State model
A partitioned event log feeds time-windowed operational state. Per-entity sequence and watermark prevent late or duplicate events from appearing current.
Token streaming improves responsiveness but can expose an unsafe claim, disclosure or action promise before the complete answer is available for validation.
The response path separates low-risk conversational acknowledgement from held material segments, validates citations and policy at sentence boundaries, and releases only committed segments in order.
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.
A shared authority core coordinates independently owned capability cells, domain systems and operating evidence.
Permitted workThe router selects a certified route and context envelope. It does not inherit the selected specialist's business authority.
Consistency rulePreserve atomic intents, active journey, current custodian and route version across every turn and modality change.
Hard boundaryThe model is not a system of record, identity provider, policy authority or proof that an external effect occurred.
Events or conversation turns require a decision before the underlying situation changes, with strict latency and back-pressure constraints.
Amounts, eligibility, deadlines, legal effect and action confirmations remain buffered until deterministic validation succeeds.
A deterministic outer workflow contains model-led work inside typed, observable calls. Dashed messages remain proposals until policy or a human grants authority.
A partitioned event log feeds time-windowed operational state. Per-entity sequence and watermark prevent late or duplicate events from appearing current.
Use streaming ingestion for signals, low-latency feature or state reads for the hot path, and asynchronous enrichment outside the decision budget.
Point-in-time correctness takes precedence over the newest unqualified value. Late events trigger correction or review instead of mutating the old decision invisibly.
These roles are deliberately vendor-neutral. Each can be independently owned, versioned and replaced.
Evaluates identity, purpose, capability, amount, risk tier and policy version; returns allow, deny, step-up or human-review with reasons.
Chooses an approved model route by task, risk, evidence quality, latency budget and cost ceiling; enforces structured outputs.
Serves owned, audience-qualified and effective-dated content; exposes supersession, withdrawal and dependency metadata.
Runs component, route, trajectory, failure, harm and outcome tests against the versioned system manifest.
Appends request, versions, policy result, model proposal, approval, action receipt, readback, correction and custody events under one correlation key.
Durable records carry provenance, authority, effect and custody without turning a transcript into an uncontrolled memory store.
Preserve atomic intents, active journey, current custodian and route version across every turn and modality change.
The selected design is not universally superior. It is the safer fit for this boundary and failure cost.
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.
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.
Federate authoring while centralising lifecycle metadata, validation and serving rules.
Create one centrally authored knowledge corpus.
Cost acceptedFederation requires stronger contracts and owner discipline, but preserves domain accountability and release velocity.
Test prompts, models, tools, knowledge, policies, state transitions and human paths together.
Use a static answer-quality benchmark as the release gate.
Cost acceptedSystem evaluation takes longer and needs synthetic environments, but detects authority and recovery failures that answer scoring misses.
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.
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.
partition_rate = peak_events_per_second / active_partitionsdecision_budget = ingest + state_read + policy + score + action_commitbacklog_clear_time = queued_events / recovery_throughputFilter candidates and context before model classification so unauthorised capability descriptions or customer facts do not enter the prompt.
A design is production-ready only when teams can prove what happened, recover it and change it safely.
multi-intent and overlap test
low-confidence clarification
loop and maximum-hop test
unsupported-route and queue-rejection recovery
Draft segments, risk labels, validation results, citation coverage, release timestamps, suppressed content and final response hash.
Content and conduct owners define which semantic classes require pre-commit validation.
Start with advisory intervention and measured shadow scoring. Increase automation only when peak-load, late-event and fallback tests preserve the control outcome.
Journey, routing, conduct, channel and assisted-service owners define route and handoff policy.