TECHNICAL PROTOCOL · REPRODUCIBLE RESEARCH STANDARDS REVISED: OCTOBER 2026 VERSION: 2026.10-PUBLIC

VERDICT

Methodology & Verification Protocol — How We Reconcile Public Records

Institutional Research Standards

Research Methodology & Forensic Verification

A comprehensive guide to how VERDICT transforms fragmented public data into auditable, reproducible, and legally defensible intelligence records.

1. Start with an Auditable Question

Investigative integrity begins with question design. Inquiries that embed unproven conclusions corrupt data gathering before collection starts. VERDICT enforces a strict distinction between falsifiable factual queries and speculative accusations.

Methodological Test:
✓ Auditable: "Which publicly documented organizations has Entity X worked for between 2020 and 2026?"
✕ Biased / Speculative: "How is Entity X secretly controlled by political sponsors?"

The second question embeds a conclusion of clandestine control before any evidence is reviewed. VERDICT rejects speculative framing; our research engine only answers questions whose elements can be corroborated against primary records.

2. The 9-Stage Verification Ladder

All published records must traverse an immutable 9-stage verification ladder. No automated agent can bypass this ladder, and raw findings never reach the public interface without editorial review.

  1. Objective Decomposition: The research target is broken into distinct investigative facets (Identity, Directorships, Affiliations, Public Statements, Legal Filings, Timelines).
  2. Independent Source Sweeps: Specialized adapters query gazettes, corporate registries (MCA), public court dockets, archived web pages, and media reports.
  3. Atomic Normalization: Raw source material is parsed into discrete, single-assertion observations with publication timestamps and author attribution.
  4. Identity Resolution: Multi-token Jaccard similarity and context scoring verify whether the subject of an article is the target or an unrelated namesake.
  5. Corroboration Cross-Check: Independent secondary and primary sources are matched to determine whether a claim is corroborated.
  6. Contradiction Review: Lexical and negation polarity analysis identifies conflicting accounts between credible publishers.
  7. Human-in-the-Loop Review: Senior researchers inspect source provenance, context, and legal defensibility.
  8. Publication Gating: Private data, credentials, and unverified allegations are purged; public-safe projections are generated.
  9. Immutable Audit Logging: Every state change, reviewer decision, and timestamp is sealed in an append-only audit trail.

3. Source Reliability Classification (Tiers A–E)

VERDICT assesses reliability per source and per claim. We reject universal "reliability scores" because a reputable newspaper can err on an unverified quote, and an official social media post is excellent evidence of what someone claimed while being insufficient proof that the claim is true.

Tier A — Primary / Official Records: Government gazettes, certified corporate filings (MCA/RoC), court orders, statutory declarations, and direct organizational charters.
Tier B — Established Reporting: Attributable reporting by professional news organizations with transparent bylines and editorial corrections policies.
Tier C — Secondary / Analytical: Aggregators, academic working papers, republished commentaries, and research briefs citing external material.
Tier D — Self-Published / Public Statements: Verified social accounts, personal campaign blogs, and press releases. Valid for establishing that a subject made a statement, but never treated as self-evident proof of truth.
Tier E — Unverified / Unattributed: Anonymous forum leaks, unauthenticated screenshots, and unsourced social chatter. These are held in quarantine and cannot be published.

4. Identity Resolution Architecture

Name collisions are the single largest source of false investigative accusations in India. With millions of citizens sharing common surnames, identical names do not imply identical people.

Our engine compares candidate targets across six distinct dimensions: exact normalized name, token set overlap, corporate directorships (CIN/DIN), shared public domains, geographical jurisdiction, and temporal activity windows.

CONFIRMED

Score ≥ 0.78 & Zero Contradictions

Multiple primary signals match (e.g. corporate DIN, official charter name, and verified public domain). High-impact claims permitted.

PROBABLE

Score ≥ 0.52 & Zero Contradictions

Strong token overlap and organizational context, but lacking statutory ID confirmation. Displayed with explicit caveats.

UNRESOLVED

Score < 0.52

Common name match with insufficient contextual overlap. Entity remains unmerged; claims cannot be cross-attributed.

CONFLICTED

Material Signals Contradict

Sources point to two different people with identical names in conflicting professions or cities. Prohibits automatic merging.

