Benchmark
Benchmark proof
Live benchmark view for insurance extraction schemas. Thirty-four schemas now have scored diagnostic real-public-document results on named corpora; future lines publish with the same evidence and release-gate discipline.
Diagnostic Snapshot
Accuracy result by completed schema.
Commercial GL, Commercial Auto, Workers Compensation, Commercial Property, Umbrella / Excess, Professional Lines, Cyber Liability, Directors and Officers, Employment Practices Liability, Crime, Inland Marine, Cargo, Environmental / Pollution, Surety / Bonds, Builders Risk, Personal Auto, Homeowners, Renters, Condo, Personal Umbrella, Life, Health, Disability, Travel, Pet, COI, Policy Declaration, Endorsement, Schedule, Binder, Quote, Renewal, Cancellation, Application, and First Notice of Loss now show scored diagnostic results from named real-public-document corpora. Public 99% claims still require larger corpora and the release gate.
Commercial GL benchmark
Limits, entities, additional insured posture, waiver, primary/non-contributory, and endorsement effects.
Scored diagnostic snapshotCommercial GLState: diagnosticShip bar: passed
Current scored diagnostic result for Commercial GL: 100% on 4 real public documents. This is a small proof set, not an insurance-wide accuracy claim, and it is not labeled published until the full release gate is cleared.
Public release gate: pass rate >= 99%, required fields >= 99%, evidence >= 99%, deferred docs = 0.
Wake County Commercial General Liability Declarations (Commercial General Liability view)
Source URLPortella Bellisimo Unit 2 Homeowners Association General Liability Policy (Commercial General Liability view)
Source URLGreenhills Homeowners Association Commercial General Liability Policy (Commercial General Liability view)
Source URLSean Scarmack Commercial General Liability Policy (Commercial General Liability view)
Source URLBenchmark Rules
What appears here.
Publication
Diagnostic before published
A scored diagnostic result can appear before a schema is marked published. Published accuracy still requires the full release gate.
Schema-specific proof
Each line publishes independently. One schema passing does not imply another schema has cleared the same bar.
Scoring
99% release gate
A schema does not publish as passed until document pass rate, required fields, and evidence checks clear the 99% bar.
Evidence matters
A value without evidence is not treated like a defended extraction. Evidence quality stays part of the release bar.
Boundaries
No blended insurance number
Accuracy is tracked by schema, document profile, and required-field applicability.
Real insurance documents
The current hardened proof sets use real public documents. The broader 99% claim requires larger gold-labeled corpora by schema family.
Core Lines
The first thirty-five proof-gated hardening tracks.
Commercial GL, Commercial Auto, Workers Compensation, Commercial Property, Umbrella / Excess, Professional Lines, Cyber Liability, Directors and Officers, Employment Practices Liability, Crime, Inland Marine, Cargo, Environmental / Pollution, Surety / Bonds, Builders Risk, Personal Auto, Homeowners, Renters, Condo, Personal Umbrella, Life, Health, Disability, Travel, Pet, COI, Policy Declaration, Endorsement, Schedule, Binder, Quote, Renewal, Cancellation, Application, and First Notice of Loss are hardened named-corpus schemas. The remaining tracks are active hardening targets and cannot publish accuracy until their own evidence gates clear.
Commercial GL
Limits, entities, additional insured posture, waiver, primary/non-contributory, and endorsement effects.
Commercial Auto
CSL, symbols, hired/non-owned signals, deductibles, and fleet-facing coverage structure.
Workers Comp
Statutory states, employers liability limits, mod factors, and class-code related signals.
Umbrella / Excess
Occurrence, aggregate, retention, follow-form posture, and underlying schedule references.
Commercial Property
Building, BPP, business income, deductibles, valuation basis, and causes-of-loss signals.
Professional Lines
Professional liability, cyber, D&O, pollution, claims-made posture, retro dates, and sublimits.
Cyber
Cyber aggregate limits, each-claim/event limits, retentions, retroactive dates, and breach-response signals.
