Jev screens Jev
Pilot-fit screen. Five gates, twelve typed checks, one request. YES means recommend this scoped shadow pilot on this evidence, not approve production use.
A demo. It screens one company against five gates, and asks Jev twelve typed questions about the supplied evidence.
- Company
- Gallows Freight (fictional) (gallows-freight.invalid)
- Note
- Fictional company. Demo data only.
- Data
- Company data
- Workflow
- Freight damage claim payout decision (wf-claim-payout)
- Input unit
- claim narrative
- Example
- Driver reports a pallet shifted in transit and three cartons were crushed. Shipper is claiming the full invoice value of the pallet plus the cost of the return leg. The bill of lading notes the pallet was accepted without exception at pickup.
Labels
pay_full- Pay the claimed amount in full → Issue payment from the claims accountpay_partial- Pay the depreciated or apportioned amount → Issue reduced payment from the claims accountdeny- Deny the claim → Send the denial letter
Evidence (5)
The claims team handles about 95 damage claims a month at roughly 18 minutes each.
95 claims per month, mean 18 minutes per decision, which is 28.5 hours.
The decision issues payment or sends a denial letter directly, so the model output would move money and deny service.
Each label triggers a payment from the claims account or a denial letter to the shipper. There is no shadow mode in the current design.
Reference labels are whatever the handler paid at the time, with no adjudication pass.
Labels are the historical payout amounts as recorded by the handler. No second reader, no adjudication, no written rubric at the time of handling.
Legal has not authorised sending claim narratives to a hosted API, and the narratives contain shipper and driver names.
Processing authorisation for a hosted API has not been granted. Claim narratives carry shipper contact details and driver names, and no redaction step exists.
The development and held-out sets were drawn from the same export without removing duplicates, so they overlap.
Both splits were sampled from the same claims export with replacement. Overlap has not been removed and is known to exist.
View JSON
{
"company": {
"name": "Gallows Freight (fictional)",
"domain": "gallows-freight.invalid",
"fictional": true
},
"workflow": {
"id": "wf-claim-payout",
"name": "Freight damage claim payout decision",
"input_unit": "claim narrative",
"example_text": "Driver reports a pallet shifted in transit and three cartons were crushed. Shipper is claiming the full invoice value of the pallet plus the cost of the return leg. The bill of lading notes the pallet was accepted without exception at pickup.",
"labels": [
{
"id": "pay_full",
"meaning": "Pay the claimed amount in full",
"downstream_action": "Issue payment from the claims account"
},
{
"id": "pay_partial",
"meaning": "Pay the depreciated or apportioned amount",
"downstream_action": "Issue reduced payment from the claims account"
},
{
"id": "deny",
"meaning": "Deny the claim",
"downstream_action": "Send the denial letter"
}
]
},
"baseline": {
"minutes_per_decision": 18,
"decisions_per_month": 95
},
"data": {
"mode": "company_data",
"development_count": 120,
"heldout_count": 240,
"development_label_counts": {
"pay_full": 40,
"pay_partial": 40,
"deny": 40
},
"heldout_label_counts": {
"pay_full": 80,
"pay_partial": 80,
"deny": 80
},
"disjoint_splits": false,
"processing_authorized": false,
"redaction_policy": "None documented.",
"label_provenance": "Historical payout amounts, recorded by whoever handled the claim at the time."
},
"risk": {
"per_label": [
{
"label_id": "pay_full",
"false_positive_consequence": "The claims account pays a claim that policy would have denied or reduced.",
"false_negative_consequence": "A valid claim is underpaid and the shipper disputes it.",
"max_direct_loss_usd": 4200,
"recovery": "Recovery requires clawing back an issued payment, which the claims team describes as rarely successful."
},
{
"label_id": "pay_partial",
"false_positive_consequence": "A full claim is underpaid, triggering a dispute and a second review.",
"false_negative_consequence": "A reducible claim is paid in full.",
"max_direct_loss_usd": 4200,
"recovery": "Reissue the difference, after the shipper disputes."
},
{
"label_id": "deny",
"false_positive_consequence": "A valid claim is denied and a denial letter goes to the shipper.",
"false_negative_consequence": "An invalid claim is paid.",
"max_direct_loss_usd": 4200,
"recovery": "The shipper must appeal. The denial letter has already been sent."
}
],
"excluded_unreviewed_consequence": true,
"production_side_effects": true
},
"escalation": {
"reviewer_role": "Claims supervisor",
"queue": "Claims escalation inbox",
"reviewer_confirmed": false,
"review_deadline_business_days": 2,
"uncertainty_policy": "Unresolved cases default to the existing payout recommendation.",
"fallback": null
},
"facts": [
{
"id": "f-volume",
"text": "The claims team handles about 95 damage claims a month at roughly 18 minutes each.",
"kind": "operator_attested",
"source_url": null,
"source_title": "Fixture narrative, operator-written",
"excerpt": "95 claims per month, mean 18 minutes per decision, which is 28.5 hours.",
"as_of": "2026-10-01"
},
{
"id": "f-payout-effect",
"text": "The decision issues payment or sends a denial letter directly, so the model output would move money and deny service.",
"kind": "operator_attested",
"source_url": null,
"source_title": "Fixture narrative, operator-written",
"excerpt": "Each label triggers a payment from the claims account or a denial letter to the shipper. There is no shadow mode in the current design.",
"as_of": "2026-10-01"
},
{
"id": "f-labels-historical",
"text": "Reference labels are whatever the handler paid at the time, with no adjudication pass.",
"kind": "operator_attested",
"source_url": null,
"source_title": "Fixture narrative, operator-written",
"excerpt": "Labels are the historical payout amounts as recorded by the handler. No second reader, no adjudication, no written rubric at the time of handling.",
"as_of": "2026-10-01"
},
{
"id": "f-no-rights",
"text": "Legal has not authorised sending claim narratives to a hosted API, and the narratives contain shipper and driver names.",
"kind": "operator_attested",
"source_url": null,
"source_title": "Fixture narrative, operator-written",
"excerpt": "Processing authorisation for a hosted API has not been granted. Claim narratives carry shipper contact details and driver names, and no redaction step exists.",
"as_of": "2026-10-01"
},
{
"id": "f-splits-overlap",
"text": "The development and held-out sets were drawn from the same export without removing duplicates, so they overlap.",
"kind": "operator_attested",
"source_url": null,
"source_title": "Fixture narrative, operator-written",
"excerpt": "Both splits were sampled from the same claims export with replacement. Overlap has not been removed and is known to exist.",
"as_of": "2026-10-01"
}
],
"check_fact_ids": {
"closed_decision": [],
"text_judgment": [],
"action_mapping": [],
"baseline_observed": [
"f-volume"
],
"comparison_valid": [],
"context_sufficient": [
"f-splits-overlap"
],
"labels_valid": [
"f-labels-historical"
],
"processing_ready": [
"f-no-rights"
],
"consequences_bounded": [
"f-payout-effect"
],
"recovery_valid": [
"f-payout-effect"
],
"reviewer_ready": [],
"routing_complete": []
},
"prediction_scope": {
"entity_name": "Gallows Freight (fictional)",
"workflow_description": "freight damage claim payout decisions"
}
}