Part 10 — System Blueprint

BRE Functional Specification

This page specifies the functional contract for the BRE and decision service named in module map. It computes the policy outcomes described in underwriting decisioning and the credit appraisal package, then returns a reproducible decision object to LOS, credit workbench, the co-lending module and monitoring workflows.

The representation should align with Decision Model and Notation (DMN), the Object Management Group standard whose current formal version is DMN 1.4, adopted in April 2023 (OMG DMN 1.4). DMN is useful here because it separates a decision requirements graph from decision logic, decision tables, hit policies and expression language. The runtime may use a commercial or open-source engine, but the authoring model should preserve this structure so policy_risk, credit_manager, credit_committee_member and internal_auditor can read the policy without reverse-engineering application code.

Decision Domains

Domain	Caller stage	Main inputs	Main outputs	SLA class
Pre-screen	`prescreen_pending` in LOS workflow	Product, channel, PAN/GSTIN, pincode, negative lists, consent status, repeat-customer flags	`pre_screen_passed`, `hard_reject_triggered`, missing-data list, coarse reason codes	Synchronous, sub-second engineering target after features are present
Eligibility and pricing	`data_pending` to `credit_in_review`, straight-through offers, renewal/top-up	Bureau, banking, GST, ITR/financials, Udyam/KYC, obligations, collateral, product grid	Amount, tenor, FOIR/DSCR/turnover/MPBF eligibility, rate, fees, risk grade, binding constraint	Synchronous for complete data; async when scorecard/model call is unavailable
Deviation classification	`credit_in_review`, `deviation_pending`	Rule failures, severity matrix, exposure, channel, product, risk grade	`deviation` rows with severity, required `authority_level`, waiver expiry and mitigant prompt	Synchronous, because queue routing depends on it
Sanction-condition derivation	`approved`, `sanctioned`, `documentation_pending`	Approved structure, collateral, KYC/document gaps, product, partner conditions	`sanction_condition` rows for CPs, CSs, covenants, monitoring requirements	Synchronous during sanction generation
Monitoring/EWS re-score	LMS `active_current`, `active_overdue`, monthly or event-driven monitoring	Repayment conduct, bureau refresh, GST/banking feeds, stock statements, anchor data, overdue history	Watchlist tier, limit-freeze flag, covenant breach, renewal route, collections alert	Async batch plus near-real-time triggers
Collections treatment	`collection_case` state changes	DPD, exposure, product, collateral, PTP history, contactability, legal eligibility	Tele/field/legal strategy, settlement eligibility band, next action and contact controls	Async or same-day batch, not borrower-facing instant decision

Rule Representation

A ruleset is a versioned bundle of policies per product, lender and channel. The minimum bundle key is product_code, product_variant_code, business_line, regulated_entity_party_id, channel_type, effective_from, effective_to, ruleset_version, and status. For co-lending, the bundle also includes cla_id, both RE party ids, both policy versions and allowed split structure.

Representation element	Required fields	Build rule
Decision table	`decision_table_id`, input columns, output columns, hit policy, priority, effective dates	Use for policy grids: geography, vintage, bureau band, GST turnover, FOIR caps, LTV caps, fee bands and deviation matrix.
Hit policy	`hit_policy`, `aggregation`, conflict handling	Supported values: `UNIQUE`, `FIRST`, `PRIORITY`, `COLLECT`, `RULE_ORDER`. Reject publish when a table marked `UNIQUE` has overlapping rows.
Expression condition	`expression_id`, expression language, variables, null policy	Use FEEL-style expressions where practical. If the engine uses another expression language, store a DMN-readable equivalent and test vectors.
Scorecard node	`scorecard_id`, input features, bins, points/model version, score bands	Use for application or behavioural scorecards. Store score output and rank-ordered drivers, not only the final score.
Decision requirements graph	`drg_id`, nodes, dependencies, input data, knowledge sources	Model dependencies from data quality to eligibility, from eligibility to pricing, and from policy failures to deviations.
Ruleset bundle	`rule_set_id`, `rule_version_id`, artifact hash, approver, validation report	Immutable after publish. Any change creates a new version and canary pointer.