D&O
Aggregate limits, Side A posture, retentions, continuity dates, and claims-made evidence.
EPLI
Each-claim and aggregate limits, retentions, continuity dates, and claims-made evidence for EPLI schedules.
Crime
Employee theft, computer fraud, forgery, funds transfer fraud, deductibles, and crime/fidelity schedules.
Inland Marine
Contractor's equipment, equipment floaters, transit limits, installation risk, scheduled property, and deductibles.
Cargo
Ocean cargo, marine cargo, stockthroughput, shipment limits, transit deductibles, policy numbers, and territory signals.
Environmental
Contractors pollution liability, pollution legal liability, premises pollution liability, per-pollution-condition limits, aggregate limits, retentions, and retroactive-date language.
Surety
Bond numbers, penal sums, obligees, principals, surety names, effective dates, and performance/payment/reclamation bond evidence.
Builders Risk
Project limits, soft-cost limits, transit and temporary-storage limits, deductibles, valuation basis, and completed-value/all-risk form evidence.
Personal Auto
Policy numbers, policy periods, bodily injury and property damage limits, UM/UIM limits, and vehicle schedules from declarations and public court packets.
Homeowners
Coverage A dwelling, other structures, personal property, loss of use, personal liability, medical payments, deductibles, and policy periods from homeowners declarations.
Renters
Personal property, loss of use, personal liability, medical payments, deductibles, and policy periods from renters declarations.
Condo
HO-6 / condominium unit-owner Coverage A/C, loss assessment, liability, medical payments, deductibles, and policy period evidence.
Personal Umbrella
Umbrella liability limits, policy periods, retained limits, and underlying insurance requirements from declarations and public court records.
Life
Policy numbers, issue dates, face amounts, insured names, and beneficiary evidence from life policies and public court records.
Health
SBC plan names, covered-member tiers, overall deductibles, and out-of-pocket maximums from public health plan documents.
Disability
Long-term disability monthly benefit amounts, elimination periods, and maximum benefit-period evidence from public certificates and summaries.
Travel
Travel insurance trip-cost, medical-limit, cancellation-limit, and coverage-effective trigger evidence from public travel protection documents.
Pet
Pet insurance annual policy limits, reimbursement percentages, deductibles, and insured-pet names from public policy packets.
COI
Certificate producer, insured, certificate-holder, and coverage-row extraction from real public ACORD certificate packets.
Declaration
Declaration-page policy numbers, named insureds, policy periods, and limit evidence from real public policy evidence documents.
Endorsement
Endorsement form numbers, edition dates, effective dates, and changed-term titles from real public policy packages.
Schedule
Schedule extraction for policy numbers, schedule types, scheduled rows, and form-or-limit schedule evidence from real public policy packages.
Binder
Binder extraction for binder numbers, effective and expiration dates, and bound coverages from real public insurance binders.
Quote
Quote extraction for quote numbers, quoted premiums, proposed limits, and subjectivities from real public quote schedules.
Renewal
Renewal extraction for renewal terms, expiring policy numbers, renewal premiums, and change-condition evidence from real public renewal invites.
Cancellation
Cancellation extraction for policy numbers, effective cancellation dates, and stated cancellation reasons from real public cancellation notices.
Application
Application extraction for applicants, operations descriptions, exposure statements, and requested coverages from real public application records.
FNOL
First Notice of Loss extraction for loss dates, claimant names, loss locations, and claim descriptions from real public claim notice records.
Methodology
Honest, bounded benchmark claims.
No single blended number
PolDex does not publish a single blended "insurance accuracy" number. Accuracy is tracked by schema, by document profile, and by required-field applicability.
Public corpus target
The current Commercial GL diagnostic set is the first Worker-reachable proof set. The broader corpus already has 11 gold-labeled real public documents, and each core schema still targets 50 real public or approved documents before it can carry a broader published accuracy claim.
Proof before integration
Use Live Proof to inspect a real extraction, then use Docs to inspect the underlying API.
Validate the product directly.
Run live proof, inspect schema contracts, then initialize access when you are ready.