All case studies
RESEARCH REFERENCE / LEGAL INTELLIGENCE
Customer name redacted

Document intelligence.
Better economics.

A high-volume document workflow shows how a specialist model can improve cited answers while reducing the cost per review-table cell.

An anonymized adaptation of public research. This is not a Finetuned customer engagement, endorsement or performance promise.
54.8%

Lower cost per cell

Compared with Claude Sonnet 5 on the same tasks.

0.903

Answer-quality score

A combined correctness and citation metric on a 0–1 scale.

22.4%

Citation-quality improvement

Relative improvement against the base GLM 5.2 checkpoint.

THE WORKLOAD

Each answer needs
the right evidence.

High-volume review tables multiply the impact of every quality and cost improvement.

The research used a synthetic dataset built from public legal documents; no customer data was used for training. The model was trained to produce correct, structured answers with supporting citations and appropriate abstention.

Specialization improved document answers in an existing semantic-retrieval workflow. A separate agentic-retrieval experiment studied token efficiency; its results should not be combined with the primary benchmark.

Reported study design
WorkflowLarge-scale document review tables
Training dataSynthetic examples from public legal data
Training targetCorrectness, schema and cited evidence
EvaluationPer-cell answer and citation quality
Operational focusAnswer quality and cost per cell
THE FINETUNED APPROACH

Use the pattern.
Validate the outcome.

We would scope the training data, select a licensed base model, establish held-out evaluations and agree on rollout gates with your experts. Your results would be measured independently.

Discuss a similar workflow
CONSIDERATIONS FOR YOUR PILOT
01

Define the answer contract

Agree on output schemas, evidence rules and abstention behavior.

02

Test on difficult documents

Hold out edge cases and measure unsupported answers.

03

Measure unit economics

Track cost per accepted answer alongside quality and latency.

SOURCE & COMPARISONS

Metrics are the original authors’ reported, workload-specific findings. The baselines differ across metrics; they are not a universal comparison of specialist and general models. Names are redacted in this presentation while original sources remain available for verification.

Read the original research
YOUR WORKFLOW IS NEXT.

Find your
specialist advantage.

Explore the economics for your company.

Calculate your value