Local explanations support accepting or contesting a single result

What it is

An insurance claim is denied. What the person has to decide is: accept this one, or contest this one. What can hold that decision up is an account locked to this case — which documents were used, which item did not match, which page is missing. The use of a local explanation is to support accepting or contesting a single result, not to teach how the model usually thinks.

Without that layer, acceptance is a blind signature and contestation is a blind essay.

Why it happens

Both accepting and contesting face a ruling that has already happened. To accept, a person must be able to judge whether the facts of this case were used rightly. To contest, they must be able to point to a piece of this case that can be supplemented or rebutted. The input to those acts is the grounds of this instance, not a group pass rate, and not an architecture diagram. Neighbourhood methods such as LIME fit this contestation setting because they try to mark what mattered near this instance — which is where a point of dispute may sit.

Global temperament gets in the way here: telling an applicant “the system is strict overall” neither lets them accept this one in peace nor tells them which page to add. Temperament is a deployment question, not a case procedure.

Studying it

Pair denials with a local account that locates facts of this case, or with overall policy copy only, then open “accept / submit supplement / formal appeal.” Independent variables: whether local grounds can be checked against a specific field, whether the appeal form is prefilled with the marked points of dispute. Dependent variables: relevance of supplements to those points, rate of groundless appeals, rate of appealing even when the facts were used rightly.

Relevance matters more than appeal rate. If the account only goads people into appealing with materials that do not match, the local explanation has not done its procedural job.

Where it stops holding

Settings that cannot be contested (the next song) do not need this procedure; if a local account is still there, its use is not accept-or-appeal. In high-stakes settings with no human review path, a local explanation fabricates a “you can contest this” appearance, which is worse than no explanation. This entry assumes a channel for accepting or objecting exists; how that channel is designed is another matter.

Applying it

• Beside a denial, list the facts used in this case, each openable to the source. Prefill the appeal form with those items and let people tick “wrong / can supplement,” rather than a blank essay.
• Near the accept control, restate in one sentence the items this conclusion depended on, not “this system is 92% accurate.”
• Policy or model introductions that were not grounds in this case go elsewhere, so applicants do not spend them as points of dispute.
• Check: sample appeal packs and see whether they aim at the fields the account marked. If not, the local explanation is not supporting contestation. Then see whether people who accepted can restate the grounds of this case; if they cannot, they signed blind.

Related

• Same group: L5.02.1 A local explanation accounts for a single output · L5.02.2 A global explanation accounts for overall behavioural tendency · L5.02.3 Users need different explanation levels in different situations · L5.02.5 Global explanations support judging whether to keep relying on the system · L5.02.6 An overall impression induced from a few local explanations is often wrong · L5.02.7 Feature importance is a correlation ranking, not a causal account · L5.02.8 Developers need debugging information; end users need grounds for action — they cannot share one explanation
• Nearby: L5.01 Types of Explainability · L5.08 Counterfactual Explanations · L5.05 The Moderation Principle of Transparency
• Search terms: local explanation for appeal · contest a decision · instance-level grounds