SallyIP

Home / Raw LLM vs verification-first legal AI: an ablation study

Raw LLM vs verification-first legal AI: an ablation study

What happens when fluent fabrications meet substance scoring alone versus the full verification pipeline? On fixed fixtures: 1/4 and 0/22 versus 4/4 and 26/26 — with zero false positives on good answers.

Method

Hand-written synthetic fabrications (invented cases, fake statutes, altered quotations) plus 2 good answers, checked against a fixed 5-passage statute fixture. Layers: A substance (phrase checks), B retrieval attachment, C citation-integrity guard, D quote verification, E full pipeline. Two sets: original 8-item (4 fabrications) and expanded 25 (22 new fabrications). Zero live model calls — a guard-behavior simulation, not a model benchmark.

Results

SetSubstance alone (A)Full pipeline (E)Good answers harmed
Original (4 fab, 2 good)1/44/40/2
Expanded (22 fab)0/2222/22— (combined 0/2)

Why substance scoring fails

The fabrications were designed fluent: each contains its expected phrase, so phrase checks pass by construction. Only evidence-starved guards (dangling labels), word-for-word quote checks, and entailment grading catch them — three independent signals, which is why the pipeline separates 26 flagged from 2 supported with nothing in between.

Limitations

Fixed passages stand in for retrieval; ranking, gating and answer modes are not exercised. Real retrieval adds noise the simulation lacks. The fabrications are hand-written, so the study measures guard mechanics against known trap types (13 covered), not discovery of novel failure modes. Reproduce: node scripts/ablation-bench.mjs, node scripts/ablation-25.mjs.

Sources: benchmarks/ablation-report.md, benchmarks/ablation-25-report.md (both 2026-09-10).

More external validation