Research specification and initial experiments - November 1, 2026. This release contains a literature review, proposed architecture and training protocol, a real public-data descriptive analysis, and an executed synthetic transformer study. It does not report a trained production foundation model, observed borrower outcomes, or completed comparisons against commercial or frontier models. Each result identifies its evidence class.
Abstract
Financial records describe different stages of the same economic activity. A sale can appear as an order, invoice, payment, settlement, bank deposit, and accounting entry. Combining those records without resolving their relationships overstates activity. Examining only bank deposits can hide margins, obligations, reversals, and changes that precede settlement. The proposed AMLOK architecture therefore separates transaction interpretation, economic-event reconciliation, temporal entity resolution, financial calculation, and outcome prediction. The core hypothesis is that this separation improves evidence quality and permits a shared representation to transfer across underwriting and monitoring. It remains a hypothesis until tested on authorized, representative Canadian cohorts. A secondary hypothesis is that a relatively compact specialist model can match or exceed a general-purpose language model on a tightly defined task at lower cost. Model size and technical novelty are not substitutes for this evidence.Research questions
Read the program
Scientific context
What prior work establishes and which comparisons remain open.
Data and evidence
Source rights, point-in-time data, labels, and Canadian context.
Model and training
The proposed representation, training stages, and ablations.
Evaluation and results
Task-specific benchmarks and the initial synthetic experiment.