Editorial policy
This site describes any unresolved enforcement action as an allegation, distinguishes allegations from findings throughout, updates pages when a matter is dismissed or resolved in a respondent's favour, and logs every correction publicly.
The rule that governs everything else
An allegation is not a finding, and this site never blurs the two.
A regulator filing an action is asserting something it intends to prove. It has not proved it. Most enforcement actions settle without any admission, which establishes an outcome rather than a fact. Some are dismissed. Some are lost.
Every page on this site is written to reflect that. Where a case’s status is filed, the layout
renders an allegations notice above the narrative, and the language used is “alleged”, “accused” or
“charged” — never “committed”, “guilty” or “defrauded”. This is enforced in the templates rather
than left to the writer, and a validation step fails the build if a case narrative uses
finding-language inconsistent with its recorded status.
How cases are selected
The library records enforcement actions that involve market manipulation as this site defines it: conduct that interferes with the honest formation of a price.
Actions are ingested from the regulator indexes listed on the sources page, and each is classified against the technique taxonomy by keyword rules run over the regulator’s own release text. An action that scores no technique tag is logged and not recorded, because this is a manipulation library rather than a mirror of any agency’s full enforcement programme.
The consequences of that choice are worth stating:
- This is not a complete record of any regulator’s enforcement. Insider trading, accounting fraud, adviser misconduct and Ponzi schemes appear only where they intersect a manipulation technique.
- The classifier will miss things. An action described in unusual language will not match the keyword rules and will be silently absent. We do not know the size of this gap.
- The classifier will also over-include. A release that mentions a technique in passing may be tagged with it.
Both failure modes are why every case page links prominently to the primary document, and why records that have not been checked against the primary document carry a visible marker saying so.
How technique tags are assigned
Technique tags are ours, not the regulator’s. Agencies charge statutory provisions — Section 10(b), Rule 180.1, Section 9(a)(2) — not technique names. Mapping a charged provision and a described course of conduct onto a named technique is an editorial judgement, and it is one this site makes explicitly rather than pretending the taxonomy is given.
Tags are first proposed by keyword rules with weights, published at
data/technique-keywords.json in the repository, and are then confirmed, changed or removed by an
AI agent reading the regulator’s document (see how AI is used). A proposal
below the score threshold is not applied. Everything is logged.
A case tagged with several techniques appears in the statistics for each. This means that summing across technique pages double-counts, which is why the site never presents such a sum.
Review status
Records are written with reviewed: false and are published in that state.
This is a deliberate choice, and there is a real trade-off in it. Publishing an unreviewed record risks propagating an extraction error. Not publishing it means the action is invisible until someone gets to it, which for a growing library means indefinitely. We judged that a record clearly marked as machine-extracted, with a prominent link to the primary source, is more useful than absence and less risky than an unmarked one.
Unreviewed records render a notice saying so, and reviewed records carry a “checked” line instead (see How we check records). The distinction is also carried in the JSON export, so anyone using the data can filter on it.
What happens when a case resolves
Dismissals and acquittals are published above the fold. Where a case is dismissed, the case page renders that outcome at the top, before the narrative describing what was alleged. The same is true on any person page linked to it. This is the one place where the site deliberately breaks its own reverse-chronological logic, because a reader who lands on the page and leaves after two sentences should leave knowing the allegation did not stand.
We keep dismissed matters online rather than deleting them. Deleting a page makes the original allegation harder to find but does not remove it from the internet; keeping the page with the outcome on it is the only version that actually corrects the record.
Settlements are described as settlements. A consent judgment entered without admission is
recorded as settled, and the site does not describe the underlying conduct as established.
Sourcing and quotation
Every factual claim about a case traces to a primary document, linked on the page.
No release text is copied onto this site. The summaries and narratives on case pages are composed from the structured fields extracted from a release, not from its prose. Where quotation is genuinely necessary because exact wording is legally load-bearing, it is limited to a short phrase and attributed. This holds even though US federal releases are government works and not subject to copyright — the reason is accuracy and independence, not licensing.
Where our summary and the primary document disagree, the primary document is right, and we say so on every case page.
Language
Plain English. Terms are defined on first use and linked to the glossary. Figures are given with their units and their basis. Where a number is a lower bound, it says so.
