ARES_POC is a mathematically grounded misinformation verification system.
Unlike LLM-based fact checkers, ARES computes truth using a deterministic evidence functional:
This repository is the Proof-of-Concept implementation accompanying the research manuscript.
Given a claim C, ARES:
- Retrieves evidence passages (R)
- Selects Top-M semantically relevant passages (S)
- Evaluates entailment / contradiction (N)
- Applies source credibility priors (W)
- Computes a reproducible truth score
pip install -r requirements.txtARES also includes a research-grade visualization interface for inspecting the mathematical verification process.
The verdict panel shows the final decision along with the computed Truth Score and Confidence derived from the functional.
Every evidence passage contributing to the verdict is shown with:
- Semantic Similarity (S(C,E))
- Entailment Score (N(C,E))
- Source Weight (W(E))
- Individual Contribution to Truth'
This ensures full transparency of the decision process.
py main.py "The Earth orbits the Sun" --k 10 --m 5py evaluate.py --dataset liar --samples 20| File | Operator | Role |
|---|---|---|
| retriever.py | R(C) | Evidence retrieval |
| similarity.py | S(C,E) | Top-M ranking |
| entailment.py | N(C,E) | Logical validation |
| credibility.py | W(E) | Source prior |
| verifier.py | V(C) | Truth functional |
ARES uses a local cache/ to ensure that the same claim with same parameters always produces the same truth score.
ARES does not generate answers.
It derives verdicts strictly from discovered evidence.
MIT License


