Build1 publisher2 min readPublished
Reversibility, blast radius and stakes together sort a legal agent's actions into four grades before controls attach
A dev.to guide scores every action an AI legal agent takes by reversibility, blast radius and stakes, then attaches controls by grade: maker-checker on anything executed or filed, attorney review on all substantive work product.
The Engineer · Build desk
What happened
- A dev.to guide for AI legal and contract agents grades every action G0 to G3 on three factors, reversibility, blast radius and stakes, and reversibility is described as the most important of the three.
- It starts from whether a human can realistically catch the mistake in time, and answers that for a fabricated citation in a fluent brief or a contract about to be signed, the answer is usually no.
- Above G1 it attaches five controls: attorney review of substance, source-checking of every citation, maker-checker on anything executed or filed, untrusted handling of ingested documents, and logging.
- G1 work is left to run on its own with a log, which keeps the expensive human controls confined to the G2 and G3 end of the scale.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Routing moves off document type and onto whether the action can be undone, so the drafting step and the send step of the same NDA can land in different grades and get different controls.
- constraint Making citation checking a resolution task limits what the reviewer is asked to do with their eyes; a cite either resolves to an on-point source or it does not, and that job can be queued or automated.
- cost At the filing end a team pays for two humans, not one: the substantive reviewer and a separate checker, both inside whatever time is left before the deadline.
- exposure The counterparty's markup is input the agent is told to read, so the untrusted-document rule applies squarely to the redline path.
A reviewer reading a fluent brief cannot see the hole in it. The fabricated citation sits in the same register as the real ones, inside prose the reviewer is being asked to approve on substance. The post's fix is to stop relying on attention and require resolution instead: each cited case, statute or clause must resolve to a real, on-point source before the output is used, and an unsourced legal assertion from an agent is unverified by default [12].
The same idea shapes what the human is shown. Putting the sources and citations next to the claim, so a person can click through and confirm, is the Show step in the post's Grade, Guard, Show, Prove sequence [13][5]. Attorney review carries a display requirement too. The review has to give the attorney the real artifact and the context to judge it, not a tidy summary that hides what changed [8]. The summary is the cheaper view to build. Plenty of review screens probably ship one for that reason.
Grading happens per action and per context. The post says you grade the specific action and its context, and that there is no one blanket rule for legal AI [10]. Extracting a clause from a routine NDA and extracting indemnification terms from an eight-figure deal do not land on the same grade [11]. Inside an agent those are the same extraction call, so the grade has to come from context the caller passes in.
Grading does not thin the human layer at the top of the scale. Five controls attach at G2 and G3 [7]. Only maker-checker is scoped to actions that execute or file; substantive review by an accountable, licensed attorney covers anything that constitutes legal work product, and the post says the attorney owns the legal judgment [14][15]. So a filing against a hard deadline needs the substantive review and a second person as checker before it goes.
Ingested documents are treated as untrusted [7]. In a contract loop the ingested document is frequently the counterparty's markup, which is what the agent is pointed at when it proposes a redline [9].
The available text of the post breaks off in the middle of the maker-checker section, so the requirements on the second signer are not spelled out. LoopRails publishes an interactive grader alongside separate G2 and G3 guides [17]. G1 work is allowed to run, and it gets logged [6].
What to watch
- Whether the G2 and G3 guides specify who can act as checker and what the checker must be shown before signing.