Input Contract

The evaluator receives a canonical request. It must not fetch arbitrary data during rule evaluation. Missing features are resolved before the call, or declared using explicit null semantics.

Field	Type	Source system	Freshness / consent	Nullability semantics
`application_id`	`uuid`	LOS `application`	Current application version	Never null
`application_number`	`text`	LOS	Current	Never null
`decision_point`	`enum`	Caller	`prescreen`, `eligibility`, `pricing`, `deviation`, `sanction_condition`, `monitoring`, `collections`	Never null
`idempotency_key`	`text`	Caller	One per snapshot, point and ruleset	Never null
`evaluation_timestamp`	`timestamptz`	Caller	Injected, not read from wall clock inside rules	Never null
`ruleset_version_requested`	`text`	Policy service	Active or explicit historical version	Never null
`product_code`	`text`	`application`	Current	Never null
`product_variant_code`	`text`	`application`	Current or default variant	Nullable only when product has no variants
`channel_type`	`enum`	`application`	Current	Never null
`source_partner_id`	`uuid`	`application`	Current	Required for `dsa`, `lsp`, `anchor`, `co_lending`
`primary_applicant_party_id`	`uuid`	`application`	Current	Never null
`party_features`	`jsonb`	KYC and AML service	KYC status current; BO status current	Hard fail if borrower KYC status absent at KYC-gated points
`bureau_features`	`jsonb`	Bureau integration / `external_verification`	Timestamp and `consent_id`; product policy defines max age	Refer or hard fail by policy when mandatory report missing
`banking_features`	`jsonb`	AA, BankConnect, upload analyzer	Statement end date, fetch timestamp, consent scope	Refer when mandatory; default only for explicitly optional metrics
`gst_features`	`jsonb`	GST provider / borrower upload	Return period and fetch timestamp	Refer for GST-surrogate programs; optional for non-GST entities
`financial_features`	`jsonb`	Credit workbench / spreads	Audited/provisional period and analyst version	Refer when method is `audited_financials`; nullable for `scorecard_only`
`udyam_kyc_features`	`jsonb`	Udyam/KYC service	Verification timestamp	Refer if MSME/PSL classification is required
`platform_anchor_features`	`jsonb`	`anchor_data_feed`, partner API	Batch id, consent scope, snapshot date	Optional unless partner program requires anchor data
`collateral_features`	`jsonb`	Collateral and security service	Valuation date, legal status, insurance status	Hard fail for secured disbursement; refer during early eligibility
`obligation_features`	`jsonb`	Bureau, banking, declared SOA	Snapshot timestamp	Refer if duplication unresolved
`co_lending_context`	`jsonb`	Co-lending module	Active `cla_id` and share constraints	Required when `co_lending_flag=true`
`manual_overrides_requested`	`jsonb`	Credit workbench	Maker, reason, evidence	Does not alter decision; only raises deviations

Null handling is policy, not code. A feature has one of four meanings: not_applicable, unknown_but_optional, missing_required, or provider_unavailable. missing_required creates a refer or hard fail by decision point. provider_unavailable may use cached data only if the output is marked stale and the ruleset permits it. Silent defaulting to zero is forbidden for income, turnover, DPD, bounces, GST delays, collateral value or obligations.

Output Contract

The decision object is persisted as bre_result.result_json, with selected values projected into typed fields on bre_result, deviation, credit_assessment, sanction_condition, partner_status_event and co-lending records.

