How accurately Kavach reads real policies

We run Kavach's document reader on real, public Indian health insurance wordings and compare every term it reads with labels a person checked against the PDF.

Accuracy per field

Each bar is the share of wordings where Kavach read that term exactly as labelled.

• Room rent limit per dayno labels yet
• ICU limit per dayno labels yet
• Co-pay percentno labels yet
• Age the co-pay startsno labels yet
• Proportionate deduction appliesno labels yet
• Deductibleno labels yet
• Initial waiting periodno labels yet
• Pre-existing disease waiting periodno labels yet
• Specific illness waiting periodno labels yet
• Non-payable items listno labels yet

The wordings

Public documents filed with IRDAI or published by the insurer, identified by their UIN.

• Arogya Sanjeevani Policy

  The New India Assurance Co. Ltd.

  UIN NIAHLIP25044V022425

  IRDAI standard productReadLabel not checked yet

• Senior Citizens Red Carpet Health Insurance Policy

  Star Health and Allied Insurance Co. Ltd.

  UIN SHAHLIP19101V031819

  Senior citizenReadLabel not checked yet

• my:health Medisure Super Top Up Insurance

  HDFC ERGO General Insurance Co. Ltd.

  UIN HDFHLIP22021V042122

  Super top-upReadLabel not checked yet

• Super Health Insurance

  SBI General Insurance Co. Ltd.

  UIN SBIHLIP23050V012223

  IndividualReadLabel not checked yet

• Complete Healthcare Insurance

  Universal Sompo General Insurance Co. Ltd.

  UIN UNIHLIP23006V032223

  Family floaterReadLabel not checked yet

• Group Health Insurance

  ICICI Lombard General Insurance Co. Ltd.

  UIN ICIHLGP22207V032122

  Group (employer)ReadLabel not checked yet

• Group Health Policy

  United India Insurance Co. Ltd.

  UIN UIIHLGP21226V022021

  Group (employer)ReadLabel not checked yet

• Group Health Insurance (Revised)

  Future Generali India Insurance Co. Ltd.

  UIN FGIHLGP21165V022021

  Group (employer)ReadLabel not checked yet

The rules behind every rupee

Every amount Kavach shows comes from these rules, not from the AI. A rule is marked verified only after a person checks it against the primary source.

17 of 31 rules verified

• R-HL-001To verify

  Proportionate deduction only on room-priced heads

• R-HL-002To verify

  ICU is never proportionately deducted

• R-HL-003Verified

  Co-pay on the admissible amount

• R-HL-005To verify

  Non-payable items excluded

• R-HL-006To verify

  Choice of policy order

• R-HL-008Verified

  Room rent capped at the limit

• R-HL-009Verified

  Super top-up pays only above its deductible

• R-HL-010Verified

  Billing type unknown: both branches and one question

• R-HL-014Verified

  Package rate: no ratio, ask for the lower-category price

• R-CLM-001To verify

  Deduction when the room was within the limit

• R-CLM-002To verify

  Deduction on excluded heads

• R-CLM-004To verify

  Consumables deducted

• R-CLM-005Verified

  Co-pay base and rate

• R-CLM-006To verify

  Vague deduction reason

• R-CLM-008To verify

  Deduction on ICU or without differential billing

• R-MT-001To verify

  Maternity capped at the defined limit

• R-MT-002To verify

  Maternity waiting period

• R-MT-003To verify

  Newborn covered inside the maternity limit

• R-CLM-009To verify

  Deduction for the maternity limit

• R-HN-010Verified

  Nuclear family health need

• R-HN-011Verified

  Parents' health need

• R-HN-012Verified

  Employer cover is contingent

• R-HN-013Verified

  Super top-up needs a base

• R-LF-001Verified

  Term life need

• R-LF-002Verified

  Savings plan is not term cover

• R-PA-001Verified

  Personal accident need

• R-DV-001Verified

  Device protection

• R-BG-001Verified

  Monthly protection budget

• R-OP-001Verified

  Most protection per rupee, most urgent first

• R-SC-001Verified

  Safety score

• R-MG-001To verify

  Migrate group cover before the last working day

How we measure

• Dataset: 8 public policy wordings from 8 insurers, including group, senior citizen, super top-up and the IRDAI standard product. The PDFs are not stored in our repository; a script fetches them by URL and checks their fingerprint.
• Labels: for each wording an AI pre-fills about ten terms with the page and a quote; a person then checks every one against the PDF. Only checked labels count.
• Scoring: exact match per field. Limits written as a percentage of the sum insured are compared at a sum insured of ₹5,00,000.
• Spoken amounts: rehearsal recordings are transcribed and every rupee amount spoken is compared with the engine's answer, offline.
• The eval code, labels and results are in the evals folder on GitHub.

What this does not show yet

Eight wordings is a small set; we say so rather than round it up. Citation accuracy and the verifier catch rate are not measured on this page yet.