Audit-Grade Multi-Model Fact-Checking API

Fact-check AI output before your users see it.

Built for AI products that cannot afford to publish hallucinations. Eight frontier models across a 5-stage adversarial pipeline with immutable primary source citations.

Verify Any Claim or Statement
8-Model Pipeline
Quick examples:
63%
Frontier LLM Disagreement
8 Models
In Adversarial Pipeline
5 Stages
With Immutable Citations
1,000+ /day
API Extractions
Depth Ladder

Four Primitives. Choose Your Latency & Depth.

From synchronous 5-second pre-send filters to 90-second deep adversarial audit investigations.

POST /extractFREE · 1K/DAY

Atomic Claim Extraction

Pulls testable factual claims out of unstructured human or AI text. Resolves pronouns and strips rhetorical filler.

Latency: ~500ms · Deterministic
POST /assessFAST GATE

3-Model Frontier Panel

Fast multi-model verdict for synchronous user experiences and runtime pre-send checks before delivering drafts.

Latency: ~5–10s · Synchronous
POST /verifyDEEP AUDIT

8-Model Adversarial Pipeline

Full 5-stage pipeline: live research, 2-round Pro vs. Con debate, 3-expert panel review, and 20+ primary citations.

Latency: ~60–90s · Webhooks
POST /ask/{id}GROUNDED Q&A

Verification Follow-Ups

Conversational exploration strictly grounded in the verified evidence corpus. Ask about edge cases and jurisdictions.

Latency: ~1–2s · Streaming
8 Models · 5 Stages · Real Sources

The Multi-Model Adversarial Pipeline

Click each stage below to explore how raw claims turn into audit-grade verdicts.

PIPELINE ENGINE ACTIVE
Stage 03: 2-Round Adversarial Debate

Adversarial Debate

Two AI advocates argue opposing sides across two rounds. One builds the affirmative case, the other constructs the counterargument, followed by cross-rebuttals.

Proponent constructs strongest affirmative case
Opponent writes targeted counter-argument
Mutual cross-rebuttals expose weak logic
Execution TopologySTAGE 3/5
Proponent (Model A)Constructs affirmative argument from primary evidence.
Opponent (Model B)Constructs counter-argument and tests edge cases.
Original Acuityio Research

Why One LLM Is Never Enough

Standard benchmarks test multiple-choice trivia. When we tested 5 frontier models on 1,000 real-world claims, they failed to reach consensus on 63% of claims.

Each model self-reported 9/10 or 10/10 confidence over 75% of the time, even when completely disagreeing with peer models.

Open Evaluation Benchmark · 1,000 Real Claims

Frontier LLM Agreement & Confidence Variance

Models evaluated on 1,000 user-submitted claims. 63% resulted in split verdicts.

63%
Disagreement Rate
GPT-4.5 / 4oOpenAI
Agreement: 84.2%Avg Conf: 9.2/10
Claude 3.5 SonnetAnthropic
Agreement: 86.8%Avg Conf: 8.9/10
Gemini 1.5 ProGoogle
Agreement: 82.5%Avg Conf: 9.1/10
Llama 3.1 405BMeta
Agreement: 79.4%Avg Conf: 8.7/10
Mistral Large 2Mistral AI
Agreement: 78.1%Avg Conf: 8.5/10
Rated True
Rated Mixed
Rated False
1,000 Claims · Acuityio Empirical Benchmark v1.0
Developer-First Architecture

Integrate in Minutes with Typed SDKs

Available on PyPI (pip install acuityio) and npm (npm install acuityio). Built-in idempotency keys, automatic retries, and HMAC webhook verification.

Remote MCP Server for Claude, Cursor, and ChatGPT
Official n8n & Zapier workflow automation nodes
CI/CD regression runner for golden evaluation sets
Acuityio SDK Integration
from acuityio import Acuityio

client = Acuityio(api_key="acuity_live_...")

# 1. Extract atomic claims from unstructured text
claims = client.extract("Einstein won the 1921 Nobel Prize for general relativity.")

# 2. Run full 8-model adversarial verification
result = client.verify_and_wait(claims[0])

print(result.verdict)      # "False"
print(result.acuity_score) # 1 (out of 10)
print(result.summary)      # "Won for the photoelectric effect, not relativity."

Stop AI hallucinations before they reach production.

Start verifying claims with high-throughput multi-model panels and comprehensive citation trails today.