References and source scope

Sources were reviewed for the November 1, 2026 research specification. A citation identifies related work or a source’s own claim. It does not establish access to proprietary data, permission to train on a source, independent replication, or an AMLOK result. Record an exact revision when a source or model becomes an experimental dependency.

Financial event and language models

Source	Role in this program
Slope TransFormer: an LLM for banks	Author-reported descriptor labeling method and narrow comparison, published in 2023.
Revolut PRAGMA, arXiv 2604.08649v1	Typed-event representation, self-supervision, and adaptation. Version 1 is the supplied reference.
RBC Borealis ATOM	Canadian financial foundation-model prior work; not an AMLOK benchmark.
Your Spending Needs Attention: Modeling Financial Habits with Transformers	Nubank nuFormer research; recommendation and representation scope.
TransactionGPT, arXiv 2511.08939	Transaction-native representation and prediction research.
BloombergGPT	Financial-language pretraining context; different from borrower-outcome validation.
LoRA	Parameter-efficient adaptation method.
Ditto: Deep Entity Matching with Pre-Trained Language Models	Learned entity-matching baseline and experimental context.

Cash-flow scoring and monitoring

Source	Role in this program
Prism: creation of CashScore v4	Vendor account of score development. The gated full whitepaper was not acquired.
Prism: five essential development standards	Multi-lender development and evaluation considerations.
Prism Pulse	Longitudinal financial monitoring context.
Prism support for Canada and Neo Financial	Existing Canadian cash-flow analytics activity.

Evaluation and baselines

Source	Role in this program
FinBen	Broad financial task taxonomy; no substitute for Canadian lending labels.
FinanceBench	Financial-document question-answering evaluation.
FinQA	Numerical reasoning with financial evidence.
TAT-QA	Reasoning over financial text and tables.
Revisiting Deep Learning Models for Tabular Data	FT-Transformer and strong tabular baselines.
TabPFN 3.5 technical report	Current tabular-model candidate; access and permitted use require checking for an actual run.
On Calibration of Modern Neural Networks	Probability calibration and validation methodology.
OpenAI model announcement, Anthropic model page, Google Gemini catalogue	Sources for candidate model availability. No hosted frontier model was evaluated in this package.

Canadian sources and governance

Source	Role in this program
Corporations Canada data services and API documentation	Federal registry scope and access mechanism.
Office of the Superintendent of Bankruptcy statistics	Aggregate insolvency context, not individual labels.
Statistics Canada WDS and open licence guidance	Public statistical acquisition and reuse context.
Bank of Canada Valet documentation and terms	Source for the executed FX descriptive analysis.
BDC debt-service coverage definition	Financial definition reference; agreement-specific ratios can differ.
CMHC GDS and TDS calculation	Mortgage affordability definitions within CMHC’s stated scope.
OSFI Guideline E-23 (2027)	Model-risk governance; revised guideline effective May 1, 2027.
Canadian privacy authorities’ generative-AI principles	Privacy, purpose, and accountability reference.
Québec private-sector privacy statute, section 12.1	Exclusively automated decisions within the provision’s scope; linked official version must be rechecked for an actual deployment.

The initial research consulted public source pages and selected AMLOK starter documents. It did not acquire competitor datasets, commercial model weights, private borrower histories, or a Canadian loan-outcome corpus. Public-data availability and source quality are evaluated separately from permission to collect, redistribute, or train.

Original records and acquisition sources

The public-source inventory contains the full source-by-source access and field register. The executed source analysis uses original EDC disclosures, Investissement Québec intervention records, CSBFP data, Kinross financing disclosure, Lundin Mining’s amendment, and BCE’s financial report. The SBA data portal supplies the inspected US loan schema. The original FinQA repository supplies the executed retrieval task; TAT-QA supplies development examples. FinanceBench’s publisher page identifies the open sample and its use conditions. CUAD is a candidate contract-clause benchmark, not an executed result here. The acquisition manifest records the exact file URLs and hashes. A citation alone is not evidence that the entire archive was collected, labels were adjudicated or a model was trained.

Research program

Data and models

Risk analysis

Evaluation and evidence

References and source scope

Financial event and language models

Cash-flow scoring and monitoring

Evaluation and baselines

Canadian sources and governance

Original records and acquisition sources

​Financial event and language models

​Cash-flow scoring and monitoring

​Evaluation and baselines

​Canadian sources and governance

​Original records and acquisition sources

Financial event and language models

Cash-flow scoring and monitoring

Evaluation and baselines

Canadian sources and governance

Original records and acquisition sources