Federal crimes: using Mesa to sift through the U.S. Code at scale

It’s a federal crime to sell margarine with “more than one and six-tenths degrees of yellow,” to import honeybees with “genetically undesirable germ plasm,” and to put an advertisement on a flag in Washington, D.C. Provisions like these are spread across 47 titles of the U.S. Code, which is one reason no one has an official count of federal crimes.

People have tried. In the early 1980s, a team of Justice Department lawyers spent two years on it and came back with an estimate of roughly 3,000. Ronald Gainer, who oversaw the effort, said that anyone looking for an exact number “will have died and resurrected three times” before finding one. Now the Count the Crimes to Cut Act, which passed the House and was reported to the Senate on April 14, 2026, would give the Attorney General a year to produce the list.

An American Bar Association (ABA) task force offered three reasons the count is so hard: criminal provisions are scattered throughout the U.S. Code, one statutory section can contain multiple crimes, and criminal sanctions can exist outside the statutory code.

I tackled the former two issues (and some of the third) with Mesa. First, I identified sections likely to have criminal provisions across the whole United States Code, as amended through Pub. L. 119-73. Then, I used batch inference to spin up many agents on a small, efficient model to each read identified sections simultaneously. Mesa gave each agent the same instructions to identify crimes within a section if present, follow cross-references to penalty and definition sections, and record relevant information about each crime, splitting those defined in the same section.

Mesa compiled a dataset of 4,640 distinct offenses and associated information. It’s accessible through the visuals and download options below.

Building the dataset

I built the dataset through four steps: screening provisions in the U.S. Code for those with penalty language, having Mesa read each qualifying section, instructing batch agents to record each separate offense and related information, and comparing results to earlier estimates.

When recording each offense, I had Mesa’s batch agents identify specific fields: prohibited conduct, mens rea requirements, maximum penalty, penalty section, whether the conduct is defined by a regulation, and a confidence rating that the provision is criminal rather than civil. Each row in the dataset also has agent reasoning and quoted statutory text.

Where federal crimes live

The dataset records 4,640 crimes in total. While Title 18 is the criminal code, around 68% of federal crimes (3,162) were identified across 46 other titles.

Title 18’s chapters, on average, consistently define crimes; chapters like Embezzlement and Theft have a crime in every section. Chapters in other titles regularly have crimes in penalty clauses inside long civil or regulatory chapters.

The titles with the most crimes following Title 18 are Commerce and Trade (Title 15, 339 offenses), Agriculture (Title 7, 293), Conservation (Title 16, 286), Food and Drugs (Title 21, 276), and Armed Forces (Title 10, 236).

The chapters with the most offenses are:

• The UCMJ (10 U.S.C. ch. 47, 175 offenses). It houses punishments for desertion to dueling (a service member who fights or promotes a duel commits an offense).
• Fraud and false statements (18 U.S.C. ch. 47, 99 offenses).
• Counterfeiting and forgery (18 U.S.C. ch. 25, 95 offenses).
• The alcohol tax chapter (26 U.S.C. ch. 51, 85 offenses). It’s a legacy of Prohibition-era enforcement.
• The Food, Drug, and Cosmetic Act (21 U.S.C. ch. 9, 78 offenses).

Regulatory crimes

For 35% of offenses, the conduct defined as a crime is set by a regulation, order, or permit instead of the statute. Among titles with at least 50 offenses, Food and Drugs (Title 21) has the highest regulatory share (65%), followed by Foreign Relations (Title 22), Conservation (Title 16), and War and National Defense (Title 50). In those titles, the statute usually makes it a crime to violate “any regulation issued under this chapter,” so the conduct actually prohibited is defined in the Code of Federal Regulations.

Uncertain cases: criminal or civil?

Each row carries a confidence rating of whether it is a criminal rather than a civil provision: 4,340 high, 268 medium, and 32 low. The low- and medium-confidence rows fall mostly into three recurring patterns:

• Contempt clauses: Many agency statutes say that a person who disobeys a court order enforcing a subpoena “may be punished” as contempt. Many statutes don’t define contempt as civil or criminal. Most of these clauses don’t set a punishment of their own, instead relying on the courts’ general contempt power in 18 U.S.C. § 401.
• “Subject to a fine”: The ban on slaughtering dogs and cats for food says a violator is subject to a fine of up to $5,000; it isn’t defined as a civil or criminal provision.
• Fines imposed inside another proceeding: For example, bankruptcy petition-preparer fines and copyright deposit fines have no criminal procedure attached.