5. Epistemic Claim States

VERDICT does not reduce journalism to a binary "True / False" verdict. Real evidence exists in varying states of documentation and dispute. Every published claim is explicitly tagged with one of six epistemic states:

  • DOCUMENTED: Directly supported by a primary official record, statutory gazette, or authenticated charter.
  • REPORTED: Credibly reported by established news organizations with on-record bylines, but uncorroborated by primary statutory filings.
  • ALLEGED: A person or organization has formally asserted a claim; factual verification remains incomplete.
  • DISPUTED: Two or more credible sources materially contradict each other; both records are preserved side-by-side.
  • REFUTED: Direct primary evidence or authoritative judicial findings positively disprove the claim.
  • UNRESOLVED / GAP: Research has investigated the question but identified no positive corroboration; recorded as an explicit evidence frontier.

6. Contradiction Detection Engine

When two credible news outlets or public sources disagree, bad algorithms silently discard the earlier report or pick whichever outlet aligns with an editorial bias.

VERDICT’s contradiction engine surfaces conflicts automatically: when two claims share entity subjects and lexical overlap but exhibit opposite negation polarity, the system creates a ContradictionFinding with requiresReview=true. The conflicting records are presented transparently to the reader rather than being swept away.

7. Public Money-Flow Boundaries

Financial investigations frequently suffer from over-interpretation. VERDICT enforces three critical separation rules:

  1. Announcement ≠ Receipt: A public announcement (e.g. a ₹1 crore legal defense pledge) is recorded strictly as a commitment until banking or statutory returns substantiate receipt.
  2. Entity ≠ Individual: Organizational donations cannot be conflated with personal compensation to an organizer or spokesperson.
  3. Transaction Path ≠ Wallet Ownership: On-chain public transactions demonstrate cryptographic asset flows, but graph topology alone never establishes legal ownership without off-chain attribution.

8. Social Interaction Edges vs. Personal Relationships

Public digital activity (replies, reposts, mentions, quotes) demonstrates visible public interaction. It does not establish private coordination, organizational command, or secret affiliation.

Network Rule: Two public figures who interact with the same third party share network overlap, not a proven personal relationship. The system strictly prohibits automated inferences of conspiracy or control based on graph proximity alone.

9. Bounded Negative Evidence: What "Not Found" Means

One of the most dangerous fallacies in public research is treating the absence of evidence as proof of non-existence.

When our research fleet finds no records for a claim (for example, foreign remittances or undeclared corporate holdings), VERDICT never declares "No such funding exists." Instead, the system states: "Within the bounded corpus of Ministry of External Affairs statements, MCA records, and searched archives, no corroborating record was identified."

10. The Micro-Agent Research Fleet

VERDICT is orchestrated by 13 specialized research roles. Each agent possesses a narrow capability and operates under strict epistemic constraints:

1. Identity & Resolution: Normalizes names, corporate DINs, and handles across public registries.
2. Web Discovery: Discovers long-tail articles, institutional publications, and press releases.
3. Wayback & Archives: Recovers historical versions of altered websites and deleted organizational charters.
4. Official Records: Searches gazettes, election affidavits, and parliamentary proceedings.
5. Corporate Registry: Maps Ministry of Corporate Affairs filings, active charges, and directorships.
6. Legal & Court Records: Inspects public court orders, cause lists, and legal aid disclosures.
7. Social Cross-Check: Verifies public statements on X and Reddit without inferring private ties.
8. Timeline Reconstruction: Assembles chronological sequences with documented ISO timestamps.
9. Relationship Mapping: Structures direct public affiliations and institutional ties.
10. Public Money Flow: Traces bounded transaction movement without asserting unverified ownership.
11. Contradiction Engine: Detects opposite polarity assertions between credible publishers.
12. Evidence Review: Evaluates multi-source corroboration against publication thresholds.
13. Publication Gate: Enforces privacy boundaries, redacts internal vault tokens, and blocks unqualified claims.

11. The Publication Gate & Privacy Boundary

VERDICT does not turn public accountability research into a license to violate personal privacy. Our automated publication projection enforces strict redacting filters:

  • Zero private communications, hacked data, leaked password vaults, or phone numbers are published.
  • Internal research task weights, API keys, and model parameters are stripped prior to public projection.
  • Only observations with confirmed source provenance and human reviewer clearance can be rendered.

12. Corrections & Immutable Audit Trail

Every published record retains a complete historical revision trail. If a primary record proves an existing finding inaccurate:

1. The original claim is marked with state REFUTED or SUPERSEDED.
2. The correcting source is permanently attached with attribution.
3. A public correction entry is added to the research ledger detailing the modification date, reason, and reviewing researcher.

To submit a factual correction, visit our Public Contribution Desk.