The problem
Lawyers have been sanctioned in real courts for filing briefs containing case law an AI invented. The citations looked perfect. They did not exist.
That is not a bug in one model. It is what happens when a system whose mechanism is generating text is asked a question about a document. A plausible fabrication is always available to it, and it has no way to tell you which kind of answer you just received.
What Anchor does
Anchor never writes a sentence about your contract. It only locates text that is already there, labels what kind of term it is, and quotes it back with a clause citation.
There is no step at which a fabrication could enter, because there is no step that produces new text. Click any finding and the document scrolls to that clause with the matched words highlighted. You are never asked to trust the tool. You are shown the evidence and you check it yourself in about two seconds.
Measured on real contracts, not on my own
The first version was tested against a contract I wrote myself, which only proves that rules match text written to be matched. So I pulled 20 real agreements filed as SEC exhibits, from Adobe, Vonage, Jazz Pharmaceuticals, BellRing Brands and others, drafted by lawyers who had never heard of this tool.
Recall is measured against a plain keyword search of the same documents. If a term is present and Anchor does not report it, that counts as a miss.
The first run scored 54.5 percent. That number was worth more than any feature, because the failures were specific. Confidentiality found 0 of 11 contracts, because my rule demanded a duration clause while real contracts put the duty in one sentence and the duration in another. Liability caps found 1 of 8, because my rule required a money figure where real caps are written as a multiple of fees paid. Governing law missed 5 of 8 over the exact words "the laws of".
After widening those patterns against real drafting: 84.4 percent recall, 10.4 findings per contract.
Citation accuracy stayed at 100 percent across 197 quotes on documents the tool had never seen. That is the property that actually matters, and it is the one that did not move.
The harness is in the repository and reproduces the table, including the terms it is still weak on: assignment restrictions and IP ownership, both on small samples.
On the absence of a language model
This is an AI safety submission that contains no language model, which is the point rather than an omission. The failure it addresses is fabricated citation, and fabrication is a property of generation. A system that only locates and quotes existing text cannot produce a citation that does not exist, and no amount of prompting or fine-tuning gives a generative system that same guarantee.
The 84.4 percent recall is the honest cost of that choice. A model would likely find more, and would also be able to invent things, which for a document someone signs is the worse trade.
What it finds
Twenty-four term types across six categories: time, money, exit, risk, restrictions and law. Automatic renewal, notice periods, liability caps, uncapped liability, indemnities, unilateral amendment rights, non-competes, governing law and more. Each is graded risk, watch or info, with a short explanation written for someone who is not a lawyer.
Asymmetry detection
The finding people miss is not a bad clause. It is a clause that is fine for one side and absent for the other, because each half reads reasonably on its own. Anchor counts which party holds each one-sided right and flags rights only one party appears to hold.
How it was built
Plain JavaScript. No dependencies, no framework, no build toolchain beyond a short concatenator. Everything runs in the browser tab, so the contract is never uploaded anywhere. Turn off your network and it still works.
Problems worth mentioning
The first segmentation pass merged clause 7.2 into 7.1. Contracts write whole paragraphs on the same line as the clause number, and my heading detector rejected long lines. Every citation after a long clause pointed at the wrong place, which is fatal for a tool whose entire value is citation accuracy. It is now a regression test.
Honest limits
Plain text only, so PDF contracts need converting first. Rules are English-language patterns, so an unusually drafted clause can be missed, and a miss shows as nothing rather than as false comfort. It finds and explains terms. It does not tell you whether to sign, and it is not legal advice.
Log in or sign up for Devpost to join the conversation.