Skip to main content
The proposed AMLOK system uses different components for different scientific tasks. A foundation model claim would require evidence of reusable pretraining and transfer across multiple tasks. A transformer layer alone does not establish that claim, and a financial event encoder is not necessarily a generative LLM.

​
Component contracts

​
Representation design

Represent an event as source, type, time, and typed key-value fields. Use learned categorical embeddings, scale-aware numeric encoding, text representations, missing-value indicators, and time-gap/calendar features. Preserve raw values outside the embedding so every reported calculation remains reproducible. For long histories, compare hierarchical event-to-day-to-month aggregation, retrieval of relevant evidence, and bounded local attention. Preserve material outliers, uncommon obligations, and original evidence references. Evaluate how each compression strategy affects low-frequency risk signals. The customer record allowance is not a model context length: 75,000 source records need not become 75,000 attention tokens.

​
Proposed model families

Evaluate a compact specialist text model, an encoder-only event model, a graph-enhanced variant, and a general-purpose model with identical retrieval and calculator tools. Select a base checkpoint only after verifying its licence, language coverage, provenance, deployment constraints, and available compute. Record its exact revision and digest. A pilot architecture search can vary event width, depth, windowing, and text adaptation. Treat configurations as candidates; do not invent production parameter counts, hardware inventories, training budgets, or achieved scaling laws. The executed public-data study establishes retrieval baselines for the document pathway. The frontier protocol specifies comparisons of the larger architecture without assuming those results.

​
What the design does not assume

A single latent vector cannot establish complete financial truth. Public sources do not reveal private transaction histories. Model scores need calibration and meaningful outcome definitions. A graph does not make correlations causal. An explanation can be fluent and still be wrong. These limitations determine the evaluation design, rather than being deferred until deployment.