Deterministic core
A deterministic core is the layer of a lending platform whose outputs are fully reproducible from inputs and policy: no probabilistic models, no LLM calls, no hidden randomness.
- Recorded inputs
- Rule version
- Reproducible result
Two auditors, a year apart, ask why a particular account was classified NPA on a particular day. A deterministic core answers both the same way: here are the repayments as posted, here is the day-end days-past-due count, here is the classification rule in force, here is the result. If the answer could differ between the two visits, the lender has an explanation, not a record.
Interest calculations, repayment allocation, NPA classification, and the audit log all live here. A deterministic core gives the lender a system of record that auditors and regulators can interrogate. AI layers (workflow assistants, agents) sit above it and call into it; they never replace it. The architectural rule: the deterministic core is the truth, the AI layer is the help. Read as the sequence a single action runs through: AI proposes, the core decides, the record proves it. The essay AI in lending, the deterministic core makes the full argument.
On the live book the core is what turns an agent's suggestion into something a lender can stand behind. The agent may read a borrower conversation and propose a waiver, a date change or a settlement. The core recomputes the schedule, checks the proposal against the policy version in force and the approvals it requires, posts the result or rejects it, and writes the whole sequence to the audit log. Nothing probabilistic touches the ledger.