The working A prototype: use its controls to test the build question.
Show where consequential risk enters and propagates through a real AI workflow.
RN conceived the question, structured the evidence boundaries, designed the experience, implemented the prototype and documented its limits.
This independent prototype demonstrates an approach; it is not a deployed client system, professional advice or proof of real-world outcomes.
Work through the controls with a situation of your own rather than the sample values. The tool responds to what you put in, so the useful output comes from real input, and everything is processed in your browser as you go. Any sample content you find already loaded is there to show the shape of a filled-in state, and you can clear it and start again at any point.
The build overview sets out the question, the purpose, the role and the evidence status in one place, and the public record behind it states what changed, what another person can reuse and what supports the system. If you want to judge how far this build should be trusted, the record is the page to read, not this one.
BUILD 013-A
AI Workflow Consequence Scanner
Trace risk from entry point to downstream action, then expose which owner, control, stop power, or monitoring link is missing.
WORKFLOW COVERAGE
100%INHERITED BUILD 012 STATE
CONTROLLED USERISK CONTAINED
Applicant intake
- Owner
- Intake operations lead
- Control
- Required-field and consent validation
- Stop authority
- Intake lead may halt incomplete or invalid cases
- Monitoring
- Daily missingness and consent exception review
Data preparation
- Owner
- Data stewardship lead
- Control
- Purpose limitation, minimization, and quality checks
- Stop authority
- Data steward may reject or quarantine a batch
- Monitoring
- Drift, missingness, provenance, and subgroup coverage
Model recommendation
- Owner
- Model assurance lead
- Control
- Versioned model, thresholds, evaluation, and uncertainty display
- Stop authority
- Assurance lead may disable a version or threshold
- Monitoring
- Error, calibration, bias, robustness, and override rates
Human eligibility review
- Owner
- Eligibility review supervisor
- Control
- Independent evidence review and documented override power
- Stop authority
- Reviewer may change, pause, or escalate any decision
- Monitoring
- Adverse decisions, overrides, disagreement, and review time
Notice and benefit action
- Owner
- Benefits operations director
- Control
- Pre-action approval, plain-language notice, and appeal route
- Stop authority
- Director may stop transmission and restore manual action
- Monitoring
- Delivery, appeal, correction, delay, and remedy outcomes
Outcome monitoring
- Owner
- Program accountability officer
- Control
- Thresholds trigger review, incident response, and control change
- Stop authority
- Officer may pause the workflow and initiate rollback
- Monitoring
- Continuous outcome review with subgroup and incident trends
Propagation paths
Incorrect or incomplete applicant data
intake → data → model → decision → action → monitor
EXPOSURE 48/100 · 0 uncontrolled stagesReviewer anchors on model recommendation
model → decision → action → monitor
EXPOSURE 80/100 · 0 uncontrolled stagesAdverse action notice or appeal route fails
action → monitor
EXPOSURE 60/100 · 0 uncontrolled stagesSCAN COMPLETE
Every mapped stage names an owner, control, stop authority, and monitoring rule.
Boundary: exposure is a prioritization heuristic—not a probability, legal conclusion, formal risk assessment, safety approval, or permission to deploy.
No workflow or personal data is transmitted or stored by this local demonstration.