COLLABORATIVE AGENTIC REASONING • VERIFICATION ENGINE

FactLens Multi-Agent Verification Prototype

An autonomous fact-checking pipeline built on LangChain and LangGraph. Decomposes multimodal claims into atomic propositions, executes parallel tri-channel evidence retrieval across institutional archives, runs deterministic credibility scoring, and arbitrates verdicts with an adversarial skeptic loop.

Orchestrator: LangGraph StateGraph Parallelism: LangChain RunnableParallel LLM Inference: Gemini 2.5 Flash + Grounding Backend: FastAPI ASGI (Async Pipeline) Memory & Cache: SQLite Drift Tracking

INPUT INTAKE

Run Claim Verification

Select an empirical benchmark claim or enter a custom claim to execute the 12-agent LangGraph workflow:

PRESET BENCHMARK CLAIMS:

CLAIM TEXT (NORMALIZED INTAKE):

RETRIEVAL POLICY:

CACHE MODE:

STATUS: IDLE · READY TO RUN 00:00.0

STAGE 02 · PARALLEL TRIO

Archive Hunter Agent

Institutional primary source scraper

LANGGRAPH LOGIC & RULES:

Executes targeted site-scoped queries to high-trust institutional publishers (.gov, .edu, WHO, IMF, World Bank, Nature). Filters media commentary.

INPUT SCHEMA: normalized_claim, domain, query_plan

OUTPUT SCHEMA: EvidenceItem[] (Archive Tiers T1-T2)

APIs & PROVIDERS:

Vertex Grounding Google CSE API World Bank API

LANGGRAPH STATEGRAPH ARCHITECTURE

CLICK ANY AGENT TO INSPECT CODE & PROMPTS

PHASE 01 INTAKE & ROUTING

Intake Agent

Sanitizes & Normalizes

Domain Router

Semantic Route Classifier

➔

PHASE 02 PARALLEL RETRIEVAL TRIO

Web Scout Agent

Multi-Provider Cascade

Archive Hunter

Gov/Edu Institutional Scrape

Data Extractor

World Bank & Deep Scrape

➔

PHASE 03 CONSISTENCY & FUSION

Tri-Consistency

Jaccard Coherence Metric

Evidence Aggregator

Deduplication & MMR Re-rank

➔

PHASE 04 ADVERSARIAL AUDIT

Skeptic Agent

Contradiction Gap Finder

Evidence Auditor

7-Tier Credibility Scoring

Stat Comparator

Numeric & Rank Guardrail

➔

PHASE 05 ARBITRATION & REBUTTAL

Consensus Moderator

Final Explainable Verdict

Rebuttal & Memory Tail

Contradiction Feedback Loop

↺ CYCLIC REBUTTAL CONDITIONAL EDGE: If Skeptic flags contradiction & confidence < gate ➔ re-fetch archive evidence and re-audit

CONTRADICTED CONFIDENCE: 94.6%

JACCARD CONSISTENCY: 0.84 AUDIT SCORE (T): 92.4/100 EXECUTION: 2.78s

EXPLAINABLE ARBITRATION RATIONALE:

Claim asserts India is the 4th largest global economy by nominal GDP in 2026. While official IMF and World Bank datasets project India reaching the 4th rank in nominal terms between late 2026 and 2027 ($4.18T), Japan and Germany currently occupy ranks 4 and 3 respectively as of latest verified quarterly statistics. Counter-evidence from IMF World Economic Outlook (October 2025 release) confirms rank 5 status with active transition velocity.

AUDITED EVIDENCE CORROBORATION (7-TIER QUALITY AUDITED):

3 Sources Corroborated

TIER	DOMAIN	RETRIEVAL CHANNEL	SNIPPET / ASSERTION	WEIGHT
T1 (1.00)	imf.org	Archive Hunter (Official)	IMF World Economic Outlook: India nominal GDP projected at $4.187T, maintaining Rank 5 behind Japan ($4.31T).	0.96
T1 (1.00)	worldbank.org	Data Extractor API	World Development Indicators: NY.GDP.MKTP.CD series confirms $3.94T baseline for previous fiscal year.	0.94
T3 (0.85)	reuters.com	Web Scout Cascade	Analysis: India on trajectory to overtake Japan by fiscal 2027; currently verified at fifth spot.	0.81

ENGINEERING SPECIFICATION

Deterministic Quality Scoring & LangGraph Primitives

FactLens eliminates hallucinations by decoupling reasoning into an auditable directed graph with deterministic mathematical scoring.

