Worked case · Controlled response · 12 figures

Actual adverse-action reasons

Generate accurate adverse-action reasons

Figure 01 / 12

Determine the eligible population first

Determine the eligible population first — Actual adverse-action reasons. Count; records. Exact values are in the figure data below.
Count; records

Of 4,250 source records, 3,060 are within the stated synthetic scope and 1,190 are outside it. Eligibility here is an explicit teaching input, not a legal conclusion. Production classification must use the actual entity, product, activity, jurisdiction, and facts.

Figure data and text version
Scope stateRecords
Within stated scope3,060
Outside stated scope1,190

A notice must reflect the actual reasons for the action under the applicable requirements. A plausible explanation for another component is not the decision path.

Trace inputs, rules, model results, overrides, and accurate customer-facing reasons.

All amounts, rates, capacity limits, and outcomes in this case are synthetic. The three conditions are separate assumptions for comparison. A better result in the response condition is not measured proof that the proposed control causes that improvement. The figures expose the calculation and its limits; a real deployment needs its own evidence.

Read the result

The case identifies 3,060 eligible records from a source population of 4,250. The required workflow completes for 3,045, but 15 completed records miss the illustrative internal target. Another 15 remain incomplete. Communication evidence covers 3,039 generated notices. Scope, completion, timeliness, and delivery are four separate properties of the customer outcome.

Model inputs and calculated values

Inputs below are the case-specific values. Each figure states the condition-specific assumptions and units used in its calculation. Calculated values are rounded for display.

InputValue
population4,250
eligibility0.72
Calculated valueResult
population4,250
eligible3,060
excluded1,190
complete3,045
incomplete15
late15
ontime3,030
notices3,039
undelivered6
pending0
reviewed15
Figure 02 / 12

A control can miss eligible records

A control can miss eligible records — Actual adverse-action reasons. Count at a fixed observation cutoff. Exact values are in the figure data below.
Count at a fixed observation cutoff

The required workflow completes for 3,045 of the 3,060 eligible records. The 15 remainder needs an owned exception path. Reporting completion as a percentage of all source records would answer a different question and could hide the actual coverage gap.

Figure data and text version
MeasureRecords
Eligible records3,060
Workflow completed3,045
Workflow incomplete15
Figure 03 / 12

Completion and timeliness are distinct outcomes

Completion and timeliness are distinct outcomes — Actual adverse-action reasons. Count; exclusive states within eligible population. Exact values are in the figure data below.
Count; exclusive states within eligible population

The illustration applies an internal target, not a statutory deadline. Of 3,045 completed records, 15 miss that target and 3,030 meet it. The 15 open records are a third state; do not automatically classify them as timely merely because their final outcome is unknown.

Figure data and text version
OutcomeEligible records
Complete within target3,030
Complete after target15
Still incomplete15
Figure 04 / 12

An obligation record connects authority to behavior

An obligation record connects authority to behavior — Actual adverse-action reasons. Control contract. Exact values are in the figure data below.
Control contract

This case implements generate accurate adverse-action reasons. The record separates scope, trigger, required action, ownership, and retained proof. Exact legal duties belong to the applicable source and interpretation; the timing and counts in this worked example are synthetic.

Figure data and text version
ElementIllustrative value
Control subjectActual adverse-action reasons
ScopeThe eligible population defined above
TriggerA model explanation is sent even though a policy override caused the final result.
Required behaviorgenerate accurate adverse-action reasons
Owneradverse-action process owner
EvidenceVersioned event, action, and communication records
Figure 05 / 12

Different clocks start from different facts

Different clocks start from different facts — Actual adverse-action reasons. Internal teaching timeline; not a legal deadline schedule. Exact values are in the figure data below.
Internal teaching timeline; not a legal deadline schedule

These relative times are illustrative service targets. They deliberately distinguish customer contact, receipt by the institution, classification, investigation, and communication. A routing delay must not silently replace the original receipt time when that fact matters.

Figure data and text version
EventIllustrative timeRecord
Customer reportT0Original channel and words
Institution receiptT0 + 5 minutesRetained receipt timestamp
ClassificationT0 + 20 minutesApplicable process and owner
Internal review targetT0 + 1 dayInternal target only
Outcome communicationAt decisionContent, destination, and delivery state
Figure 06 / 12

Evidence fields fail independently

Evidence fields fail independently — Actual adverse-action reasons. Count; overlapping field-level checks. Exact values are in the figure data below.
Count; overlapping field-level checks

Each row is one evidence requirement over the eligible population. The same record can fail several checks, so the absent counts across rows must not be added as though they were distinct customers. Completeness does not itself prove that a field is accurate.

