ProofPath

The educational problem

Students are often asked to judge whether a conclusion follows from evidence, but the feedback they receive is usually only whether an answer is right or wrong. ProofPath turns that hidden reasoning step into something visible: the learner supplies a claim and evidence, sees whether the relationship is SUPPORTED, CONTRADICTED, or INSUFFICIENT, and can inspect the model's class probabilities.

The learning goal is narrow and concrete: practice separating what the evidence says from what someone wants to conclude from it. ProofPath is not a truth detector; it is an educational tool for evidence reasoning.

What it does

ProofPath is an educational AI/ML prototype that compares a claim with a piece of evidence and classifies their relationship into three categories:

SUPPORTED — the evidence supports the claim
CONTRADICTED — the evidence conflicts with the claim
INSUFFICIENT — the evidence is not strong enough to justify the claim

The goal is not to act as a universal truth detector or automatic fact checker. ProofPath is designed specifically as a learning tool for evidence assessment.

How we built it

ProofPath runs entirely in the browser using JavaScript, HTML, and CSS.

At startup, the application trains a compact three-class softmax/logistic regression model from bundled labeled claim-and-evidence examples. The model uses weighted features derived from each claim/evidence pair to estimate probabilities for the three possible outcomes.

No external AI API, paid model, account, cloud inference service, CDN, or model download is required. The project can therefore demonstrate the full ML reasoning pipeline locally and transparently.

The interface was designed to make the model's reasoning easier to inspect by showing the predicted class together with its class probabilities.

Why AI/ML is core

The verdict is not hard-coded per example. The model learns from labeled claim/evidence pairs and combines text-pair features such as lexical overlap, coverage, negation mismatch, numeric conflict, antonym conflict, contradiction cues, uncertainty cues, overclaim cues, and evidence/claim length relationships. The ML component therefore performs the central educational classification task rather than acting as a decorative add-on.

Challenges

The biggest challenge was keeping the project both technically meaningful and honest about what the model can prove.

A small educational model can perform well on carefully designed examples without being a general-purpose fact checker. We therefore built explicit claim limits into the project and avoided presenting the model as something it is not.

Another challenge was creating examples that clearly distinguish support, contradiction, and insufficient evidence while remaining understandable to learners.

Accomplishments

The completed prototype is fully functional in the browser and requires no external inference infrastructure.

A fixed bundled holdout set contains 10 examples that were excluded from training. The current build correctly classifies all 10 of those examples.

This 10/10 result is a bounded demonstration on a small curated holdout set, not evidence of general-world accuracy.

What we learned

Building ProofPath reinforced an important lesson: useful AI is not only about producing answers. It can also help people understand the limits of evidence and recognize when a conclusion goes beyond what the available information supports.

What's next

Future work could expand the training and evaluation corpus, test the model with more diverse evidence relationships, improve explanations of individual predictions, and study how tools like ProofPath affect students' evidence-reasoning skills.