Deterministic Evidence Auditor Formula

Evidence quality is never left to subjective LLM opinion. Every extracted item passes through the mathematical scoring equation:

T = ( Σ(w_i × R_i) / n ) × (1 - ρ) × 100

w_i: Domain credibility weight from the 7-tier hierarchy (T1 Government: 1.00 to T7 Social Media: 0.05).
R_i: Semantic relevance token overlap score (0 to 1).
ρ: Duplicate penalty ratio discounting single-domain concentration.
Temporal Decay: Multiplier penalizing stale data (1.0 for current year down to 0.4 for 10+ years).

Why LangGraph vs. Linear Chains?

Standard LangChain sequential chains fail when verification requires cyclic feedback. LangGraph provides:

StateGraph Typed State: VerificationState carries normalized propositions, evidence pools, and skeptic notes across nodes.
RunnableParallel Fan-Out: Web Scout, Archive Hunter, and Data Extractor execute simultaneously across threads, reducing latency from 7.2s to 2.8s.
Conditional Cyclic Edge: The Rebuttal Loop re-triggers evidence retrieval if the Skeptic Agent detects contradictions and the confidence gate is breached.
Deterministic Memory Tail: Persists run states and evaluates historical drift in local SQLite.

7-Tier Source Credibility Taxonomy

T1 • 1.00 Government & Official Records (.gov, .edu, WHO, NASA, RBI)

T2 • 0.95 International Bodies & Scientific Journals (UN, IMF, World Bank, Nature, Lancet)

T3 • 0.85 Global Wire Services & Major Financials (Reuters, AP News, Bloomberg, BBC, WSJ)

T4 • 0.75 Reputable National Broadsheets (The Hindu, NDTV, Economic Times)

T5 • 0.65 Dedicated Fact-Checking Organizations (Snopes, PolitiFact, BoomLive)

T6 • 0.15 Clickbait, Tabloids & Satire Outlets

T7 • 0.05 Social Media & Messaging (X, Facebook, TikTok, Reddit - Excluded)

REPO: /f/fact-lens-lablab

# factlens_crew/orchestrator.py
from langgraph.graph import StateGraph, END
from langchain_core.runnables import RunnableParallel
from factlens_crew.schemas import VerificationState

def build_factlens_graph():
    workflow = StateGraph(VerificationState)

    # Register the 12 Stage Nodes
    workflow.add_node("intake", intake_agent)
    workflow.add_node("router", domain_router_agent)
    workflow.add_node("retrieval_trio", RunnableParallel({
        "web": web_scout_agent,
        "archive": archive_hunter_agent,
        "extractor": data_extractor_agent
    }))
    workflow.add_node("tri_consistency", tri_consistency_agent)
    workflow.add_node("aggregator", evidence_aggregator_agent)
    workflow.add_node("skeptic", skeptic_agent)
    workflow.add_node("auditor", evidence_auditor_agent)
    workflow.add_node("stat_comparator", stat_comparator_agent)
    workflow.add_node("moderator", consensus_moderator_agent)
    workflow.add_node("rebuttal_loop", rebuttal_agent)
    workflow.add_node("memory_tail", memory_tail_node)

    # Edge Sequence
    workflow.set_entry_point("intake")
    workflow.add_edge("intake", "router")
    workflow.add_edge("router", "retrieval_trio")
    workflow.add_edge("retrieval_trio", "tri_consistency")
    workflow.add_edge("tri_consistency", "aggregator")
    workflow.add_edge("aggregator", "skeptic")
    workflow.add_edge("skeptic", "auditor")
    workflow.add_edge("auditor", "stat_comparator")
    workflow.add_edge("stat_comparator", "moderator")

    # Conditional Rebuttal Edge: If Skeptic flags contradiction & confidence < 80%
    workflow.add_conditional_edges(
        "moderator",
        should_rebut,
        {
            "rebuttal": "rebuttal_loop",
            "finalize": "memory_tail"
        }
    )
    workflow.add_edge("rebuttal_loop", "auditor")
    workflow.add_edge("memory_tail", END)

    return workflow.compile()