Field	Type	Semantics
`bre_result_id`	`uuid`	Primary persisted result id.
`application_id`	`uuid`	Same as input.
`decision_point`	`enum`	Same as input.
`outcome`	`enum`	`approve`, `refer`, `decline`, `manual_review`, `error`. For legacy `bre_result.decision`, map to `pass`, `refer`, `fail`, `manual_review`, `error`.
`outcome_confidence`	`enum`	`deterministic`, `scorecard_band`, `model_assisted`, `fallback_cached`, `incomplete`.
`eligibility_by_method`	`array`	Each row: method `foir`, `dscr`, `turnover`, `mpbf`, `ltv`, `policy_cap`; input amount; calculated capacity; applied haircut; result.
`binding_constraint`	`text`	The method or policy cap that limited the final amount. Required when approved or referred.
`max_eligible_amount`	`numeric(18,2)`	Amount to project to `bre_result.max_eligible_amount`.
`max_tenor_months`	`int`	Tenor allowed by product, score band and repayment method.
`pricing_components`	`array`	Base rate, risk premium, tenor premium, collateral discount, partner spread, fees, taxes, APR-relevant charges and policy min/max bounds.
`price_floor_rate`	`numeric(9,6)`	Rate to project to `bre_result.price_floor_rate`, before final human pricing where allowed.
`conditions_precedent`	`array`	Draft `sanction_condition` rows with `condition_type=condition_precedent`, owner role and blocking flag.
`conditions_subsequent`	`array`	Covenant and monitoring items after disbursement.
`deviations`	`array`	`deviation_type`, `rule_code`, `severity`, text, mitigant prompt, `required_authority_level`, expiry and whether waivable.
`reason_codes`	`array`	Ranked reasons with code, human-readable CAM text, borrower-facing text when permitted, rule id and evidence values.
`model_inventory_refs`	`array`	Model ids, scorecard ids and validation status used in this run.
`co_lending_allocation`	`object`	See below. Required for co-lending decision points.
`audit_summary`	`object`	Trace id, input snapshot ref, ruleset version, result hash and timing.

The co-lending allocation block uses exact co-lending module names:

Field	Type	Semantics
`cla_id`	`uuid`	Active `co_lending_arrangement`.
`originating_re_party_id`	`uuid`	Originating RE from `co_lending_arrangement`.
`partner_re_party_id`	`uuid`	Partner RE from `co_lending_arrangement`.
`originator_policy_verdict`	`enum`	`approve`, `refer`, `decline`, with originator ruleset version.
`partner_policy_verdict`	`enum`	`approve`, `refer`, `decline`, with partner policy version or pending marker.
`eligible_pairings`	`array`	Allowed RE pairing, product, geography and borrower segment rows.
`split_percentages`	`object`	Proposed `originator_share_percent` and `partner_share_percent`; each must be at least 10 percent under RBI co-lending directions.
`blended_rate_computation`	`object`	RE rates, share percentages, weighted blended rate, fees and APR inclusion.
`partner_share_preview`	`array`	Draft `partner_share` amounts by RE.
`dlg_variant_flags`	`object`	DLG present, `dlg_set_id`, cap remaining, covered portfolio and blocked reason if ineligible.
`partner_conditions`	`array`	Conditions to convert into `sanction_condition` or partner-facing API messages.

RBI’s 2025 co-lending directions make the dual-policy basis operational: each RE must incorporate CLA provisions into its own credit policy, changes to each RE rate follow that RE’s credit policy, each RE maintains its own borrower account for its share, and both shares must be reflected without delay and no later than 15 calendar days (RBI Co-Lending Arrangements Directions, 2025). The BRE should therefore record both policy verdicts, not only a joint blended offer.

Decision Trace And Audit Schema

Field	Type	Required	Purpose
`trace_id`	`uuid`	Yes	Correlates request, evaluation nodes, logs and persisted result.
`idempotency_key`	`text`	Yes	Deduplicates retries and replay.
`rule_set_id`	`uuid`	Yes	Versioned rule bundle.
`rule_version_id`	`uuid`	Yes	Immutable published version.
`input_snapshot_ref`	`text`	Yes	Encrypted object reference for full input snapshot.
`input_snapshot_hash`	`char(64)`	Yes	Hash for replay and tamper detection.
`node_evaluations`	`array`	Yes	Each node has node id, node type, input values, output value, start/end time and errors.
`rule_evaluations`	`array`	Yes	Each row has rule id/version, values seen, threshold, fired flag, hit policy result and DRD path.
`scorecard_evaluations`	`array`	Conditional	Score, bins, points, rejected variables and top drivers.
`timing_by_node_ms`	`jsonb`	Yes	Latency analysis and regression monitoring.
`external_model_refs`	`array`	Conditional	Version-pinned model calls, feature vector hash and response hash.
`final_decision_hash`	`char(64)`	Yes	Hash of canonical output JSON.
`pii_masking_profile`	`text`	Yes	Describes redaction in logs and UI views.
`retention_until`	`date`	Yes	Must satisfy audit, model inventory and loan-record retention policies.

