CaseC-0626

Cracked windshield

A small claim the model was nearly sure of, just short of the certainty bar.

auto claim · $844

What the claimant wrote

Auto glass claim. A rock thrown up by a truck on the highway near Fresno, CA cracked the windshield of my 2020 Kia Sorento on 2026-06-05. The crack runs across the driver's line of sight. Hartwell Restoration quoted $844 for replacement including sensor recalibration. — Mei-Ling Zhou Some of the damage may have been there before; it is hard to tell what is new.

At a glanceEvery version

This claim through all ten versions

0Today

By hand

The claim joins one queue with every other claim and waits about 5.0 days for an adjuster to read it, check the policy and decide.

W0 to W3: turning the answers into a decision

W0 · Trust the one big answer

Sent to a person: flagged as needing a specialist. The right call
The steps it went through
1. Code R-HV-50K not fired: Claims over $50,000 require mandatory human review
2. Code R-INJECTION not fired: Suspected manipulation of automated processing quarantines the claim
3. Code R-FRAUD not fired: Claims with fraud indicators require investigation
4. Code R-POLICY not fired: Claims on a policy that is not in force go to a human (denials are never automatic)
5. Code R-TYPE-CONFIDENCE not fired: Uncertain claim type cannot be routed automatically
6. Code R-SPECIALIST FIRED: Claims needing domain expertise are assigned to a specialist Decided here
7. Person Disposition human_review by R-SPECIALIST
W1 · Combine the small answers

Sent to a person: flagged as needing a specialist. The right call
The steps it went through
1. Code Claim C-0626: $844, policy active.
2. Decision model The model answers fourteen typed questions about completeness, evidence and fraud; code decides, it never does.
3. Code Claims over $50,000 require mandatory human review
4. Code Suspected manipulation of automated processing quarantines the claim
5. Code Claims with fraud indicators require investigation
6. Code Claims on a policy that is not in force go to a human (denials are never automatic)
7. Code Claims needing domain expertise are assigned to a specialist Decided here
8. Person Sent to a person: R-SPECIALIST.
W2 · Learn how much to trust each answer

Sent to a person: not certain enough for the amount at stake. The right call
The steps it went through
1. Code Claim C-0626: $844, policy active.
2. Decision model The model answers fourteen typed questions about completeness, evidence and fraud; code decides, it never does.
3. Code Claims over $50,000 require mandatory human review
4. Code Suspected manipulation of automated processing quarantines the claim
5. Code Claims with fraud indicators require investigation
6. Code Claims on a policy that is not in force go to a human (denials are never automatic)
7. Code Fourteen learned weights score approval at 58.2% and a request for information at 79.3%; the higher one is proposed.
8. Code Not certain enough for a claim this size, so it goes to a person. Decided here
9. Person Sent to a person: below the confidence cut-off.
W3 · Be stricter as amounts rise

Sent to a person: not certain enough for the amount at stake. The right call
The steps it went through
1. Code Claim C-0626: $844, policy active.
2. Decision model The model answers fourteen typed questions about completeness, evidence and fraud; code decides, it never does.
3. Code Claims over $50,000 require mandatory human review
4. Code Suspected manipulation of automated processing quarantines the claim
5. Code Claims with fraud indicators require investigation
6. Code Claims on a policy that is not in force go to a human (denials are never automatic)
7. Code Fourteen learned weights score approval at 58.2% and a request for information at 79.3%; the higher one is proposed.
8. Code Not certain enough for a claim this size, so it goes to a person. Decided here
9. Person Sent to a person: below the confidence cut-off.

Around W3: the work people still do, and what else was tried

+ A quick confirmation for near-certain claims (now for requests only)

A person confirms the prepared file. The right call
The steps it went through
1. Code Not handled automatically: below the confidence cut-off.
2. Code Stopped only by the confidence cut-off, with a proposed action: an added step may still help.
3. Code Near enough to certain, or a clean prepared suggestion over $50,000: a quick confirm is offered.
4. Person A person confirms the prepared file.
+ Grade the fraud response

A person confirms the prepared file. The right call
The steps it went through
1. Code Not handled automatically: below the confidence cut-off.
2. Code Stopped only by the confidence cut-off, with a proposed action: an added step may still help.
3. Code Near enough to certain, or a clean prepared suggestion over $50,000: a quick confirm is offered.
4. Person A person confirms the prepared file.
+ A second vote on near-misses (a negative result)

A person confirms the prepared file. The right call
The steps it went through
1. Code Not handled automatically: below the confidence cut-off.
2. Code Stopped only by the confidence cut-off, with a proposed action: an added step may still help.
3. Decision model A second model did not name the same action.
4. Code Near enough to certain, or a clean prepared suggestion over $50,000: a quick confirm is offered.
5. Person A person confirms the prepared file.
+ Ask the customer one question

Settled with no person, after the customer replied. The right call
The steps it went through
1. Code Not handled automatically: below the confidence cut-off.
2. Code Stopped only by the confidence cut-off, with a proposed action: an added step may still help.
3. Decision model A second model did not name the same action.
4. Code A stored exchange for this claim: the enlarged claim is re-routed.
5. Decision model A question, drafted by a language model, is put to the customer.
6. Code The reply is stored, and the enlarged claim is re-analysed with it.
7. Code The enlarged claim is routed through the same fitted workflow and the same cut-off, once.
8. Code The enlarged claim clears the cut-off as an approval: handled with no person.
+ Watch after the decision

A person confirms the prepared file. The right call
The steps it went through
1. Code Not handled automatically: below the confidence cut-off.
2. Code Stopped only by the confidence cut-off, with a proposed action: an added step may still help.
3. Code Near enough to certain, or a clean prepared suggestion over $50,000: a quick confirm is offered.
4. Person A person confirms the prepared file.