Apppublic

bitsabhi/phi-coherence

sourceHugging Facemitupdated 7mo agoView on Hugging Face

1likes

App README

φ-Coherence v3 — Credibility Scoring

Detect fabrication patterns in ANY text — human or AI. 88% accuracy. No knowledge base. Pure math.

The Insight

Truth and fabrication have different structural fingerprints. You don't need to know the facts to detect the fingerprints.

LLMs generate text that sounds like truth. Humans inflate resumes, pad essays, write fake reviews. Both exhibit the same patterns:

—Vague attribution ("Studies show...")
—Overclaiming ("Every scientist agrees")
—Absolutist language ("Exactly 25,000", "Always", "Never")

This tool detects the structural signature of fabrication — regardless of whether a human or AI wrote it.

Use Cases

Domain	What It Catches
AI Output Screening	LLM hallucinations before they reach users
Fake Review Detection	"This product completely changed my life. Everyone agrees it's the best."
Resume/Essay Inflation	Vague claims, overclaiming, padding
Marketing Copy	Unsubstantiated superlatives
News/Article Verification	Fabricated quotes, fake consensus claims
RAG Quality Filtering	Rank retrieved content by credibility

What It Detects

Pattern	Fabrication Example	Truth Example
Vague Attribution	"Studies show..."	"According to the 2012 WHO report..."
Overclaiming	"Every scientist agrees"	"The leading theory suggests..."
Absolutist Language	"Exactly 25,000 km"	"Approximately 21,196 km"
Stasis Claims	"Has never been questioned"	"Continues to be refined"
Excessive Negation	"Requires NO sunlight"	"Uses sunlight as energy"
Topic Drift	"Saturn... wedding rings... aliens"	Stays on subject

Why It Works

LLMs are next-token predictors. They generate sequences with high probability — "sounds right." But "sounds right" ≠ "is right."

Your tool detects when "sounds like truth" and "structured like truth" diverge.

The LLM is good at mimicking content. This tool checks the structural signature.

Benchmark

Version	Test	Accuracy
v1	Single sentences	40%
v2	Paragraphs (12 pairs)	75%
v3	Paragraphs (25 pairs)	88%
Random	Coin flip	50%

API

python

from gradio_client import Client
client = Client("bitsabhi/phi-coherence")
result = client.predict(text="Your text here...", api_name="/analyze_text")

Limitations

—Cannot distinguish swapped numbers ("299,792" vs "150,000") without knowledge
—Well-crafted lies with proper hedging will score high
—Best on paragraphs (2+ sentences), not single claims

Built by [Space (Abhishek Srivastava)](https://github.com/0x-auth/bazinga-indeed)

"Truth and fabrication have different structural fingerprints."