Kebiki · Use case
Every denial carries the clause that produced it.
Coverage, limits and deductibles evaluated against the policy wording in force on the date of loss. Reason codes that cite the clause rather than paraphrase it, and a trace that replays three years later against the same version.
Assess claim
Business OperationPOST /api/public/v1/packages/claims_adjudication/evaluate{
"case": {
"claim": {
"type": "AUTO_COLLISION",
"amount": 4200,
"incident_date": "2026-04-18",
"at_fault": false
},
"policy": { "id": "P-88431", "as_of": "2026-04-18" }
}
}
ok{
"status": "ok",
"outputs": {
"decision": "APPROVED",
"reason_codes": ["COVERAGE_ACTIVE", "DEDUCTIBLE_APPLIED"],
"citations": ["MOTOR-2026.1 §3.2", "MOTOR-2026.1 §7.1"],
"facts": { "deductible": 500, "payable": 3700 }
},
"package_version": "MOTOR-2026.1",
"trace_id": "tr_9f2c41"
}
The problem
The wording and the code stop agreeing.
Nobody sets out to adjudicate inconsistently. It happens because the policy lives in a document, the logic lives in an application, and only one of them gets maintained.
Two adjusters, two answers
The same facts reach different outcomes depending on who picked up the file and how they read a clause written for a different product.
Denials nobody can explain
The letter gives a reason in prose. The claimant appeals. Reconstructing which clause and which version actually applied takes a person a day, if it is possible at all.
Examinations become projects
A market conduct exam samples files from a period you did not choose, and the answer has to be assembled from commit history and memory.
What Kebiki does
Compile the wording. Call it. Keep the receipt.
The policy document stays the system of record. Kebiki compiles it into a decision package your claims system calls, and the package is a disposable artifact that can be rebuilt from the document at any time.
| What you get | Why it matters at the examination |
|---|---|
| Coverage, limit and deductible checks | One evaluation, called by every channel, so the first notice of loss and the adjuster screen cannot disagree. |
| Reason codes that cite the clause | The ERISA claims procedure rule asks for reference to the specific provision relied on. A citation is cheaper to produce at decision time than to reconstruct at appeal. |
| Version pinning | A 2024 loss is decided by the 2024 wording, for as long as anyone asks. Changing the policy does not rewrite the past. |
| A replayable trace | The inputs, the version and the path through the rules, re-runnable to the same answer rather than described from memory. |
| No model at runtime | AI compiles the wording; nothing infers at decision time. There is no sampled output to govern, document or defend. |
The rest of the answer
A decision is not the whole file.
Kebiki settles which rule applied. Two things sit either side of it, and a claim that has to survive an appeal needs all three.
Bring one policy and one denial.
Send us a real wording and a determination you have had to defend. We will compile the wording, run the claim, and show you the citation and the trace it produces.