Under the hood · CaseLinker

Case reports

press releases, dockets, PDFs

Public case reports from 76 agencies and task forces

ICAC task force and police releases, state attorneys general, DOJ and NCMEC reports, and federal court records from PACER. Everything was public before CaseLinker read it.

76 sources. Each listing is fetched, every URL probed, and only public documents are downloaded
Each case keeps its source, URL and fetch time, so any value can be traced back to a document

Layer 1

Ingestion

collected documents become text, one case per release

Turns collected documents into cases

# PDF and HTML → plain text → normalise → split text = extract_text(document) # pdfplumber for PDFs cases = case_batching(text, org="azicac") # one case per release # ids: {source}_{year}_{seq} · nj_ag_2017_001, wy_dci_2023_004

PDFs and web releases become one plain-text format, so every later step reads the same thing
Boilerplate and links are stripped, then each source's date and heading pattern splits the text into one case per release
A content fingerprint stops the same release entering the corpus twice
Person names are masked in everything public; only their character offsets are kept

Layer 2

Extraction

four readers, one decision, a span for every value

Reads each case into structured fields, with a span for every value

# readers run side by side on the same text regex · spaCy NER · Stanza NER · LLM residue pass # each emits (field, value, start, end, confidence) resolution → one value per field, disagreements go to review

Ages, dates, sentences, platforms, agencies, locations, investigation type and charges are deterministically extracted
Agencies resolve to a law-enforcement registry and places to the Census gazetteer, so the same name means the same thing across cases
Four readers vote on every field; where they disagree, the case goes to review instead of a guess
Audit shows every value beside the passage it came from

Layer 3

Storage

case table, provenance, graphs on disk, SPARQL store

Keeps the cases, the evidence and the graphs

PostgreSQL (Railway) · SQLite locally # cases, assertions, resolutions ontology/graph_output/*.ttl | *.jsonld # one CAC graph + one provenance graph per case Oxigraph behind /sparql # 10,802 case graphs, 10,802 provenance graphs, PACER

Every extractor's answer is stored, not only the winning value, so a resolution can be re-run or audited
Redis caches the heavy API responses, so the facet tree and cluster pages answer in milliseconds
Every graph passes SHACL validation before it enters the store, so a bad mapping never reaches a query

Layer 4

Knowledge graphs and analysis

CASE / UCO / CAC, w3id IRIs, state machines

Maps each case to CASE/UCO/CAC, with its provenance beside it

# deterministic mapping, no custom predicates features → CAC entities → Turtle + JSON-LD → SHACL → Oxigraph https://w3id.org/caselinker/resource/case/{id} # the case graph https://w3id.org/caselinker/resource/case/{id}/provenance # oa:Annotation per value

Every case, step, agency, platform and location has a persistent w3id IRI, so a result can be cited and dereferenced; Explore walks them
30 PACER cases are modelled as CAC phase state machines, so a court record can be read as phases and the affordance that enabled each step
Facet cohorts, clustering and triage run over the case table and come back through the API

CAC OntologyUCOCASEWeb AnnotationPROV

Layer 5

Interface

web app, REST API, SPARQL endpoint, MCP server

Serves the same corpus to people, scripts and agents

/analysis /patterns /explore /lifecycle /visualization /search /audit /sources /api/* # REST, OpenAPI at /docs /sparql # GET | POST, read-only /mcp /mcp-http # 55 tools for agents: corpus, facets, graphs, SPARQL, triage

Production runs on Railway with PostgreSQL; a local clone runs on SQLite against the live SPARQL store
Explore in the browser, query the public APIs from a script, or use an agent with the MCP server