Grae SI

Sandbox profile

Matter-note briefs

Medium intensity

Associates brief their own matter notes on a server in the office. The model does not see other clients, and it does not draft a filing.

Vertical
Legal
Scale
Small company
Organization
Northwind LLP
Shape
Healthcare, pharma, and insurance boutique, 25 lawyers.

The business problem

Notes across a matter are hard to re-read before a call. Lawyers will not put privileged notes into a consumer chat product, and they should not. The firm wants a brief of the notes on that matter only, on hardware in the suite, with the lawyer responsible for what is said to a client.

What Grae SI recommends

Family

Qwen 2.5 Instruct

Size

7B on one office GPU. Notes are long. A 3B model drops parties and dates.

Why this size

The value is summarization under a privilege boundary, not legal research against the open web. A local 7B model with retrieval over one matter is the right size. It is not a brief-writing product and it is not connected to a research service.

License

The firm’s professional-responsibility partner reads the license. Privilege is a duty of the firm, not a term of the model license.

Data controls

Privileged material

Matter notes are confidential and often privileged. They stay on the office server. There is no vendor processor.

De-identification

Not the primary control. The primary control is matter isolation. A training fine-tune, if any, uses only notes the responsible partner has cleared, which may be none for the first release.

Access

The ethical wall is the document system’s. The model can retrieve a note only if that wall already allows the user. No “firmwide” index.

Residency

Office server. If a lawyer works remotely, the path is the firm’s VPN to that server, not a sync to a laptop model.

Retention

Briefs are not a second client file. If a brief is saved, it is saved into the matter. Unsaved sessions expire in 7 days.

Risk classification

Intensity is a planning tier for evidence. It is not a certification and not a regulatory classification.

  • Privileged client information is in scope, so this is not low.
  • The output is a brief for the lawyer on the matter. It is not advice sent to a client and not a filing.
  • A tool that drafted a filing or mixed two clients’ facts would be re-classified, and the second case would be a professional-conduct problem before it was a model problem.

Compliance requirements to prepare for

Professional conduct

Confidentiality and supervision. The firm’s duties do not move to software. A partner remains responsible for the matter.

HIPAA, when the firm is a business associate

Some healthcare clients require a BAA. A model host inside the firm, with no outside processor, is easier to explain than a chat vendor. It is not automatically compliant.

Privilege logistics

Ethical walls and matter access. The model must not be a way around them.

GDPR or DPDP

If a matter includes EU or India personal data, residency and purpose are counsel’s call. A single office server in one country may be the wrong design for that matter.

Grae SI helps assemble the evidence. It does not certify HIPAA, FDA, ISO, SOC 2, ICH, NAIC, GDPR, DPDP, or any other framework.

Validation KPIs

These are targets to lock before go-live. They are not measured outcomes from a customer deployment.

Matter isolation

0 cross-matter facts on a 25-probe set

The wall is the product.

Note fidelity

At least 90% of brief points trace to a note on 40 matters

A polished sentence that was not in the notes is a risk.

No filing voice

The brief is labeled internal and the probe set contains no caption-ready draft

Scope.

Wall bypass

A user outside the matter receives no text from it

Tested with real permission cases, not a demo account.

Deployment architecture

DeploymentOffice server, VPN only

The document store’s permissions are checked before a passage is retrieved. The model has no route except the firm’s VPN. There is no fine-tune in the first release unless a partner clears a corpus.

  1. 01

    Document management

    Ethical walls stay here.

  2. 02

    Matter retrieval

    Only documents the user may open.

  3. 03

    7B runtime

    One GPU in the office. Ollama or vLLM on the LAN.

  4. 04

    Lawyer

    Reads the brief before a call. Saving it is a filing into the matter, which is a human choice.

Each box is a component in the customer’s network. Arrows are the order of the path, read left to right on a wide screen and top to bottom on a phone.

Where this sits in the stack

The same ten layers appear on every profile. The mark says whether Grae SI provides the layer, shares it, or leaves it with you. Person-weeks are the buyer’s effort after the runtime is deployed, not a project plan for the whole company.

  • Grae SI providesGrae SI provides the layer
  • SharedShared. Grae SI supports part of it.
  • You ownYou own the layer after install

  1. Layer 1

    Infrastructure and compute

    You own

    A server in the firm’s office or a tenant with no training egress. Matter data does not go to a shared API.

    Buyer effort after the runtime is up: 1–3 person-weeks

  2. Layer 2

    Data ingestion and prep

    Shared

    A checklist of what to collect, a sample upload, and a readiness score. The pipeline from the system of record into an approved extract.

    Buyer effort after the runtime is up: 2–5 person-weeks

  3. Layer 3

    Data controls and governance

    Shared

    Client names can be masked in the sample. Ethical walls and matter access are the firm’s.

    Buyer effort after the runtime is up: 1–3 person-weeks

  4. Layer 4

    Model selection and training

    Shared

    Grae SI supplies the packaged 7B runtime. The first release does not fine-tune. Retrieval is the firm’s index, permission-filtered.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  5. Layer 5

    Evaluation and validation

    Shared

    Medium-tier suite, matter-isolation probes, and a partner who reads a sample.

    Buyer effort after the runtime is up: 1–3 person-weeks

  6. Layer 6

    Agents and orchestration

    Shared

    Compliance agents that classify the use case and name the frameworks to prepare for. Any production agent that calls tools inside your applications.

    Buyer effort after the runtime is up: 1–2 person-weeks

  7. Layer 7

    Deployment and serving

    Grae SI provides

    Safetensors, GGUF, a model card, and an offline bundle for Ollama, vLLM, or llama.cpp. Moving that bundle into the zone and keeping the host patched.

    Buyer effort after the runtime is up: 0.5–1.5 person-weeks

  8. Layer 8

    Monitoring and drift

    Shared

    Registry lineage: version, dataset, eval result, and the package that was released. Live traffic checks, drift alarms, and the decision to pull a version.

    Buyer effort after the runtime is up: 1–3 person-weeks

  9. Layer 9

    Application and integration

    You own

    A panel in the document system, or a simple internal page if the vendor cannot embed one.

    Buyer effort after the runtime is up: 2–4 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    A partner decides what the model may see. That judgment is not delegated to the software.

    Buyer effort after the runtime is up: 1–4 person-weeks

Buyer effort for this profile

4–7 person-weeks

A 25-lawyer firm should not run a research project. Isolation and a written rule of use are the deliverable.

  • A partner for the rules of use
  • The firm’s IT admin
  • One associate to build the gold questions

Quantitative benefits

Illustrative estimate for this sandbox. It is not a measured customer result. Assumptions for this profile are listed under the figures.

Time to a reviewable package

3–5 weeks

No fine-tune. The document system already has walls.

Cost versus a cloud API

About $1k–$3k a month avoided, and a privilege problem avoided

Illustrative at this headcount. The API is the wrong design even if it were cheaper.

Compute

$80–$200 a month in power and support

One office GPU.

Prep before a call

15–30 minutes on a thick matter

Illustrative. Thin matters will not show the gain.

Accuracy target

90% of points trace to a note, zero cross-matter leaks

Release gates.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • No client-facing output and no court filing from this model.
  • Ethical walls in the document system are correct. The model does not repair a bad wall.
  • Healthcare, pharma, and insurance clients make BAAs and privilege both relevant. The profile does not assume every matter is a BAA matter.

All sandbox profiles · The full stack