US Sentencing Guidelines API: Querying the Manual That Sets Federal Exposure

The US Sentencing Guidelines Manual is available as a queryable corpus of 302 citable sections behind corpusType=SENTENCING_GUIDELINES on the Vaquill AI primary law API, refreshed annually because the Manual itself changes annually. That is the whole answer to the retrieval question. The harder question, and the one this post is really about, is what you do with it: the Guidelines have been advisory rather than mandatory since United States v. Booker, 543 U.S. 220 (2005), yet a federal judge still has to calculate the range correctly before departing from it, which makes the Manual the number every federal exposure model starts from. If you are building for criminal defense, white-collar practice, or any workflow that has to say "here is the likely federal sentence," you need the Manual as data, plus two other corpora that are not the Manual.

TL;DR

• 302 sections, corpusType=SENTENCING_GUIDELINES, annual refresh. One of the smallest federal corpora we serve, and one of the highest-leverage, because a handful of sections drive the entire arithmetic of a federal sentence.
• Annual is the correct cadence, not a shortcut. The Commission submits amendments to Congress, and absent congressional action they take effect on November 1. This corpus changes on a known date once a year instead of continuously.
• Advisory does not mean optional. A district court must still calculate the range as the starting point, so the Manual anchors the outcome even when the sentence lands elsewhere.
• Application notes are commentary with real interpretive weight, not footnote decoration. Under Stinson v. United States, 508 U.S. 36 (1993), commentary interpreting a guideline is authoritative within limits, so a tool that indexes guideline text and drops the notes is wrong at the exact moment it matters.
• Model four axes, not one: chapter structure, offense level, criminal history category, and the sentencing table that turns the first two into a range in months.
• Scope is narrow on purpose. This is the Manual text. It is not case law construing it, not the Commission's data reports, and not any state sentencing scheme.

Part of our US primary law corpus series, where each post takes one source and works it end to end.

The post that pairs most closely with this one is DOJ Justice Manual API: 1,548 Sections of Federal Prosecution Policy, since charging policy shapes exposure long before a range does. On the procedural side, see Federal Rules of Procedure API: FRCP, FRE, FRAP, FRCrP, FRBP and Supreme Court Rules.

How these numbers were measured

Every count in this post comes from a live measurement of our own corpus on 2026-08-30, and every one of them is re-checkable by anyone with a key. GET /us/statutes/coverage is free, returns per-corpus counts, and stamps a measuredAt on the response, as the coverage reference sets out. We build this API, so treat the counts as our own published figures and go verify them rather than taking our word for it.

Legal propositions in this post carry their citation inline so you can check them against the source. Case citations are given in full the first time they appear.

What the Guidelines actually are

The United States Sentencing Commission is an independent agency in the judicial branch that promulgates the Guidelines Manual. The Manual tells a federal judge how to convert a conviction into a recommended range expressed in months of imprisonment.

Before 2005, that range was mandatory. Booker severed the provision that made it binding and left the Guidelines advisory.

What survived is more consequential than most product teams assume. A sentencing court still calculates the correct range first, treats it as the starting point, and then considers the statutory factors before choosing a sentence. A miscalculated range is a procedural error that can send a case back on appeal even when the sentence imposed looks reasonable.

The amendment cycle, and why annual refresh is right

Most legal corpora change whenever a publisher pushes an update, which is why our CFR corpus refreshes daily and the Federal Register corpus refreshes weekly. The Guidelines do not work that way.

The Commission votes amendments, then submits them to Congress. Absent congressional action to reject or modify them, the amendments take effect on November 1 of that year. The mechanism sits in the Commission's enabling statute at 28 U.S.C. § 994(p).

That single fact changes the engineering. You are not polling for drift. You are waiting for a scheduled cutover on a date you know a year in advance.

The structure a developer has to model

The Manual is not a flat list of rules. Getting the shape right up front saves a rewrite later.

Chapters. The Manual runs Chapters One through Eight. Chapter One holds the introduction and application instructions, Chapter Two holds offense conduct organized by offense type, Chapter Three holds adjustments (role in the offense, obstruction, acceptance of responsibility), Chapter Four holds criminal history, Chapter Five holds the determination of the sentence including the table itself, Chapter Six holds sentencing procedures and plea agreements, Chapter Seven covers violations of probation and supervised release, and Chapter Eight covers sentencing of organizations.

The offense-level axis. Chapter Two assigns a base offense level for the offense of conviction, then specific offense characteristics add or subtract levels. Chapter Three adjustments apply on top. The result is a single integer, and the scale runs from level 1 to level 43.

The criminal-history axis. Chapter Four converts prior convictions into criminal history points, and points map to one of six categories, I through VI.

The sentencing table. Chapter Five holds a grid: offense level down one axis, criminal history category across the other, and a range in months in each cell. That grid is the artifact everyone actually wants, and it is why the two axes have to be modeled as first-class fields rather than free text pulled out of a paragraph.

Notice where the chain leaves the Manual. The last two steps run on 18 U.S.C. § 3553(a), which is statute, not guideline. A product that stops at the range has done the arithmetic and skipped the law.

