agreement type, jurisdiction, playbook level, turn, side, the parties, the author string the agent must use, the attorney letter, `responds_to`, a numeric `expert_id` for the attorney who wrote the redline and rubric set (the same number identifies the same attorney across tasks; the mapping to people is private), the rubric counts by required action and priority, the total rubric weight, and the split. ## Rubrics and grading Each attorney completed the redline before writing roughly 10 to 20 rubric items for it. A rubric item names the contract section, the **required action** (Edit, Preserve, Comment or Escalate; on Turn 2, Edit splits into Reject, Counter and New Edit, and Preserve into Accept and Leave Untouched), a **priority** (High, Medium or Low, weighted 3, 2 or 1), an **evaluation dimension** (Issue Prioritization, Negotiation Restraint, Situational Judgment, Escalation Judgment or Standard Position Execution), the condition the redline must satisfy, and the attorney's justification. Grading runs in `tests/`: 1. A **validity gate** diffs the returned document against the input and rejects a deliverable that changed text outside tracked changes or altered the counterparty's markup. 2. A **change list** is built from the document: the model's edits by section, how each counterparty change was handled, the model's comments and escalations, and the resulting text of each section. 3. A **panel of three judge models** from different providers (Claude Opus 4.8, GPT-5.5 and Gemini 3.5 Flash) reads that record and grades every rubric item Pass or Fail; a strict majority decides. 4. The **task score** is the earned rubric weight divided by the total rubric weight. A rollout that reaches the token limit or fails the gate receives no score and counts against the model's completion rate. ## Running the benchmark The tasks run with Harbor (0.18 or later) and Docker. Git LFS must be installed: the Word and PDF files are stored with LFS, and without it the task images fail to build on purpose rather than run on placeholder files. Run the whole set straight from this repository: ```bash harbor run --repo https://huggingface.co/datasets/micro1-inc/ContractBench@v1.0 -d contractbench-public@1.0 \ -a -m --env-file .env ``` Or download it first and run from the local copy: ```bash hf download micro1-inc/ContractBench --repo-type dataset --local-dir ContractBench harbor run -p ContractBench/tasks -a -m --env-file .env ``` `.env` holds the provider keys for the agent and for the judge panel: `ANTHROPIC_API_KEY` (required by the grader), `OPENAI_API_KEY` and `GEMINI_API_KEY` (a missing panel member is dropped from the vote). Use `-i ''` to run a subset of tasks, for example `-i 'msa-*-t1-*'`. Each trial's `reward.json` holds the task score; the redlined document is captured as the trial's artifact. ## Anonymization All contracts, grounding documents and redlines are fictional material written by attorneys for this benchmark; the parties, people and events in them are invented. All personal identifiers of the attorneys who produced the redlines and rubrics have been stripped from the files and the metadata. Each attorney is recorded only as a numeric `expert_id`, which identifies the same attorney across tasks.