⚖️ Contract Clause Classifier — NLP Risk Flagging Live

Every vocabulary token is tagged with one of four categories: indemnity (indemnify, hold harmless, defend, indemnitee), termination (terminate, convenience, without cause, perpetual), liability (unlimited, uncapped, consequential damages, waive, disclaim) or general (breach, penalty, forfeit, arbitration). For each clause the simulation sums the coefficients of matched tokens separately within each category, and the category with the highest positive subscore becomes that clause's "primary risk driver" tag shown in the highlighted panel. A clause can score high overall while still having a clearly dominant category, which mirrors how real contract-review tools bucket flagged clauses for a lawyer's queue.

Because the model is a linear sum of per-token coefficients, the exact contribution of any single token to the final score is just that token's coefficient (or coefficient × count, for repeated tokens) — there is no approximation or post-hoc explanation step needed, unlike attribution methods such as SHAP or integrated gradients that are required for non-linear models. The heatmap renders each token's background opacity and hue directly proportional to its signed coefficient: warm colours for positive (risk-raising) weights, cool colours for negative (risk-reducing) weights, and neutral grey for unweighted tokens. What you see is the literal arithmetic term in the sigmoid's input sum, not a simulated visual effect.

The classifier always outputs a continuous risk probability between 0 and 1 for every clause; the threshold slider sets the cutoff above which a clause is labelled "flagged" rather than "clear". Lowering the threshold flags more clauses (higher recall, more false positives on borderline clauses) while raising it flags fewer, more confidently risky clauses (higher precision, but some real risk may slip through unflagged). This is the same precision/recall trade-off every binary classifier faces in production, and legal-tech teams typically tune this threshold against a labelled sample of previously reviewed contracts rather than picking it arbitrarily.

The running chart plots the per-clause risk score in the order clauses are fed through the classifier, plus a cumulative running-average line. Because the synthetic contract deliberately interleaves low-risk boilerplate (definitions, notices, governing law) with high-risk clauses (uncapped indemnification, unilateral termination for convenience), the per-clause line oscillates sharply while the running-average line smooths those swings out and converges toward the document's overall risk profile as more clauses accumulate — exactly the pattern a real batch contract-review job produces when scanning clause by clause.

A hand-weighted bag-of-words model like this one is not something a firm would deploy as-is — the coefficients here were chosen for pedagogical clarity, not fitted to a labelled dataset of real contracts, and the vocabulary is far too small to catch the enormous variety of phrasing real contracts use. In practice this kind of linear model is used as a fast first-pass triage filter, trained on thousands of clauses labelled by paralegals, sitting in front of a slower, more accurate transformer-based re-ranker that a human reviewer checks before anything is relied upon legally. The core value of the simple linear version is that every flag is fully explainable — exactly what this simulation visualises.