The cases in examples/cases/ are fictional teaching packages. They do not
recreate a victim, suspect, defendant, company, or unresolved investigation.
They borrow narrow investigative mechanics from public, authoritative sources
and combine them with invented people, places, timelines, evidence, and
outcomes.
This distinction matters for both ethics and engine design: the source tells us which kinds of records investigators actually compare, while the authored YAML remains a deterministic test fixture rather than a claim about a real person.
The United States Department of Justice described an alleged scheme in which rare Chinese manuscripts were checked out and dummy manuscripts were returned. The complaint described library asset tags, aliases/library cards, and travel timing as parts of the investigation. The source itself stresses that a criminal complaint is an allegation and that a defendant is presumed innocent.
The example adapts only the substitution-and-provenance mechanics. Its archive, documents, people, dates, and resolution are invented.
The FBI describes business email compromise as fraud that abuses ordinary business email workflows, often by impersonating a trusted party and changing payment instructions. Its field-office alert also describes lookalike domains and fraudulent attachments carrying wire instructions.
The example uses an invented supplier, domain, bank route, and recovery window. It is designed to test message headers, invoices, exact searches, and deadlines.
The National Transportation Safety Board's Railroad Investigation Report RIR-24-12 covers an April 2023 Union Pacific collision after a train crossed a main-track switch lined toward a yard track and hit parked equipment. NTSB rail reports demonstrate the use of switch position, recorder data, operational records, and timelines in a factual accident investigation.
The example's railway, crew, switch, maintenance work, causes, and findings are fictional. It does not restate the NTSB's probable-cause finding.
Europol's OPSON operation targets counterfeit and substandard food and drink. Its olive-oil operation describes lower-grade oil being used to dilute product sold under a more valuable label, with laboratory and supply-chain evidence playing central roles.
- Europol: Operation OPSON
- Europol: 11 olive-oil counterfeiters arrested following Operation OPSON
- Europol: counterfeit and substandard food seized in OPSON XIII
The example invents its cooperative, laboratory measurements, batch numbers, and responsible party. It adapts only the label/lab/ledger comparison.
Arthur Conan Doyle's “Silver Blaze” is a public-domain example of negative evidence: a guard dog's silence matters because it implies a familiar visitor. The story also combines a missing horse, a drugged stable boy, and records that reframe the timeline.
silent-stable is not a retelling. It uses an invented modern training stable,
different people, evidence, motive, and outcome to test whether the engine can
represent an expected event that did not occur.
These examples intentionally reveal where a generic engine needs more than an author-provided conclusion:
- comparisons should be explicit proof predicates, not facts silently precomputed by a case author;
- negative evidence must distinguish “observed false” from “not observed”;
- timestamps from different devices need generic normalization and ordering;
- numeric ranges, set intersection, and lot/account joins belong in versioned capabilities rather than case-specific branches;
- media needs localized labels, alternative text, captions, and transcripts;
- preservation and chain of custody should eventually be first-class effects, not just flags.
Until a predicate or capability exists, a case may expose a signed analyst
finding as a structural observation (for example,
classification: lower_grade). The package README must say when it does so.
That keeps the current demo honest and gives future engine work a concrete
regression package to improve.
