Limits of explanation tools

Figure 01 / 12

Determine the eligible population first

Of 2,950 source records, 2,448 are within the stated synthetic scope and 502 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 state	Records
Within stated scope	2,448
Outside stated scope	502

A model explanation describes behavior under a particular method and reference. It does not itself establish causation, fairness, or that the final action used the model.

An attractive feature-importance chart substitutes for validation of the notice 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 2,448 eligible records from a source population of 2,950. The required workflow completes for 1,909, but 248 completed records miss the illustrative internal target. Another 539 remain incomplete. Communication evidence covers 1,661 generated notices. Scope, completion, timeliness, and delivery are four separate properties of the customer outcome.

Figure 02 / 12

A control can miss eligible records

The required workflow completes for 1,909 of the 2,448 eligible records. The 539 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

Measure	Records
Eligible records	2,448
Workflow completed	1,909
Workflow incomplete	539

Figure 03 / 12

Completion and timeliness are distinct outcomes

The illustration applies an internal target, not a statutory deadline. Of 1,909 completed records, 248 miss that target and 1,661 meet it. The 539 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

Outcome	Eligible records
Complete within target	1,661
Complete after target	248
Still incomplete	539

Figure 04 / 12

An obligation record connects authority to behavior

This case implements use explanation tools within their limits. 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

Element	Illustrative value
Control subject	Limits of explanation tools
Scope	The eligible population defined above
Trigger	An attractive feature-importance chart substitutes for validation of the notice reasons.
Required behavior	use explanation tools within their limits
Owner	model validation
Evidence	Versioned event, action, and communication records

Figure 05 / 12

Different clocks start from different facts

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

Event	Illustrative time	Record
Customer report	T0	Original channel and words
Institution receipt	T0 + 5 minutes	Retained receipt timestamp
Classification	T0 + 20 minutes	Applicable process and owner
Internal review target	T0 + 1 day	Internal target only
Outcome communication	At decision	Content, destination, and delivery state

Figure 06 / 12

Evidence fields fail independently

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 field	Present	Absent
scope	2,252	196
trigger	2,154	294
action	2,081	367
notice	2,203	245
evidence	1,958	490

Figure 07 / 12

A generated notice is not a delivered notice

1,909 completed records generate a modeled notice event. 1,661 have a delivered state and 248 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 state	Notices
Generated	1,909
Delivered state recorded	1,661
Delivery unresolved	248

Figure 08 / 12

Authority differs by operation

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

Role	Read evidence	Propose action	Approve release
model validation	Scoped	Yes	No
Independent approver	Scoped	No	Yes
Support	Limited	Request only	No
System administrator	Operational logs	No	No

Figure 09 / 12

Exceptions need capacity and a closing state

The control has 539 incomplete records. The available exception capacity covers 120, leaving 419 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 item	Records	Meaning
Exceptions opened	539	Eligible workflow incomplete
Capacity applied	120	Records handled in this window
Pending exceptions	419	Still require an owned response

Figure 10 / 12

A rate includes its denominator

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

Metric	Numerator	Denominator	Percent
Eligible coverage	1,909	2,448	77.98
On-time among completed	1,661	1,909	87.01
On-time among eligible	1,661	2,448	67.85
Delivered among generated	1,661	1,909	87.01

Figure 11 / 12

A change needs an evidence trail

The trigger is An attractive feature-importance chart substitutes for validation of the notice reasons.. 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

Stage	Retained proof
Interpret	Scope, source, effective date, and owner
Implement	Versioned logic, data contract, and message template
Verify	Boundary cases and affected-population comparison
Release	Approval, start time, and rollback condition
Correct	Affected records and customer outcome where required

Figure 12 / 12

Correction follows the affected population

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

From	To	Relationship
Limits of explanation tools	Affected population	Reproducible query
Affected population	Financial review	Amount and balance impact
Affected population	Customer message	Required communication
Financial review	Closure evidence	Verified adjustment
Customer message	Closure evidence	Delivery and follow-up

Connect the result to the system

Test the explanation against the actual action logic and disclose the method’s limits.

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.