StrategyDraftIn reviewApprovedHemnaathSaiBoth

Evidence Intelligence positioning, hospital-side beachhead, product layers, pilot scope, value targets, and non-goals.

MarkdownH2ListUse # headings, - lists, and plain text. Every save creates a revision.

# EvidenceGate Product Strategy v2.0 ## Positioning EvidenceGate is an Evidence Intelligence Platform for Health Insurance Operations in India. It prepares evidence, applies governed insurer rules, and produces audit-ready review packets. It does not approve, reject, diagnose, price, or settle a case. ## Beachhead The first product is a hospital-side quality gate for planned cashless pre-authorization. It checks a packet before it enters a TPA or insurer review queue. Initial users: - Hospital insurance-desk operators who assemble and submit pre-authorization packets. - TPA and insurer reviewers who need faster access to complete, traceable evidence. Initial buyer hypothesis: - Hospital revenue-cycle or insurance operations leadership for the first pilot. - TPA or insurer operations leadership for scaled deployment. ## Problem Pre-authorization packets arrive with missing documents, incomplete mandatory fields, conflicting patient or procedure details, unclear chronology, and evidence that is difficult to locate. Reviewers respond with queries. Hospitals resubmit. Every query cycle delays authorization and adds operating cost. The product does not promise an automated decision. It prevents avoidable packet defects and makes every material finding easy to verify. ## Product layers ### Evidence Engine Classifies documents, extracts typed facts, normalizes values, and creates an evidence graph. Every fact retains the immutable document version, page, region, snippet, extraction method, confidence, and hash. ### Policy Intelligence Publishes insurer-approved clauses and structured rules with effective dates, ownership, golden tests, activation history, and rollback. RAG may retrieve an approved clause. It cannot create policy or decide an outcome. ### Quality Gate Runs deterministic required-document checks, mandatory-field validation, chronology checks, cross-document comparisons, and published insurer rules. Unsupported or low-confidence inputs remain unknown and route to a human. ### Human Review Packet Shows findings, unknowns, evidence references, rule and policy versions, timestamps, and prior human actions. The human reviewer verifies, corrects, overrides, or requests more evidence. ## Pilot scope - One planned cashless pre-authorization workflow. - Two procedure families selected with a pilot hospital. - Five initial document types. - One policy and rule bundle. - One hundred permissioned historical or shadow-mode packets. - Read-only recommendations and quality findings. ## Pilot success targets Targets must be baselined with the pilot partner before they become commitments. - Reduce avoidable query cycles by at least 30 percent. - Reduce packet preparation and first-review handling time by at least 40 percent. - Achieve at least 95 percent evidence-reference coverage for material findings. - Achieve at least 98 percent precision on critical missing-document findings. - Keep false-pass rate at zero for deterministic mandatory checks. - Record a reason and evidence trail for every human override. ## Non-goals - No autonomous approval or rejection. - No medical diagnosis or treatment recommendation. - No claim adjudication, pricing, fraud disposition, or settlement. - No policy interpretation invented by a model. - No replacement of hospital, TPA, insurer, or NHCX systems of record. - No general claims platform in the first release. ## Expansion path After the hospital-side quality gate proves measurable value, EvidenceGate can expose the same evidence graph and governed rule bundles to TPA and insurer review teams. Reimbursement claims, underwriting support, and fraud investigation are later products, not current scope. ## Defensibility The moat is not OCR, RAG, or a collection of agents. It is the combination of evidence lineage, insurer-specific policy as code, deterministic quality gates, replayable runs, and the correction data created by human reviewers. ## Current gate No model implementation starts until Hemnaath and Sai have a signed data agreement, a representative packet sample, a frozen evidence contract, and a documented annotation protocol.