The site does not use “shocking”, “brazen”, “massive” or similar. Enforcement records are interesting enough without adjectives, and a reference work that editorialises about defendants is less useful and less fair.
People pages
Pages about individuals are limited to people named in publicly filed regulator or court actions, and to legal, case-linked facts, each carrying a source URL.
No addresses. No contact details. No photographs. No employment beyond what a filing states. No speculation about anyone’s motives, finances or private life. Every such page carries a standing note that allegations are allegations, and links to any dismissal or acquittal we hold.
Corrections
Errors are fixed and logged publicly at /corrections/, with the date, what was wrong and what changed. We do not edit silently. A reference work that revises itself invisibly cannot be relied on.
How we check records
A record marked Checked against the primary document has had its summary, technique tags and figures compared with the regulator’s own document, by an AI model (Claude) following written instructions, working from a cached copy of the document, with sampled and disputed records read a second time. No lawyer has reviewed them. A record without that mark has not, and its tags and figures are mechanically extracted and may contain errors. As of 14 October 2026 every record in the library carries the mark. Checked records show the date they were checked where we have one, and every case page has a “Report an error in this record” link that opens an email addressed to us with the record’s address filled in. You can also write to [email protected] with the record slug, the field that is wrong and a link to the primary document. The schema records whether and when a record was checked; every check so far was made as described above.
What the mark does not mean:
- A checked record can still be wrong. A check is a comparison, and comparisons miss things.
- The primary document can itself be wrong or ambiguous. Releases and complaints contain errors, and a later order may change a figure. We record what the document says and do not adjudicate it.
- Checked is not endorsed. It says nothing about whether an allegation is true, a settlement was fair or a technique tag is the only reasonable one. Tags remain our editorial judgement.
The share of checked records is computed when the site is built and published on the case explorer as “N of M records checked”. The explorer can filter on the mark and export the current results, including a checked yes/no column, as CSV. If you find an error in either kind of record, tell us.
How AI is used
Most of the text on this site is written by an AI model (Claude), working under written instructions that the site’s operator sets and revises, and the operator publishes it. We say so on every page, in the footer, because Google’s guidance on generative AI content and common sense both say a reader should know how a page was made. This section is the longer account.
What AI does.
- Case records. Scripts collect the regulators’ own releases and orders. An AI agent then reads each document, writes the summary and narrative in its own words, sets the status, the respondents, the amounts and the technique tags, and records what the document does and does not show. No run of twelve or more consecutive words may match the regulator’s text; a build check enforces that. Some amounts, such as a penalty spread across many respondents, are totals the agent added up from a table in the order, and the record says so.
- Explanations. The technique pages, glossary, blog posts and tools are drafted by an AI agent from the primary sources they cite and from this site’s own data. Every figure that can be computed is computed by a script from the case records when the site is built, and posts are re-run when the data changes, rather than typed in.
What AI does not do. It does not generate any image or video on this site: the diagrams and charts are drawn by code from the data, and the social-sharing cards are rendered from page text. There are no invented authors, bylines or credentials: the site is published under its own name and an email address that reaches the operator.
How accuracy is checked. The original 2,428 records had sampled second readings by a second AI agent. Later sets were checked the same way: a sample of 60 France (AMF) records agreed on every field for 54 (90%); a sample of 60 India (SEBI) records agreed on every field for 56 (93%), and 50 of the SEBI records added later agreed for 46 (92%). The disagreements are logged on the corrections page and were fixed. A second AI reading is a comparison, not proof, and no lawyer has reviewed any of it. The build also fails when a case narrative uses finding-language that its recorded status does not support, when a page links to a page that does not exist, or when a published address would start returning an error.
Why these pages exist beyond the sources. A regulator’s release describes one action. This site adds what no single release does: the same respondent across regulators, the statutes cited, how far a criminal parallel went, how long matters take, how penalties are distributed, and what the records do not show. Pages that would only restate a single, well-known definition are kept out of Google’s index (the individual glossary term pages are marked noindex, and the glossary page lists every term with its definition), and bulk additions of new page sets are paused so that the library grows by checked records and not by page count.
If you find an error, the corrections page explains how to report it and logs every fix with its date.