Application notes are not decoration

Each guideline is followed by commentary, and the part labeled Application Notes is where the Commission explains what the guideline's terms mean, how to count loss, what counts as a prior, and which adjustments stack.

Stinson v. United States, 508 U.S. 36 (1993) held that commentary interpreting or explaining a guideline is authoritative unless it violates the Constitution or a federal statute, or is inconsistent with or a plainly erroneous reading of the guideline it interprets. In practical terms, the note usually decides the question.

There has been real litigation since about how far that deference reaches, and circuits have not all landed in the same place. That argument is case law, which is not in this corpus. What you get here is the note itself, cited and retrievable, so a human or a model can read the text the argument is about.

Citation form

The Manual is cited as U.S.S.G., sometimes written USSG without periods. A guideline section looks like U.S.S.G. § 2B1.1, with the number decoding as chapter 2, part B, subpart 1, guideline 1.

Subsections carry through in parentheses, so a specific offense characteristic reads U.S.S.G. § 2B1.1(b)(1). An application note is cited as commentary, in the form U.S.S.G. § 2B1.1 cmt. n.3.

Because the Manual is republished as an edition, careful practice names the edition when the version matters. That is a real complication and it is worth a section of its own.

The version problem, stated honestly

A citation resolves here to exactly one stored text, today's. Prior editions are a separate sourcing question from the current text this corpus serves. Amendment History and Point-in-Time Law: What a Legal API Can and Cannot Tell You sets out what any legal API can honestly claim on this.

For most corpora that is a minor limitation. For the Guidelines it is a limitation worth planning around, because federal sentencing has a genuine version question baked into it. Peugh v. United States, 569 U.S. 530 (2013) held that applying a later Manual with a higher range than the one in effect when the offense was committed can violate the Ex Post Facto Clause.

So a defense workflow sometimes needs the edition in effect years ago, and this API will hand you the current one. Say that out loud in your product rather than letting a user assume otherwise. What the API does give you is amendment history per section, described in the amendment history reference, a lastAmendedYear field, yearFrom and yearTo currency filters, and change events on watched sources, which together tell you that a section moved even though they will not serve you the superseded text.

One caveat on the year filters across the whole API: they filter on the last amendment year the publisher credits, and about a fifth of sections across the corpus carry no amendment credit at all because some publishers print none. Those sections drop out once either bound is set.

Worked queries

Base URL is https://api.vaquill.ai/api/v1, auth is a bearer key, and statutes routes are mounted at /us/statutes/*. Search costs 4 credits per call, coverage is free, and failed calls are not charged.

Confirm the corpus is there and current.

curl -s https://api.vaquill.ai/api/v1/us/statutes/coverage \
  -H "Authorization: Bearer vq_key_..."

For the Guidelines the useful field is currency, which names the edition the stored text belongs to, because on this corpus the edition is the version question. The full field list covers the rest of the response. This endpoint is the one place where the API describes itself, and it costs nothing to call.

Find the guideline that governs loss in a fraud case.

curl -s -X POST https://api.vaquill.ai/api/v1/us/statutes/search \
  -H "Authorization: Bearer vq_key_..." \
  -H "Content-Type: application/json" \
  -d '{
    "query": "loss calculation for fraud offenses",
    "corpusType": "SENTENCING_GUIDELINES",
    "limit": 10
  }'

Search is hybrid, running semantic and keyword retrieval together. That matters here because Manual vocabulary ("intended loss", "sophisticated means") is jargon a pure vector search blurs and a pure keyword search misses when a user asks in plain English.

Then fetch the section. Take the actId from the search response and pass it straight through.

curl -s "https://api.vaquill.ai/api/v1/us/statutes/section/{actId}" \
  -H "Authorization: Bearer vq_key_..."

The actId is stable and not guessable, a shape the section id reference explains. Hand-built ids usually 404, which is deliberate: an id you constructed from a citation string is an id you would keep constructing after the underlying section moved. Take it from a search or resolve response and store it.

Narrow when you know where you are. The search body accepts chapter, part, titleNumber, code, source, matchType, yearFrom, yearTo, limit (max 50), offset (max 70), and fields. Unknown filter values return 422 rather than silently matching nothing, and the source filter's error message lists every valid code, so a wrong guess teaches you the vocabulary instead of returning an empty array you misread as "no such law."

Paging is cut from a single ranking, so results do not repeat or vanish between pages, and page five costs the same as page one, per the pagination rules.

What is not in this corpus

Being specific about the gaps is more useful than a coverage number.

That last row is the important one.

The three corpora a defense workflow actually needs

Federal exposure is never a single-source question. A real answer draws on three separate bodies of text that live in three separate corpora here.

The Guidelines Manual gives you the advisory range. Title 18 of the United States Code gives you the statutory maximum and any mandatory minimum, plus the § 3553(a) factors the judge weighs on top of the range. The DOJ Justice Manual gives you the department's own charging, plea, and declination policy, which shapes what the government asks for long before a judge sees the case.

