> ## Documentation Index
> Fetch the complete documentation index at: https://docs.amlok.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Research model card and evidence status

> Scope, data, implemented baselines, proposed model components and validation boundaries for the public-record research release.

This card describes the public-record research collection and completed studies through release **0.3**. Research versioning is separate from API version 1.4.2. The measured systems are structured-loan baselines and financial-document retrievers.

## System and data

| Item | Evidence in this release |
| - | - |
| Intended research use | Canadian financial-document understanding, transaction interpretation, entity relationships, underwriting and monitoring. |
| Acquired public evidence | Canadian financing disclosures, government aggregates, issuer reports, registry documentation and released financial-document datasets. |
| Executed benchmark | TF-IDF and BM25 evidence retrieval on 1,147 FinQA test questions from 380 source filenames. |
| Best measured baseline Recall\@5 | BM25 73.90%; complete annotated evidence retrieved for 61.81% of questions. |
| Supervised model training | Logistic regression and gradient-boosted trees, trained on 19,135 SBA observations; 5,264 validation and 11,996 held-out evaluation observations. |
| Supervised target and measured AUROC | Reported charge-off within 36 months; logistic regression 0.72284 and gradient boosting 0.71820. |
| Production foundation-model checkpoint | Not produced by this package. |
| Transaction-language, temporal and graph models | Research specifications and ablation protocols. |
| Hosted frontier comparison | Not run; no GPT, Claude or Gemini prediction records in the result artifacts. |
| Canadian borrower-outcome validation | Not completed; source disclosures do not constitute an outcome-linked private ledger cohort. |
| Independent validation or regulatory approval | Not established by the package. |

## Intended use

Use the source inventory to assess field feasibility, the original records to develop and challenge extraction, the measurements to reproduce a retrieval baseline, and the protocols to execute controlled model comparisons. Preserve the original distinction between signed support, commitments, drawn balances, registered rights and outcomes.

Use task-specific model cards for language extraction, entity resolution, financial reasoning, business risk, consumer risk and monitoring. A retrieval result is evidence about retrieval; it does not validate the complete lending workflow.

## Data limitations

Public issuers differ from private SMEs and consumers. Disclosure practices select which events become observable. Registry and court records can overrepresent formal security and distress. Missing information can reflect confidentiality, access conditions, publication lag or extraction failure rather than financial weakness.

FinQA is a public benchmark of financial reports, with possible exposure in general-model training. Its source filenames are the clustering unit for the reported uncertainty; they are not independent Canadian institutions. The SBA collection contains six complete files. The supervised cohort is an approximately 10% hash sample of eligible 2010-2016 approval records with year-separated and normalized-entity-separated partitions. Current-snapshot predictors still require point-in-time review. FinanceBench was inspected but not scored under commercial use. No source is counted as acquired merely because its documentation or URL exists.

## Performance and unsupported claims

Supported performance statements are the structured-loan baseline and lexical-retrieval measurements in [public results](https://docs.amlok.io/research/public-results.md). It supports neither AMLOK superiority over frontier models nor a first-in-Canada foundation-model claim. Canadian financial-model prior work appears in the [related-work review](https://docs.amlok.io/research/related-work.md).

Do not use these baseline scores as lending probabilities, approval thresholds, fraud determinations, loss forecasts or production service commitments. The research has not demonstrated causal loss reduction, approval expansion, regulatory compliance of a trained model, or national representativeness.

## Required evidence for a trained-model release

Record the actual base checkpoint and weight digest, training corpus and purposes, source licences, token/event counts, architecture, hardware, run configuration, split manifests, independent labels, predictions, calibration, costs, failures, subgroup results and validation findings. Freeze evidence cutoff rules and record each test access.

For an improved result, report the matched comparator, effect size, uncertainty and operational tradeoff. For a negative result, retain the same artifacts and explain the limitation. The [frontier protocol](https://docs.amlok.io/research/frontier-benchmark.md) defines the comparison conditions.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.