Route diligence evidence to the right workstream before summarizing

A management interview mentions a distributor's customer concentration, a sole-source supplier and a proposed ERP migration in the same ten minutes. Sending the whole transcript through every diligence agent produces repeated reading and competing summaries. The useful unit of work is the evidence segment and the question it informs.

What this means for your business

For a consulting firm, a routing layer could turn a growing data room into more targeted analysis. Analysts spend less time sorting repetitive outputs, and specialist AI workflows operate on smaller, more relevant evidence bundles. The benefit depends on preserving overlap: one fact can legitimately belong to several workstreams.

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: industrial-distributor diligence

Consider a fictional acquisition of a regional industrial distributor. An interview excerpt says the largest customer buys custom kits assembled from a component supplied by a single manufacturer, whose agreement renews in November. The data room contains sales extracts, supplier contracts and warehouse operating notes.

A local transcript parser retains timestamps and speaker attribution. CLM answers separate routing questions for commercial, procurement and operations relevance, with a review option for each. It should not force the excerpt into one exclusive category. The known engagement ID and authorised workstream membership come from the project system.

A procurement route retrieves the named manufacturer's contract and extracts renewal and termination language for an analyst. A commercial route retrieves the relevant customer's sales history for a concentration calculation performed by code. An operations route is invoked only if the segment contains assembly or fulfilment evidence. The synthesis model sees the accepted findings and their sources after those steps.

This removes the habit of asking every specialist to read every transcript section. Reusable descriptions of the approved workstreams can be encoded in advance. Actual savings depend on how selective routing is and how often multiple workstreams are needed. An excerpt routed to two teams is a correct outcome when both need it.

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.

Host inside the engagement's controlled environment

An on-site deployment is a strong fit for engagements where confidential data-room materials should remain under the client or firm's direct control. Place the service on an approved network beside the document connector, and keep transcription, retrieval and synthesis within the same boundary when required. A client-hosted workstation can serve the engagement without granting an external model provider access to its source documents.

Use engagement-specific permissions and caches. Separate reusable generic workstream definitions from confidential client embeddings. Preserve document version and interview timestamp in every proposed finding, and include cached data in the engagement's retention and deletion procedure. The commercial case must include setup and support across short projects, not just inference cost.

Test whether analysts get to evidence sooner

Use a previously completed engagement with analyst-approved workstream labels. Include contradictory interview statements, multiple versions of the same contract and excerpts that belong to several teams. Compare specialist invocations, source-checking time and missed material evidence. Treat the router as successful only if the resulting workstream packs are at least as complete as the baseline.

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.