04:12 — The Second Night
The first night took a war room. In this designed target state the same class of allegation is contained, verified and packaged automatically overnight — and the MLRO receives exactly one message with three choices. Approve, reject into human review, or pick an alternative. The human is always allowed to be slower than the machine.
About this demonstration
Every scenario is a deterministic, synthetic demonstration set in a fictional universe (Aster Market, Lumen Digital, PSP-A/PSP-B). It shows a designed decision discipline — not a live deployment, a customer result or a production claim. Nothing is submitted, stored or tracked; the full boundaries sit under ‘What this does not prove’ at the end of each story.
- 01CaptureFix the signal, source and purpose before interpreting them.
- 02GovernApply the relevant scope, policy and mandatory boundary.
- 03DecideChoose a posture without hiding held or unresolved work.
- 04ChallengeExpose review, exceptions and the next accountable owner.
- 05ProveAssemble an inspectable artifact with its limits still visible.
Case DCK-DEMO-2044 — Lumen Digital
Allegation match on 1 listing
policy prohibited-items@demo-4.1
model risk-screen@demo-2 · conf 0.94 · DESIGNED
- Listing quarantined
- Payouts scoped-held
- KYB re-verified — no change
- Evidence pack EVP-DEMO-2044 assembled
What executed after your tap
- Scoped hold released · category ban applied
- 30-day enhanced monitoring scheduled
- Bank notified via the recorded template
- Decision record DR-DEMO-2044 signed with your approval
A maker action behind the checker gate
- Your pick becomes a maker proposal, not an execution
- The configured checker gate still applies
- Both options carry their predicted consequences into the record
Nothing executed
- The case routed to the morning queue with your reason
- Every automatic action stays in its contained state
- The record shows a human override — the machine waits
Hold the one decision that matters while the designed machine holds everything else.
The night ran to its single human gate: the release decision waits on your phone.
Seven automated steps executed and recorded; nothing beyond the gate has moved.
The designed model recommends; only the MLRO's decision releases anything.
The MLRO owns the release; the machine owns the record of what it did.
Every automated step carries its policy version, model version, confidence and input snapshot.
Why this matters
Automation earns trust when its one human gate is unmistakably real.
One decision, one tap, one record
Whatever you choose, the record shows what the machine did, what it recommended, and where the human stayed in charge.
Inspect Decision Record DR · DEMO
- case
- DCK-DEMO-2044
- merchant
- MER-DEMO-204 · Lumen Digital
- policy
- prohibited-items@demo-4.1
- model
- risk-screen@demo-2 · conf 0.94 · DESIGNED
- auto_actions
- quarantine · scoped hold · KYB re-check · evidence pack
- human_gate
- release decision · MLRO
- evidence_pack
- EVP-DEMO-2044
- correlation_id
- corr_demo_2044
Not recorded by engine — No value is inferred from surrounding data.
Night Log — one decision, one tap
The seven automated steps, the one human gate, and the decision you took at 04:12.
Designed operating model — target state, conditional on model deployment, validation and governance. Synthetic DEMO.
The night, automated
- Signal → score → quarantine → scoped hold → re-verify → case → evidence
- Every step stamped with policy and model versions
- Input snapshots preserved for reconstruction
The human gate
- The release decision waited for the MLRO
- Reject always routes to human review with nothing executed
- Alternatives remain maker proposals behind the checker gate
Design contract
- Model validation, drift monitoring and explainability are conditions, not claims
- No autonomous termination exists in the designed flow
- The ML indicator in the reviewed registry is a stub
The morning after
- A summary report stands ready for the 09:00 review
- The case and pack answer any bank or partner question
- The first night took a war room; this one took one thumb
What this shows
- How a designed automation path can keep every human control point explicit — and provable.
- How policy versions, model versions and input snapshots make an automated night reconstructable.
- Why reject is a first-class outcome: the human is always allowed to be slower than the machine.
What this does not prove
- No live model, measured false-positive rate or production automation is demonstrated; the registry's ML indicator is a stub.
- No autonomous account termination is claimed, designed or implied.
- Model validation, drift monitoring and governance are presented as required conditions, not as delivered evidence.
Aster Market
Each scenario adds a different synthetic artifact without changing the shared facts.
-
03:17Continuity Drill Brief03:17 — The Continuity RoomContinue with the same synthetic facts
-
09:00Shared Merchant Decision MemoOne Merchant — Five RealitiesContinue with the same synthetic facts
-
14:00Evidence Review ReportProof Lab — Try to Break the DecisionContinue with the same synthetic facts
-
Day 3RFI Response PackDay 3 — The FreezeContinue with the same synthetic facts
-
Day 9090-Day Pilot BlueprintBuild Your Control MapContinue with the same synthetic facts
-
04:12Night Log — one decision, one tap04:12 — The Second NightCurrent chapter
Explore all six chapters to close the night.
You made it to morning.
Six chapters, one connected record. This is the dossier of your visit — compiled in this browser from nothing but your allow-listed choices.
Compiled locally from allow-listed DEMO selections — nothing was transmitted or stored.
Chapters explored
Committed posture on this page
Boundaries honored
Nothing was force-routed; every stop stayed visible; the source records remained unchanged.
Pilot blueprint parameters
You just watched decisions become evidence. Now imagine it is your traffic.