Full snapshots contain PII, bureau data, bank data and GST data. Store them encrypted, access-controlled through actors permissions, and masked by default. Operational logs should carry only ids, hashes, coarse bands and rule codes. A DSA, LSP or field user must never see raw bureau variables, CAM comments or hidden rule thresholds.

Ruleset Lifecycle

Step	Owner role	Required evidence	Exit control
Author draft	`policy_risk`	Business objective, affected product, regulatory basis, sample rows, expected approval/reject mix	Draft ruleset compiles and all tables pass overlap checks
Technical review	`system_service` / engineering	Schema compatibility, expression safety, performance test, no forbidden nondeterminism	Artifact build succeeds
Maker-checker approval	`policy_risk` maker, `credit_manager` or `national_credit_head` checker; `aml_compliance` for regulatory rules	Change note, old/new diff, expected portfolio impact	`approval_request.status=approved`
Simulation/backtest	`policy_risk`, risk analytics	Historical applications, outcomes, confusion matrix, approval mix, bad-rate estimate, reason-code distribution	Thresholds met or exception approved
Champion-challenger setup	`policy_risk`	Champion version, challenger version, sample split, stop-loss guardrail, monitoring dashboard	Publish as canary or shadow only
Publish	`system_service`	Immutable artifact hash, effective date, rollback pointer	Version pointer updated atomically
Monitor	`policy_risk`, `internal_auditor` read-only	Approval/refer/decline mix, override rate, drift, latency, rule-hit distribution	Alerts routed to owner
Rollback	`policy_risk`, `system_service`	Incident reason, affected versions, replay sample	Pointer reverts; decisions already issued remain immutable
Periodic review	`policy_risk`, independent validation function, RMCB/delegated authority for high-risk models	Validation report, model inventory status, monitoring, exceptions	Review recorded at least annually or earlier on trigger

The June 24, 2026 RBI draft Guidance on Regulatory Principles for Model Risk Management applies broadly to banks, NBFC layers, AIFIs, ARCs and CICs. It covers internal, third-party and AI/ML models, and treats decision-based rules and computational tools as models when they materially affect decisions (RBI draft model risk guidance transcript). The BRE model inventory should therefore include scorecards, pricing calculators, ruleset bundles, feature transformations and third-party model calls.

Explainability And Governance

Reason codes are first-class output. A reason code has a stable code, internal text, borrower-facing text if permitted, rule or scorecard driver, evidence values, adverse-action severity and language version. The top reasons should be ranked by actual contribution to the outcome. Do not return generic messages such as POLICY_FAIL. For a decline, use specific but safe reasons such as high recent delinquency, insufficient verified turnover, unsettled bureau overdue, unsupported DSCR, ineligible geography, prohibited industry or KYC defect.

KFS and APR interplay is downstream but triggered here. Pricing outputs must label whether a charge is APR-relevant and whether it is borrower-payable, partner-payable or waived. Under co-lending, the borrower-facing rate is a weighted blended rate, and additional fees must be included in APR disclosure as applicable under the co-lending directions and KFS circular. The BRE should not generate the KFS, but it must provide structured pricing components so the KFS archive can reproduce the disclosed economics.

Model governance controls:

Control	BRE implementation
Model inventory	Every ruleset, scorecard, pricing calculator and external model has an owner, validator, approver, risk tier and intended use.
Independent validation	Required before deployment, after material changes, on trigger events and periodically per MRMF. Third-party models still need RE validation.
Documentation	Store objective, assumptions, limitations, feature definitions, sample decisions, validation, approval, change history and decommissioning record.
Explainability	Store rule trace, score drivers, reason codes and borrower-safe text.
Human oversight	`refer` and `manual_review` outputs route to `credit_analyst`, `credit_manager` or the resolved DoA level; high-risk AI outputs need override/suspension controls.
Review cadence	At least annual tier review for all models, with earlier review on drift, material change, regulatory change, data-source change or performance breach.