What happened
Harvey’s harvey-labs repository contains a tasks/firm-knowledge directory. The supplied primary-source view shows two subdirectories, dms/ and matters/, along with a tasks/ directory. The story description identifies the resource as “Harvey C&H,” a synthetic law firm containing roughly 100 million tokens of client matters. The source material provided here does not expand the initials C&H or describe the corpus beyond that short description and the visible folder names.
The release is hosted on GitHub under Harvey’s organization account. The supplied page view identifies the repository path and branch as main; it does not show a version number, release date, tagged snapshot, or commit-specific identifier. Because this is a directory view, the available evidence also does not establish whether the corpus can be downloaded as a single package, how large it is on disk, or whether users need any particular software to work with it.
The “synthetic” label matters: Harvey presents the material as constructed rather than as a dump of actual client matters. That could make it a candidate for experimenting with legal document workflows without relying on a real client file set. But the supplied source does not explain how the examples were generated, whether they imitate particular jurisdictions or matter types, how much variation they contain, or what quality checks were applied. The token count is approximate, and the source excerpt supplies no breakdown by document, matter, or task.
No pricing or access restrictions are stated in the supplied material. The GitHub page shows the repository directory as public, but the evidence here does not include a licence file or an explicit grant for reuse. It also does not establish that Harvey ships an evaluation harness, baseline results, prompts, or recommended tests alongside the data. Builders should inspect the repository itself for the full files and applicable terms before treating the release as an unrestricted dataset or reproducible benchmark.
What the archive adds
An earlier Field Note covered Harvey’s separate open-source sales agent, which runs through Claude or Codex CLIs without API keys: Harvey’s open-source autonomous sales agent. That is a different product and not prior coverage of this dataset. This release is a new event for the archive, rather than a new fact on that known story.
Trend signal
This release is consistent with the idea that scientific-agent progress should be evaluated through inspectable artifacts, but the supplied evidence does not show an evaluation package or reproducible results. The closest pattern is Scientific agents need artifact evidence; its fit is limited because this is a synthetic legal corpus, not a reported scientific-agent experiment.
Why it matters
Legal AI builders often need realistic-looking, structured inputs to test document retrieval and matter-level context handling before they can safely use client material. A large synthetic corpus could help expose failures that small toy examples miss: broken links between documents, weak retrieval across a matter, or workflows that lose context as a file set grows. Harvey’s approximate 100-million-token scale is therefore a potentially practical starting point for stress testing.
The scale alone does not make the dataset representative. Teams need to know whether the documents resemble the workflows, jurisdictions, and document types they handle, and whether the synthetic construction introduces artifacts that make tasks easier or harder than real work. Nor does a large corpus prove that a legal agent can reason accurately or produce reliable work product. A useful internal test would define tasks and expected evidence, then measure retrieval and verification separately rather than treating corpus size as model quality.
Limits and uncertainty
The available source material provides only a brief description and directory listing. It does not establish the dataset’s generation process, contents in detail, jurisdictional scope, distribution of matter types, document count, or validation. It also does not state an explicit licence, pricing, access conditions beyond the visible public GitHub location, or whether the corpus includes tools for evaluation. The “roughly 100 million tokens” figure is not accompanied by a counting method. Nothing in the supplied evidence demonstrates performance gains, privacy guarantees, or suitability for a particular legal workflow.
Skill to build now
Before using the corpus as a benchmark, inspect one representative slice and write down a small task specification: the query, the evidence a successful answer must cite, and the expected failure cases. Check repository files and licence terms first. This turns a large collection into a bounded retrieval test and helps reveal whether its synthetic structure matches your application.
Reader takeaway
Harvey has made a large synthetic legal knowledge set visible in its GitHub repository, and its scale could make it useful for testing document workflows. Treat it as a candidate corpus, not a validated benchmark: inspect the data, terms, and task structure before drawing conclusions about real-world legal agents.