#114
Monthly Rank
#262
Lifetime Rank
Share
10 September 2026
V

Verify Every Fact. Keep Every Judgment

AI verifies every factual allegation against the evidence but the associate decides the legal response.

Soubhagyashree DasWebsite

Gallery

About the Project

Answering a pleading in Indian civil litigation begins with a mechanical, unskippable task: testing every factual allegation in a Plaint or Written Statement against the client's documentary evidence — 30–50 factual paragraphs against 100–200 pages, by hand, allegation by allegation. It takes 3–5+ hours of an associate's time, and most of it is searching, not judgment.

This tool inserts a verification layer between receiving a pleading and answering it. It takes an already-processed pleading and a searchable evidence set, extracts and decomposes each factual proposition, searches the whole bundle, and classifies every proposition into one of six categories :- Supported, Contradicted, Partially Supported, Not Established, Conflicting, or Ambiguous and each finding traceable to a document and page.

Its defining constraint is what it refuses to do. It never sets, suggests, or implies an admit/deny response. It never drafts Written Statement prose or gives legal advice. The Decision column stays empty until the associate records one the system prepares and presents; the associate decides. Flags and evidence gaps must be explicitly acknowledged before a review can be marked complete, so no fact is silently dropped and automation bias is structurally prevented, not merely discouraged.

Built as a proof of concept under Order VIII Rules 3–5 CPC 1908, it targets active review time under 30 minutes for a representative matter and without giving up complete coverage, source traceability, or human control of every legal decision.

Practice Areas

Key Features

  • Extracts and decomposes factual propositions from a pleading, splitting compound allegations into discrete, individually-testable ones
  • Classifies each proposition into six mutually exclusive categories against the client's evidence set
  • Every evidence-based finding cites a document and page that resolves nothing is asserted without a source
  • Surfaces contradictions, gaps, and conflicts explicitly, including where the client's own records disagree
  • Enforced human-in-the-loop: flagged items and evidence gaps cannot reach "complete" until the associate acknowledges them
  • Response-control boundary: the system never sets, suggests, or implies an admit/deny decision and the Decision field is the associate's alone
  • Overrides retained alongside the original AI classification and never replaced, never reversed
  • Matter-consistency check prevents running a pleading against the wrong evidence set

Help Needed

Looking for feedback from litigation practitioners on the six-category classification scheme and the response-control boundary and specifically whether the enforced-acknowledgment design holds up against real automation-bias risk in practice. Also open to collaborators on the parts deliberately left outside this PoC: document ingestion/OCR, privilege awareness, and production security.

About the Creator

SD
Soubhagyashree Das