Calibrated Trust, Not Sharper Prediction: An Empirical Test of Uncertainty Fusion
Fusing uncertainty tools (Bayesian odds, Dempster-Shafer, conformal prediction) into legal AI pipelines does not improve discriminative accuracy over raw frontier LLMs; AUROC remains ~0.83 regardless. Naive composition of LLMs with Bayesian-odds and Dempster-Shafer fusion more than doubles calibration error (ECE from ~0.16 to ~0.46) due to a prior-mismatch mechanism. Dempster-Shafer fusion is actively unsafe on long reasoning chains, producing confident but incorrect predictions at below-chance
Analysis
TL;DR
- Fusing uncertainty tools (Bayesian odds, Dempster-Shafer, conformal prediction) into legal AI pipelines does not improve discriminative accuracy over raw frontier LLMs; AUROC remains ~0.83 regardless.
- Naive composition of LLMs with Bayesian-odds and Dempster-Shafer fusion more than doubles calibration error (ECE from ~0.16 to ~0.46) due to a prior-mismatch mechanism.
- Dempster-Shafer fusion is actively unsafe on long reasoning chains, producing confident but incorrect predictions at below-chance accuracy; the authors recommend removing it entirely.
- The pipeline's real value is operational risk control: a conformal selective-prediction layer enables automated clearance at 96.8% accuracy with only 0.5% errors escaping, versus 85.9% for an untuned baseline.
- The core thesis: uncertainty fusion in legal AI delivers calibrated trust, not sharper prediction.
Why It Matters
This study directly challenges a popular trend in legal AI of stacking multiple uncertainty quantification methods to boost performance. For practitioners building AI-assisted legal systems, it demonstrates that raw frontier LLMs already capture most discriminative signal, and that naive fusion can actively degrade calibration. The findings are especially relevant for high-stakes domains where overconfident errors carry serious consequences.
Technical Details
- Dataset and task: 1,000 real European Court of Human Rights cases from LexGLUE and FairLex, predicting whether the Court found a Convention violation from fact paragraphs.
- Models evaluated: Claude Opus 4.8 and GPT-5.5 as per-fact evidence estimators, compared across three families: (A) raw LLM, (B) LLM routed through the fusion pipeline, (C) term-frequency baseline through the same pipeline.
- Fusion components tested: Evidence graphs with belief propagation, sequential Bayesian odds updating, Dempster-Shafer combination, and conformal prediction.
- Key metrics: AUROC for discrimination (~0.83 across all conditions), Expected Calibration Error (ECE) showing a jump from ~0.16 to ~0.46 with naive fusion, and class-conditional risk control for selective prediction.
- Final tuned engine: After removing Dempster-Shafer, recalibrating, and applying conformal selective prediction, the system auto-clears at 96.8% accuracy with 0.5% error escape rate and 96.3% caught for human review, versus 85.9% / 3.8% / 72.1% for the untuned baseline.
Industry Insight
- Legal AI systems should prioritize calibration and selective deployment over chasing marginal accuracy gains through complex uncertainty fusion; a well-tuned conformal layer on a raw LLM outperforms heavily composed pipelines.
- Dempster-Shafer theory, while theoretically appealing for evidence combination, introduces dangerous overconfidence in long-chain legal reasoning and should be avoided in production systems without rigorous safeguards.
- The "calibrated trust" paradigm—using uncertainty tools to decide when to automate versus escalate—is the more pragmatic and defensible application of uncertainty quantification in regulated domains like law.
Disclaimer: The above content is generated by AI and is for reference only.