Figure data and text version
Evidence fieldPresentAbsent
scope3,0546
trigger3,0519
action3,04515
notice3,0546
evidence3,04812
Figure 07 / 12

A generated notice is not a delivered notice

A generated notice is not a delivered notice — Actual adverse-action reasons. Count; generated equals delivered plus unresolved. Exact values are in the figure data below.
Count; generated equals delivered plus unresolved

3,045 completed records generate a modeled notice event. 3,039 have a delivered state and 6 do not. The system must distinguish generation, dispatch, delivery evidence, and any required follow-up under the actual process.

Figure data and text version
Communication stateNotices
Generated3,045
Delivered state recorded3,039
Delivery unresolved6
Figure 08 / 12

Authority differs by operation

Authority differs by operation — Actual adverse-action reasons. Illustrative permission matrix. Exact values are in the figure data below.
Illustrative permission matrix

The access matrix is a proposed teaching separation of duties. Read, propose, approve, and administer are distinct capabilities. The final policy must match the organization’s actual roles and obligations, with controlled emergency access and an audit trail.

Figure data and text version
RoleRead evidencePropose actionApprove release
adverse-action process ownerScopedYesNo
Independent approverScopedNoYes
SupportLimitedRequest onlyNo
System administratorOperational logsNoNo
Figure 09 / 12

Exceptions need capacity and a closing state

Exceptions need capacity and a closing state — Actual adverse-action reasons. Records per observation window. Exact values are in the figure data below.
Records per observation window

The control has 15 incomplete records. The available exception capacity covers 15, leaving 0 pending. A pending state requires an owner and a next action; changing a status label without resolving the required behavior does not close the gap.

Figure data and text version
Queue itemRecordsMeaning
Exceptions opened15Eligible workflow incomplete
Capacity applied15Records handled in this window
Pending exceptions0Still require an owned response
Figure 10 / 12

A rate includes its denominator

A rate includes its denominator — Actual adverse-action reasons. Percent; named populations. Exact values are in the figure data below.
Percent; named populations

These rates deliberately use different populations. Overall throughput, eligible coverage, completed-record timeliness, and delivery evidence are not interchangeable. Each needs the same cohort, cutoff, and definition every time it is compared.

Figure data and text version
MetricNumeratorDenominatorPercent
Eligible coverage3,0453,06099.51
On-time among completed3,0303,04599.51
On-time among eligible3,0303,06099.02
Delivered among generated3,0393,04599.8
Figure 11 / 12

A change needs an evidence trail

A change needs an evidence trail — Actual adverse-action reasons. Control-change lifecycle. Exact values are in the figure data below.
Control-change lifecycle

The trigger is A model explanation is sent even though a policy override caused the final result.. A controlled change connects the revised requirement or interpretation to implementation, replay, customer impact, and approval. The old version remains relevant to decisions already made under it.

Figure data and text version
StageRetained proof
InterpretScope, source, effective date, and owner
ImplementVersioned logic, data contract, and message template
VerifyBoundary cases and affected-population comparison
ReleaseApproval, start time, and rollback condition
CorrectAffected records and customer outcome where required
Figure 12 / 12

Correction follows the affected population

Correction follows the affected population — Actual adverse-action reasons. Illustrative correction responsibilities. Exact values are in the figure data below.
Illustrative correction responsibilities

A remediation map links the defect to affected records, financial consequences, communication, and closure evidence. It should retain exclusions and unresolved cases. A change that prevents future failures does not by itself correct earlier customer outcomes.

Figure data and text version
FromToRelationship
Actual adverse-action reasonsAffected populationReproducible query
Affected populationFinancial reviewAmount and balance impact
Affected populationCustomer messageRequired communication
Financial reviewClosure evidenceVerified adjustment
Customer messageClosure evidenceDelivery and follow-up

Connect the result to the system

Trace inputs, rules, model results, overrides, and accurate customer-facing reasons.

Check the population, evidence, permitted action, and actual effect together. A balanced calculation can still use the wrong population; a successful response can still leave an unknown financial outcome. The case’s numerical result applies only to its stated assumptions.

Sources and further reading

The chapter sources support the concepts and scope. They do not prescribe the synthetic model rates.

  1. Regulation B, 12 CFR 1002.6: evaluation of applications
  2. Regulation B, 12 CFR 1002.9: notifications
  3. NIST: AI Risk Management Framework