Route construction-dispute evidence before the legal AI starts writing

A subcontractor emails a revised drawing, a daily site report and a disputed change-order request in the same message. A legal AI that treats the whole bundle as one summarization task creates work the team must untangle later. The more useful first question is which review each piece of evidence belongs in.

What this means for your business

For a construction-litigation practice, this could reduce repeated document processing and get evidence into the correct chronology sooner. The commercial benefit is less associate time repairing intake and less model capacity spent on irrelevant analysis. The firm's lawyers still decide relevance, privilege and legal significance.

CLM-v0.1-8B scores supplied choices instead of writing an answer. Its two projection heads use a frozen Qwen3-8B encoder. That makes it a candidate for deciding which workflow should run next. It does not extract a complete record, execute a tool or prove that a document is correct. The model card and implementation are linked below.

A specific workflow: late steel-detail revisions

Consider a fictional hospital-extension dispute. Email subject: “RFI-184 / Level 3 connection revision / CO-027.” The attachment text mentions revised steel connection details, a crane crew standing down on 14 May, and a request for additional installation hours. A second attachment is a site diary with the day's labour allocation.

First, local OCR and parsers produce page-linked text. Matter ID and access permissions come from the document system. CLM then scores a small, firm-defined set of review paths for each document section: design-change chronology, delay-event review, quantum-support review, routine project correspondence, or manual review. Use separate questions where a section may support both delay and cost; a single exclusive label would discard useful overlap.

For the drawing-related email, the proposed design-change route retrieves RFI-184 and the drawing revision register. For the site diary, the delay route retrieves the approved programme and other records for 14 May. A specialist drafting model receives only the relevant bundle and produces a proposed chronology entry with page references. The cost-support path can separately reconcile labour records; CLM itself does none of that arithmetic.

Without routing, a broad agent might run all three specialist prompts on every section. With routing, straightforward sections invoke the relevant specialist and ambiguous sections go to review. That removes unnecessary generation and retrieval calls; it does not make the underlying legal analysis intrinsically faster. Privilege screening remains an independent control and must not be skipped because the router chose a routine category.

A compact on-site deployment

The lightweight part is the scoring layer. The full service still needs the Qwen3-8B encoder. As a sizing calculation, eight billion weights at two bytes each occupy roughly 16 GB before runtime memory, activations and serving overhead. A single Linux workstation with a supported 24–32 GB GPU is a reasonable starting configuration to evaluate for short inputs and modest concurrency, not a tested minimum or a throughput promise.

The reference implementation pairs a vLLM pooling service with the CLM API. Keep decision inputs short, precompute reusable candidate embeddings, and batch arrivals within an acceptable waiting window. The repository's example caps inputs at 2,048 tokens; split longer documents into meaningful sections and retain page references instead of silently truncating evidence. One encoder can serve several department-specific sets of choices.

Four-bit weights would be roughly 4 GB in raw weight storage by arithmetic, but that is not the total memory requirement. Quantization changes embeddings and may change rankings. Validate the exact encoder, pooling method, precision and heads together before buying smaller hardware. The supplied heads cannot simply be attached to any small embedding model. A tiny CPU appliance or a Mac mini is not a verified drop-in deployment for this reference stack.

Why this belongs inside the firm's data boundary

For a practice with confidential matter repositories and an IT team able to operate the service, on-site hosting is a strong fit: document text, embeddings, proposed routes and audit records can stay beside the matter system. Apply the same matter access rules before retrieval and when opening cached results. Keep local OCR and the downstream drafting model inside that boundary too; a local classifier alone does not make a cloud drafting workflow private.

Store the model version, candidate definitions, source page IDs and reviewer correction for each routed item. Give the service read-only access to evidence and a separate permission to create a proposed work item. It should not alter originals or send correspondence. Budget patching, encrypted backups and a manual intake queue for outages.

Test the handoff that matters

Replay a lawyer-labelled sample containing overlapping delay and quantum evidence, missing attachments and poor scans. Compare routing errors, reviewer corrections, specialist calls per section and time to an accepted chronology entry against the existing process. Weight a missed material document more heavily than an unnecessary review. Start in shadow mode on one matter type and retain the source evidence at every step.

The release is an early candidate-ranking model. Published agent benchmarks do not establish accuracy or speed on this workflow. Scores are relative to the alternatives supplied, not calibrated probabilities that a business decision is correct. Include an explicit manual-review choice, then set abstention rules using held-out examples; adding that choice alone does not make the system safe.