CaseNoesis aggregates, structures, and decomposes case data across offense types, modeling exploitation as state machines and Markov decision processes to understand how platform affordances are misused, how crime types evolve alongside technology, and where intervention is possible.
The theoretical foundation for this project is "Affordances for Harm: How Offenders Misuse Platform Capabilities to Exploit Children, and Where to Intervene"
AfH develops a formal affordance–misuse–harm framework (φ/η/ψ mapping) and validates it against a corpus of ICAC enforcement cases. CaseNoesis is the empirical test of whether that framework generalizes across offense types beyond the one it was originally derived from.
- Multi-offense ingestion (fraud, trafficking, cyber-enabled crime, ICAC/CSEA)
- Local-machine collector (
collector/) plus local stdio MCP to orchestrate it (casenoesis_mcp) - Website/docs may run on Railway. There is no public MCP. CaseLinker remains the public ICAC collector and query API.
- Local FastAPI + sqlite for the researcher site and corpus tools (
python3 run/main.py)
45,877 distinct public records are collected (46,619 rows in the lookup under data/collected/public/, before repeated addresses and identical files are removed). The extraction engine has read all of them: 3,126,758 span-anchored assertions, 992,008 settled values, 58,583 mentions left for a person. Defendants are linked across the records of a case, and each term and amount is placed with its defendant. What happened in each case is ordered into a dated timeline: 108,740 events in 44,152 cases. Process mining over those timelines finds the paths prosecutions take and flags cases that depart from procedural order. All 864 court filings and 68 reference records, and a 300-record press sample, also have a CASE/UCO record graph from the pinned SDK with names removed. The engine writes its own graph per record, with defendants, charges, terms, and amounts, validated against the same SDK.
Accuracy is measured on a gold set of 200 invented records written from known answers (gold/). On its held-out half the engine scores F1 of 0.98 to 1.00 on terms, amounts, dates, citations, dockets and defendants, and 0.85 to 0.94 on victim counts and scheme and channel labels.
The Exploitation State Machine is fitted to the acts read from records and tested, not assumed. Its states are data in two layers: milestones aligned to the CAC backbone, and tracks beneath them, with the SDK's domain alphabets placed under them, so any trajectory graph runs through the same machine. Domains are compared at each layer and at the act and technology levels, and a goal-conditioned decision model is checked against observed choices. See state_machines/.
Not done: no sample of collected records has been checked by hand, so there is no accuracy figure on real records. No public data release yet.
Public enforcement records become span-anchored values, then validated knowledge graphs, then (not yet) state machines. Full design: Architecture design.md.
flowchart TD
P["PRESS / AGENCY <br/> scale · discovery"]
C["FREE RECAP / PACER <br/> lifecycle fidelity"]
F["REFERENCE <br/> statutes · platform policy"]
P --> IDX["RECORD INDEX <br/> content-hashed ids · domains"]
C --> IDX
F --> IDX
IDX --> SDK["SDK PROCESSOR <br/> pinned CASE/UCO SDK · OCR <br/> record graph, names removed"]
IDX --> ENG
SDK -->|"text + anchors"| ENG["EXTRACTION ENGINE <br/> readers → span gate → vote by family <br/> review queue · domain profiles"]
ENG --> STORE["STORE <br/> assertions · decisions · cases <br/> people · events"]
STORE --> PM["PROCESS MINING <br/> timelines per defendant · variants <br/> Inductive Miner · conformance"]
GOLD["GOLD SET <br/> 200 invented records, known answers"] -.->|"scores"| ENG
STORE --> KG["ENGINE GRAPH <br/> CASE/UCO + legalproc <br/> defendants · charges · terms · amounts <br/> SHACL gate"]
STORE --> ESM["STATE MACHINE <br/> M = (S, A, T, R_G, s₀, F) <br/> layered states · invariance · decision model · laws"]
ESM --> KG
KG --> OX["LOCAL TRIPLE STORE <br/> Oxigraph + graphs.nq · SPARQL"]
STORE --> API["python -m extraction · python -m process_mining <br/> local stdio MCP · SPARQL on 127.0.0.1"]
PM --> API
OX --> API
Three collection paths. Press for scale. Free RECAP for court filings; PACER only with a spend cap. Reference records for platform rules, statutes, and calibration. They meet at processing. See Collection.
How a value gets stored. Each reader writes its own assertion with a character span. One gate checks the span and re-reads the value from the text. Readers vote by evidence family: two families that agree settle a value, two that disagree veto it, and the rest goes to a review queue. The core knows fields and votes; fraud, trafficking, forced labor, cyber, and CSEA are vocabulary files. Code: extraction/.
People, events, paths. Readers also vote on whose each term and amount is and on what each record says happened. Defendants are linked across a case's records, events are dated and ordered per defendant, and process_mining/ mines the common paths and checks each case against procedural order. Code: extraction/people.py, extraction/events.py.
No names. Source text stays on the machine. Assertions, graphs, and tool output carry no person's name. A defendant is Defendant 1 with a role, linked to charges and terms.
All local. Clone, set up, run. The store is a SQLite file, the graphs load into a local Oxigraph store and one N-Quads file, and the MCP server and SPARQL endpoint run on this machine. Nothing here needs a hosted database.
Graphs. CASE/UCO with the SDK's legal-process extension, built and validated with the SDK pinned in this repo. SHACL is a publish gate, and inferred analytics are never typed as observed facts. The SDK path by itself: docs/sdk-encoding.md.
What the system is for. The primary purpose is to build formal models, specifically the Exploitation State Machine. Analysis tests Theorem 1 and Laws 1–4 across domains, annotates affordances, and ranks intervention points. The local website and stdio MCP expose machines and collection tools to the researcher — not to the public internet. CaseLinker remains the public ICAC collector and query API.
Collection tools live under collector/. The pipeline is the same whether you run those scripts on the CLI or an agent calls them through local MCP (casenoesis_mcp/, stdio only). Outputs land in data/collected/ and do not auto-ingest.
Press, free court records, and reference records are the three paths. A new topic is a profile, not a new pipeline. The public index is data/collected/public/*_lookup.jsonl: one row per record, no article text, no PDF path. Article text and PDFs stay on the machine. To fetch those rows again: python3 -m collector.reproduce.
Court filings enrich a press case. Search is CourtListener. Download is a filing already in free RECAP. PACER is the paid backup under collector/pacer/, and only with a spend cap. State open-records requests (FOIA, Georgia, Florida) are opt-in and are not scrapers.
Local MCP is stdio only (python -m casenoesis_mcp.server). It is not on Railway. Tool catalog: casenoesis_mcp/tool_registry.md. Corpus tools need python3 run/main.py. Collector tools do not.
Case data is drawn exclusively from publicly available enforcement records (press releases, court filings already in the public domain) and from open-records releases obtained through lawful request. CaseNoesis public-record collection is completed in accordance with UMass HRPO NHSR #8252 (16 Sep 2026).
Contributors can help by:
- Proposing offense categories and taxonomy structure
- Ingestion pipeline design for new offense categories
- Code implementation
CaseNoesis builds on ideas developed in CaseLinker, a CSEA-focused case analysis platform. The extraction engine keeps CaseLinker's design (one assertion per reader, a span gate, a vote by evidence family) and makes it domain-general. What was carried over and what was not: Architecture design.md, section 15.