The Claim That a Confidence Score Could Not Close

The story

08:42 — a high score enters the claim

Qualified adjuster Rina Patel receives a claim packet at Harbor Lantern Mutual. Its image model reports 0.94 confidence for one narrow class: “wind-pattern damage present in the supplied image set.” The result, CV-OUT-440, comes from model ROOF-CV-8 running against all twelve images in inspection revision INSPECT-R8.

The score may help Rina prioritize the review. It does not establish that the images belong to the covered property, that wind caused the damage, that the policy covers it, that the estimate is reasonable, that anyone acted improperly, or that Rina may settle the claim. Before she can make a coverage decision, the packet must show what the score actually describes and which evidence belongs to the dwelling.

08:45 — a plausible subtraction is rejected

A draft automation proposes:

$18,640 draft repair scope - $2,500 deductible = $16,140 proposed payable

Grid rejects the transition because the left-hand value is not an authorized covered amount. Before payment arithmetic can run, the packet still needs image-to-property mapping, cause and policy review, accepted scope, data-integrity state, adjuster identity and delegation, an exact human decision, and—if the payable amount exceeds the authored $10,000 limit—a supervisor disposition.

A deductible is not a bridge from an estimate to coverage. Before qualified authorization, payable is unresolved—not zero and not $16,140.

09:03 — missing explanation evidence narrows permitted use

The image vendor's explanation artifact is unavailable for fourteen minutes. The 0.94 result remains retained, but the declared region-of-interest map and input-quality notes required by authored review policy AI-CLAIM-5 cannot be retrieved.

The packet becomes incomplete-for-adverse-or-terminal-use. Grid does not assume that missing evidence supports the score or generate a substitute explanation.

09:17 — new inspection evidence changes the image set

Inspection steward Mateo Diaz publishes INSPECT-R9. Images IMG-09 through IMG-12 depict a detached shed, not the insured dwelling. The revision preserves R8, identifies each changed image, and records the correction basis.

ROOF-CV-8 is rerun only on the eight dwelling images and returns 0.71 confidence for the same narrow class. The score changed because the exact input changed. Its authority remains zero.

09:19 — one correction changes the whole packet

The correction advances a declared affected cone:

1. Detached-structure endorsement END-17 now requires separate human interpretation, making the prior dwelling-only coverage summary stale.
2. Four $1,100 shed lines move out of the dwelling draft, reducing it from $18,640 to $14,240.
3. The internal planning reserve falls from $18,640 × 1.15 = $21,436 to $14,240 × 1.15 = $16,376.
4. A provenance mismatch opens MANUAL DATA-INTEGRITY REVIEW without asserting misrepresentation, intent, claimant fault, or fraud.

The claimant notice draft, settlement proposal, reserve explanation, and model summary become stale. Historical R8 artifacts remain immutable.

09:24 — three paths preserve consumer protections

The authored model compares exactly three paths:

Path	Result	Why
Close at $16,140 using 0.94 and the original estimate	Rejected	Score scope, image membership, coverage authority, and amount basis are invalid.
Deny or refer as fraud because four images were misassigned	Rejected	A data mismatch does not establish intent, fraud, noncoverage, or an adverse decision.
Correct the evidence, hold terminal action, and complete qualified coverage and integrity reviews	Reviewable path	It preserves source correction, consumer-impact controls, authority, and external state.

Reviewable does not mean covered, denied, payable, or settled.

10:02–10:08 — human coverage authority remains separate from payment

The data-integrity reviewer attributes the mismatch to the scenario's inspection export mapping and closes the provenance exception without a fraud allegation. That disposition does not determine coverage.

Rina reviews the corrected inspection, the eight-image result, the applicable policy and endorsement, the $14,240 scope, the remaining source gaps, the model limits, the integrity finding, and her represented delegation. Under the scenario rules, she determines that exactly $11,900 is covered.

Only then may Grid calculate:

authorized covered amount = $11,900
deductible                = $2,500
payable                   = max(0, $11,900 - $2,500) = $9,400

Because $9,400 is below the authored $10,000 supervisor threshold, no supervisor signature is required. A negative case above that limit must stop and may not split payment to evade authority.

Rina signs DECISION-204-R5. The external payment system acknowledges instruction PAY-REQ-710. Coverage is authorized, but payment execution, bank settlement, and claimant receipt remain unresolved.

The 0.94 score helped identify what required review. It did not establish the covered amount, authorize the payment, or close the claim. Corrected evidence and Rina's decision established the covered amount and authorized the instruction; execution, settlement, notice, and administrative closure remained external.

Next day — external reports close only their own loops

Payment-system report PAY-EXEC-711 records execution. Bank report BANK-SETTLE-90 later records settlement. Communication record NOTICE-320 reports the reviewed notice delivered.

Those facts may satisfy the authored administrative close gate. They do not prove fair treatment, correct coverage, claimant understanding, repair completion, absence of harm, or a real-world result. Complaints, reconsideration, appeal, litigation, and regulator processes remain external.

Why this setting is recognizable

The NAIC Model Bulletin on the Use of Artificial Intelligence Systems by Insurers addresses governance, documentation, data quality, testing, transparency, and human involvement around AI-supported consumer-impacting insurance decisions. The NAIC Unfair Claims Settlement Practices Act, Model 900 provides a model framework for claim investigation and disposition while anticipating jurisdiction-specific adoption. The voluntary NIST AI Risk Management Framework 1.0 describes rights-preserving risk management across AI design, deployment, use, and evaluation.

Those materials ground only bounded claims and AI-governance context. The NAIC documents are model materials, not universal law, and NIST does not certify a model or insurer. None of the sources establishes this claim, validates ROOF-CV-8, decides coverage or fraud, authorizes payment, proves legal compliance, substantiates Grid, or shows an outcome.

Where the model stops

Grid can represent exact source and model revisions, scope a score, calculate deterministic consequences, expose missing or contradictory evidence, trace a correction, compare authored paths, assemble a review packet, and calculate payment after a qualified human supplies the covered amount. It cannot determine cause, coverage, fraud, liability, fair treatment, legal meaning, claim closure, settlement authority, payment, repair, or claimant outcome.

This page contains no real claimant, address, property, policy text, claim number, image, protected characteristic, payment account, insurer, jurisdiction, vendor secret, production threshold, fraud feature, credential, or live operating instruction. Every amount and rule is fixture-bound.

What remains to prove

Source revision 2026-08-27.1 has reviewed domain context, product boundaries, a high-consequence evidence profile, and a public-suitability finding, so this authored, non-advisory scenario can be published at R2. The fixture itself remains authored.

The next proof is an exact integer-cent implementation of the twelve-image and eight-image states, 0.94 and 0.71 scoped outputs, fourteen-minute explanation gap, $4,400 correction, $14,240 scope, $16,376 reserve candidate, three paths, data-integrity stop, $11,900 human covered amount, $2,500 deductible, $9,400 payable, $10,000 authority threshold, and separate payment, settlement, and notice reports. It needs an independent oracle, negative cases, retained execution and verification evidence, and a high-consequence assurance review. Intended-environment work would additionally require insurer-owned legal, policy, fairness, consumer-rights, model-risk, identity, payment, records, security, and complaint controls. Nothing in that path turns a confidence score into permission to close a claim.