Source layers
“Third party” describes the external system, not the source of authority. Borrower permission to connect a system is distinct from permission to train a shared model. Record permitted purposes for operational processing, institution-specific adaptation, shared training, evaluation, retention, and redistribution separately. No cross-workspace training is assumed.
Required acquisition manifest
For every dataset, record source and owner, delivery mechanism, schema/version, acquisition time, content digest, covered period, jurisdiction, language, consent or other authorized basis, license version, allowed purposes, derived-data restrictions, retention, and deletion obligations. Keep secrets and direct identifiers outside the experiment manifest. Pseudonymous subject IDs remain potentially personal information. The source inventory records actual acquired files separately from documented acquisition routes. The deal-corpus protocol defines record, event, facility and outcome counts. The data-point specification maps fields to supporting evidence. A provider integration or literature citation alone does not establish possession of a training corpus.Cohorts and denominators
Define business and consumer cohorts independently. Stratify businesses by industry, size, province, language, lender, product, and available systems. Define consumer cohorts by lending product, income pattern, financial-history coverage, and authorized demographic audit variables. Newcomers, thin-file applicants, seasonal businesses, and self-employed consumers need explicit coverage analyses. Maintain a population flow: eligible applications, permission requested, permission granted, sources connected, adequate history, successfully processed, decision observed, and outcome matured. Report exclusions and reasons at each step. Connection success is not a random sample. Compare included and excluded groups using permissible variables; missing data must not silently become evidence of financial distress.Independent labels
Use separate labels for counterparty identity, transaction purpose, economic-event linkage, contractual delinquency, charge-off, insolvency, confirmed fraud, covenant breach, and liquidity stress. A registered security interest is not a default. A loan approval is not evidence of successful repayment. A model-generated label is a weak label until independently checked. For identity and reconciliation, use two qualified reviewers blinded to model predictions and resolve disagreements with a third adjudication step. Preserveunknown and multiple plausible candidates. For credit outcomes, record the horizon, grace periods, restructuring treatment, cure, prepaid loans, censoring, and policy changes. Do not treat an applicant declined by the lender as a known default or a successful loan.
Time and leakage
Retain both event time and the earliest verified availability time. At evaluation cutofft, require event time and availability time to be no later than t. Apply this rule to aliases, registry versions, accounting restatements, reviewer notes, macro releases, graph edges, and labels used as features. Later information can establish an outcome, but cannot enter a historical prediction.
Split by time and related-subject group; hold out institutions separately. Freeze training normalization, dictionaries, candidate indexes, prompts, and retrieval corpora before scoring the holdout. Pretraining on the holdout’s unlabeled histories is transductive access and must not be concealed as an inductive test.