Rules that govern how orders are handled are written in prose and enforced in software. Between the two sits a reading: someone decides what a phrase means and writes code that acts on it. That reading is easy to lose. It lives in a ticket, a code comment or one engineer's memory, and it is hard to audit when the rule is amended or a reviewer asks why a system behaved as it did.
This note sets out a method that keeps the reading in the open. An LLM extracts a rule's operative fields, each linked to the exact phrase it came from. A person reviews every field. We write down the property the rule should guarantee, check a model against it and trace any counterexample back to the words of the rule. The worked example is Rule 201 of SEC Regulation SHO, the short sale price test, read from the current text in the Electronic Code of Federal Regulations [1].
Why start from the text
The published text is the one artifact every party shares. Summaries drift and implementations differ, but the text changes only by amendment, and amendments are dated. Starting there gives every later step a fixed reference. When someone questions a field, the answer is a pointer to a phrase. When the rule changes, the difference between two versions of the text says which fields to revisit.
Others have worked in this direction. The OECD's Rules as Code work proposes machine-consumable versions of some government rules that sit alongside the natural-language text [14]. Catala, a programming language for statutory law, gives lawyers and programmers a shared medium for turning statutes into executable code, and its authors report finding a bug in an official implementation of French family benefits [10]. Our contribution is narrower: a link from every field to its span of text, an LLM that proposes and a person who decides.
The rule as it reads today
Rule 201(b)(1) requires a trading center to "establish, maintain, and enforce written policies and procedures reasonably designed to":
(i) Prevent the execution or display of a short sale order of a covered security at a price that is less than or equal to the current national best bid if the price of that covered security decreases by 10% or more from the covered security's closing price as determined by the listing market for the covered security as of the end of regular trading hours on the prior day; and
(ii) Impose the requirements of paragraph (b)(1)(i) of this section for the remainder of the day and the following day when a national best bid for the covered security is calculated and disseminated on a current and continuing basis pursuant to an effective national market system plan.
That is the eCFR text, current as of 7 October 2026 [1]. Paragraph (b)(1)(iii) adds two exceptions: a displayed order that was priced above the national best bid when first displayed, and an order marked "short exempt". Paragraph (b)(3) says the listing market makes the determination and "shall immediately make such information available as provided in § 242.603(b)" [1].
Fields, each tied to a phrase
The LLM reads the passage and proposes one field for each operative phrase. Each record stores the value, the exact span it came from, the date of the text version, the model version and prompt that proposed it, and the reviewer's decision.
| Field | Value in the model | Source phrase in 17 CFR 242.201 |
|---|---|---|
| trigger | price at or below 0.90 × baseline | "decreases by 10% or more" |
| baseline | listing market's close, prior day | "closing price as determined by the listing market ... as of the end of regular trading hours on the prior day" |
| restriction | no short sale executed or displayed at or below the national best bid | "at a price that is less than or equal to the current national best bid" |
| duration | rest of the trigger day and the following day | "for the remainder of the day and the following day" |
| condition | while a national best bid is disseminated under an NMS plan | "calculated and disseminated on a current and continuing basis pursuant to an effective national market system plan" |
| determination | made by the listing market, then made available | "shall be made by the listing market" |
| exceptions | (b)(1)(iii)(A) and (B), one record each | "Provided, however" |
Several fields lean on text elsewhere. "Covered security", "listing market", "national best bid" and "regular trading hours" are defined in paragraph (a), which points into the Regulation NMS definitions in § 242.600 [1, 2]. The model stores each as a dependency with its own source span, so a reviewer reads both passages side by side.
The text moves
Rule 201 was adopted in February 2010 [5]. The section's source note lists three later amendments, published in 2018, 2021 and 2024, and the eCFR version history shows each one [1, 4]. All three renumbered the cross-references into § 242.600 [7, 8, 9]. The paragraph defining "national best bid" has moved from (b)(42) to (b)(43), (b)(50) and now (b)(60) [4, 1].
The 2021 Market Data Infrastructure amendments also changed the words [8]. Before them, (b)(1)(ii) read "disseminated on a current and continuing basis by a plan processor pursuant to an effective national market system plan", and (b)(3) required the listing market to "immediately notify the single plan processor", which then disseminated the information [3]. Today "by a plan processor" is gone, and (b)(3) points to § 242.603(b) [1].
Staff guidance has not moved with the text. The SEC staff's Rule 201 FAQ, whose latest dated entry is marked 4 June 2019, quotes the earlier wording of (b)(1)(ii) and cites regular trading hours as Rule 600(b)(64) [6]. The current cross-reference is § 242.600(b)(88) [1]. The FAQ still answers questions the text leaves open, so the model keeps it, as a separate class of evidence with its own date. The staff say the answers "are not rules, regulations, or statements" of the Commission [6]. A field that rests on guidance says so.
A property, then a check
With the fields reviewed, we write down what the rule should guarantee, in plain words first. Once the listing market's determination is made available, no short sale order in that security executes or is displayed at or below the current national best bid for the rest of that day and the following day, unless an exception applies. A new 10% decline on a restricted day starts the window again. The property is fixed before any check runs.
The check runs the model against generated sequences of trades, quotes and orders, weighted toward the places where the text draws lines. If a sequence breaks the property, the check returns it as a counterexample, and each step names the fields it touched.
Here is one, constructed against a deliberately flawed version of the model. That version reads "the following day" as the next calendar day. A trigger on a Friday then leaves Monday unrestricted, and the check returns a two-step sequence: the trigger on Friday, then a short sale executed at the national best bid on Monday. The trace points to the duration field and its span in (b)(1)(ii). The text alone does not settle the question. FAQ answer 2.1 does: "following day" refers to the next trading day, so a Friday trigger carries into Monday [6]. The reviewer corrects the field and attaches both sources.
The full report gives two more: a second decline on the following day, which the Commission's adopting release says re-triggers the restriction [5], and a trading halt that leaves no close for the prior calendar day [6].
People decide
The LLM proposes and a person decides. A published evaluation shows why that split matters: a 2024 study of commercial AI legal research tools found they "hallucinate between 17% and 33% of the time" [13]. Each field is accepted, edited or rejected by a reviewer, and the decision is logged with the field. Model versions and prompts are pinned, so any proposal can be traced to its inputs. When the rule is amended, extraction runs again on the new text and the reviewer sees a field-by-field difference against the reviewed model.
Grellum, our agentic legal research product, applies a related rule: every citation must match retrieved evidence.
Read the full report
The full report, available as a PDF, sets out the field records, the version history of every cross-reference, the open questions with their sources, the property in model terms and three traced counterexamples.