State model
Append-only case events feed a current-state projection. Evidence rows keep source, observation time, validity, correction history and narrative usage.
Application, valuation, income, identity and property documents can disagree in ways that are clerical, explainable or material.
Narrow verification agents compare typed facts across artefacts, classify the disagreement and produce questions supported by both sides of the evidence.
Organisations, systems and operating conditions are intentionally anonymised and recomposed. The design demonstrates engineering and banking-domain reasoning; it does not represent a named client estate, vendor product or measured production result.
Typed commands, state changes and receipts cross an event spine without surrendering domain ownership.
Permitted workThe system may organise evidence, run calculations and propose scenarios. Credit judgement, affordability, exceptions and legal effects remain authorised decisions.
Consistency ruleBind evidence to applicant, facility, legal entity, period and policy version; preserve submitted, verified and inferred values separately.
Hard boundaryThe model is not a system of record, identity provider, policy authority or proof that an external effect occurred.
An alert or case requires evidence fan-out, deterministic triage, residual model reasoning and a defensible human disposition.
The system cannot infer misrepresentation or decide eligibility; low-confidence extraction stops the downstream check.
A deterministic outer workflow contains model-led work inside typed, observable calls. Dashed messages remain proposals until policy or a human grants authority.
Append-only case events feed a current-state projection. Evidence rows keep source, observation time, validity, correction history and narrative usage.
Use asynchronous ingestion and bounded parallel reads. Slow providers have independent timeout and circuit-breaker policy so one source cannot exhaust the case tier.
Pin policy and evidence versions at case open. Later corrections append and may trigger reassessment; they do not rewrite the historic basis silently.
These roles are deliberately vendor-neutral. Each can be independently owned, versioned and replaced.
Accepts a typed evidence request and returns the minimum permitted facts with source, event time, observation time, validity and exclusions.
Appends request, versions, policy result, model proposal, approval, action receipt, readback, correction and custody events under one correlation key.
Chooses an approved model route by task, risk, evidence quality, latency budget and cost ceiling; enforces structured outputs.
Presents claims beside evidence, alternatives, uncertainty, missing information, permitted actions and current custody.
Evaluates identity, purpose, capability, amount, risk tier and policy version; returns allow, deny, step-up or human-review with reasons.
Durable records carry provenance, authority, effect and custody without turning a transcript into an uncontrolled memory store.
Bind evidence to applicant, facility, legal entity, period and policy version; preserve submitted, verified and inferred values separately.
The selected design is not universally superior. It is the safer fit for this boundary and failure cost.
Prefetch stable, purpose-safe facts; acquire volatile facts on demand against a time-qualified snapshot.
Load a broad customer profile at session start.
Cost acceptedThe selected design adds source calls and latency, but reduces stale data, excess exposure and accidental reuse.
Use append-only events plus a rebuildable current-state projection.
Overwrite the case row with its latest status.
Cost acceptedReplay and storage are more complex, but point-in-time reconstruction and correction lineage remain possible.
Reserve larger models for residual reasoning after deterministic and smaller-model gates.
Send every request to the most capable available model.
Cost acceptedRouting adds evaluation work and operational complexity, but controls cost, latency and unnecessary data exposure.
Keep people at irreversible, ambiguous and policy-exception points; sample lower-risk automated outcomes independently.
Require the same manual approval at every step.
Cost acceptedRisk-tiering reduces review load but needs calibrated thresholds, sampling and immediate withdrawal of authority when drift appears.
Compile stable decision logic and retain retrieval for explanation and residual ambiguity.
Ask a model to interpret the source document for every request.
Cost acceptedRule compilation needs controlled change, but creates repeatable decisions, regression tests and clear exceptions.
Retries are bounded by knowledge of business effect; unknown outcome remains visible, owned and independently reconciled.
Actual thresholds belong to accountable service owners. The design exposes the equations and observables that those owners must baseline.
peak_enrichment_qps = peak_case_rate x sources_per_casereview_hours = residual_cases / investigator_throughputevidence_storage = cases x evidence_rows x retention_windowLimit applicant and guarantor data to the active assessment and separate sensitive attributes from model reasoning where policy requires.
A design is production-ready only when teams can prove what happened, recover it and change it safely.
calculation reconciliation
policy supersession and open-case test
document contradiction test
adverse-impact and override review
Artefact provenance, extracted values, confidence, contradiction pair, correction history and reviewer resolution.
An underwriter resolves material inconsistencies and makes the lending decision.
Use shadow decisions and blind samples before allowing deterministic closes. Keep agent-touched cases under human disposition until operating evidence supports narrower review.
Credit, product, legal, model-risk and lending-operations owners approve decision and action boundaries.