The Justice Manual sits under corpusType=AGENCY_GUIDANCE with source=doj_justice_manual, at 1,548 sections across 9 Titles, and it cites in its own native form, like "Justice Manual 9-27.420" for the plea agreement considerations. DOJ Justice Manual API: 1,548 Sections of Federal Prosecution Policy covers querying it as data, including why its citation form breaks naive parsers; it is worth reading alongside this one if you are building a charging-stage tool.

The query sequence is three calls, and each one is cheap.

# 1. the guideline
-d '{"query":"aggravated identity theft","corpusType":"SENTENCING_GUIDELINES","limit":5}'

# 2. the statute
-d '{"query":"aggravated identity theft penalty","corpusType":"USC","titleNumber":"18","limit":5}'

# 3. the charging policy
-d '{"query":"charging aggravated identity theft","corpusType":"AGENCY_GUIDANCE","source":"doj_justice_manual","limit":5}'

Three corpora, one schema, one actId convention, and every result carries the official source URL for the publisher's own copy so a human can check the text against the government's version.

Who this is wrong for

If you need to argue about how the Second Circuit reads a specific application note, this corpus gives you the note and nothing else. Buy case law elsewhere.

If you are building state criminal defense tooling, the Guidelines are the wrong book entirely. State sentencing runs on state statutes and, in some states, a state commission's own scheme, and neither is served here.

If your product depends on retrieving the 2018 Manual to run an ex post facto comparison, treat prior editions as their own sourcing problem and size it before you build.

Where the corpus sits

Of the ten federal corpora on this API, only the US Constitution (74 sections) and the bilateral tax treaties with their Technical Explanations (119 items) are smaller. The largest is the CFR at 219,114 sections on a daily refresh.

Federal totals 618,875 sections, inside a corpus of 4,150,839 distinct sections across 53 jurisdictions, indexed as 12,003,716 retrieval passages. Each source here is the government publisher's own copy. Every source is taken from the government publisher directly rather than from a commercial aggregator. Where a publisher's terms prohibit automated access, that is declared in the API response rather than filled from a reseller.

We build the Vaquill AI primary law API. The collection layer behind it is published at open-us-law, so the sourcing is readable rather than asserted.

FAQ

Are the US Sentencing Guidelines mandatory?

No. United States v. Booker, 543 U.S. 220 (2005), made them advisory. A federal judge still has to calculate the correct range as the starting point and consider it alongside the factors in 18 U.S.C. § 3553(a), and a miscalculated range is a procedural error on appeal.

How often do the Sentencing Guidelines change?

Once a year on a scheduled date. The Commission submits amendments to Congress, and absent congressional action they take effect on November 1. That is why the corpus refreshes annually rather than daily, and why a fall review gate beats continuous polling.

How do I cite a Guidelines section?

Use U.S.S.G. followed by the section number, as in U.S.S.G. § 2B1.1. Subsections carry through in parentheses. An application note is cited as commentary, in the form U.S.S.G. § 2B1.1 cmt. n.3, and careful practice names the Manual edition where the version is contested.

Do application notes have legal force?

They carry real interpretive weight. Stinson v. United States, 508 U.S. 36 (1993), held that commentary interpreting or explaining a guideline is authoritative unless it violates the Constitution or a federal statute, or is inconsistent with or a plainly erroneous reading of the guideline. How far that deference reaches is still litigated, and that argument is case law rather than Manual text.

Can I get a prior edition of the Guidelines Manual from the API?

Versioning here runs off the amendment record rather than a date parameter. One stored text per citation, the current one, and no as_of parameter to ask for another. You do get amendment history per section, a lastAmendedYear field, yearFrom and yearTo filters, and change events on watched sources, so you can tell that a section moved even though you cannot retrieve the superseded text.

What is the sentencing table?

A grid in Chapter Five of the Manual with final offense level on one axis, running from level 1 to level 43, and criminal history category on the other, running from I to VI. Each cell holds an advisory range in months. Model both axes as typed fields, because they are the inputs everything else feeds.

Is case law about the Guidelines included?

No. The corpus is the Manual text: guidelines, commentary including application notes, and the table. Circuit and Supreme Court decisions construing a guideline are a separate body of law and are not served by this API.

What else do I need besides the Guidelines for a federal exposure model?

Two things. Title 18 of the United States Code for the statutory maximum, any mandatory minimum, and the § 3553(a) factors, and the DOJ Justice Manual (1,548 sections, source=doj_justice_manual) for the department's own charging and plea policy. All three are separate corpora on the same API and the same schema.

How much does a Guidelines query cost?

A search call is 4 credits. GET /us/statutes/coverage and GET /boards are free, failed calls are not charged, and paging costs the same on page five as on page one because pages are cut from a single ranking.

Do you cover state sentencing guidelines?

Not as a distinct corpus. State sentencing rules are mostly statutory and are served under corpusType=STATE alongside the rest of a state's code, across 52 jurisdictions. If you need a state commission's own published scheme as a separate, structured product, that does not exist here today.

Priyansh Khodiyar

Co-Founder & CTO

Priyansh leads engineering and AI at Vaquill, from the matter workbench to drafting, document comparison, document matrix, and citation-verified research.