The “Ambiguous cases” sheet in the workbook lists all 300 of these rows with the reviewer agent’s notes.

Grade and maximum sentence

39% of offenses are felonies, 25% misdemeanors, and 6% fine-only infractions. 24% are different tiers based on value, harm, or prior convictions. 6% could not be graded from the statute.

One year and five years, the usual maximums for misdemeanors and felonies respectively, are the most common grades for crimes.

Of the 56 death-eligible offenses:

• 23 are in the UCMJ, mostly offenses in time of war.
• Most of the rest are murder offenses in Title 18.
• Several are protective-killing clauses in agricultural statutes. For example, killing an egg inspector, an official enforcing the Horse Protection Act, or a nuclear safety inspector is punished under the federal murder statutes.

Mental state

Criminal statutes often define what the defendant must have known or intended to be guilty of the respective crime.

“Knowingly” appears in 40% of offenses and “willfully” in 26%. Many offenses include multiple terms, but the chart above categorizes each under the roughly most demanding term its statute names. 28% name no mental state at all. However, in practice, courts usually require one.

Unusual crimes

The U.S. Code still contains crimes that reflect the commerce and worries of earlier eras. A full inventory is what makes them easy to find.

Comparing with Count the Code

Count the Code (CTC) is the best-known earlier attempt to count federal crimes. It searched the 2019 Code for a fixed list of penalty phrases, like “shall be fined under this title,” and counted the sections where they appear. The result is a list of sections: a citation for each match, with nothing about what the section makes a crime.

This dataset is a catalog of the crimes themselves. It records offenses rather than sections, one row per crime, and describes each one: the prohibited conduct and its elements, the mental state, the maximum penalty and the section that sets it, the grade, whether a regulation defines the conduct, and a confidence rating that the provision is criminal rather than civil. Every row quotes the statutory text it rests on, along with the agent’s reasoning, so anyone can check it against the Code. That’s what makes the analysis above possible. A list of sections can’t tell you which crimes are felonies, which name no mental state, or which carry the death penalty.

Counting offenses instead of sections also changes the number. In the sections both datasets cover, CTC counts 3,786 crimes and this dataset counts 4,500, because one section often defines several offenses. This dataset also finds crimes in 776 sections CTC doesn’t count at all.

Every section CTC counts is accounted for here. Of the 91 that this dataset doesn’t flag directly, 83 are penalty sections, recorded under the section that describes the conduct with the penalty section noted alongside it. The remaining 8 show the limit of phrase matching: they contain a penalty phrase but don’t define an offense. 42 U.S.C. § 9606, for example, sets civil fines in Superfund abatement actions.

Scope and limitations

This dataset only records crimes in the U.S. Code and therefore doesn’t fully resolve the ABA’s third identified difficulty in counting federal crimes (that crimes may be defined outside the statutory code). Where a statute makes it a crime to violate an agency’s rules, this dataset only records one offense even if the Code of Federal Regulations might define many violations under it. These results also leave out uncodified statutory notes and general accessory liability.

Mesa wrote this dataset through only reading sections selected by a keyword search, meaning some sections defining a crime may have been missed. However, the dataset also counts every section that CTC does, suggesting few (if any) crimes were missed this way.

There are other reasonable ways to count crimes. For example, 21 U.S.C. § 841 makes it a crime to “manufacture, distribute, or dispense” a controlled substance. This dataset records this as one offense, though it could also be considered as three.

I also did not analyze case law for this dataset, meaning data on mental states, grades, maximum punishments, and more come from only the statutory text.

What comes next

Anyone who sets out to produce an official list of federal crimes will run into the same questions this dataset flags: whether a contempt clause is criminal, how many offenses a section like 21 U.S.C. § 841 holds, and how to count crimes whose conduct is defined in the Code of Federal Regulations. I hope this dataset is a useful starting point for that work, and for anyone who wants to check it.

The clearest lesson for me is how much of federal criminal law sits outside the criminal code. About two of every three offenses are in titles on commerce, agriculture, conservation, food and drugs, and the armed forces, often as a penalty clause inside a long regulatory chapter. That’s why counting them has been so hard: finding them means reading every candidate section and following its cross-references, which a phrase search can’t do.

The same approach works on other bodies of law. State and local governments ask similar questions of their own codes and rules: which provisions carry penalties, which fees are out of date, and which rules no longer match the statutes behind them. Reading an entire code section by section and recording what each provision does is now work that Mesa can take on across the whole code at once.

The data

The full workbook and a CSV are available to download.