Grae SI

Sandbox profile

Contract clause comparisons

Medium intensity

Associates compare a draft BAA, clinical agreement, or policy form against the firm’s playbook. A partner still negotiates.

Vertical
Legal
Scale
Enterprise
Organization
Marlowe & Grey
Shape
Panel counsel for health systems, pharma, and carriers. Several hundred lawyers, three offices.

The business problem

The firm reviews the same clauses in BAAs, CTAs, and insurance forms. The playbook lives in a set of memos. Associates re-derive it. Putting a client’s draft into a public model is not compatible with the firm’s commitments. They want a comparison: playbook position, draft position, difference, open question. Not a markup that sends itself to opposing counsel.

What Grae SI recommends

Family

Qwen 2.5 Instruct

Size

7B with long context for most drafts. A 14B step-up only if one office already has a GPU that can hold it. Do not start there.

Why this size

Clause comparison is a structured difference against a playbook the firm controls. Retrieval of the playbook plus a 7B model is the cost-effective default. A 70B model does not remove the partner. If context windows fail on long CTAs, split by clause before you buy a bigger model.

License

Recorded per office, because a later productized tool sold to clients would raise license questions a pure internal tool does not.

Data controls

Client confidential

Drafts are client confidential. They are processed in the firm’s tenant and are not used to train a model shared with another client.

De-identification

The playbook is not client data. A fine-tune, if used, trains on playbook pairs the knowledge partner has cleared, not on live client drafts.

Access

Matter team only. The playbook itself may be firmwide. The draft is not.

Residency

The office region that holds the matter. A draft for an EU affiliate is not processed in a US office node.

Retention

Comparisons saved to the matter follow the matter schedule. The playbook version is pinned on the comparison so a later playbook change does not rewrite history.

Risk classification

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

  • Client confidential documents, internal use, partner still negotiates.
  • Not low, because a missed clause can change a deal, and because healthcare and insurance forms carry regulated terms.
  • Sending the comparison to the other side without a lawyer’s review is outside the intended use.

Compliance requirements to prepare for

Professional conduct

Supervision and confidentiality. The comparison is work product of the firm.

HIPAA BAAs

When the document is a BAA, the reviewer still checks the required terms. The model’s “missing clause” flag is a prompt for that review, not the review.

Insurance forms

A policy-form comparison may later be used by a carrier client. The firm is not the filer. If the carrier wants to adopt the model, that is the carrier profile, not this one.

GDPR

EU personal data inside a draft (employee names, patient examples) needs an EU node or a redaction step. The profile assumes redaction of example personal data before indexing.

ISO/IEC 42001

Useful once the tool is firmwide. Intended use and the ban on unreviewed outbound markup belong in that record.

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.

Playbook coverage

At least 90% of playbook clauses that appear in the draft are mentioned, on 50 agreements

A silent miss is the failure associates will not notice.

False difference

Under 10% of reported differences are not differences, on the same set

Noise trains lawyers to ignore the tool.

Client isolation

A draft from matter A is not retrievable in matter B

Same wall rule as the boutique, at higher volume.

Outbound ban

The tool has no send action. Design review, plus a probe that the export is a matter memo rather than a redline email.

Workflow control.

Deployment architecture

DeploymentFirm tenant, per-region nodes

Playbook index is firmwide. Drafts are retrieved under matter permission and processed in the matter’s region. The comparison is a memo, not a message to opposing counsel.

  1. 01

    Playbook

    Versioned memos. Knowledge partner owns them.

  2. 02

    Matter draft

    Permission-checked. Personal-data examples redacted before the prompt where the EU node is not used.

  3. 03

    7B runtime

    One node per office region.

  4. 04

    Comparison memo

    Saved only into the matter, by the associate.

  5. 05

    Partner review

    Negotiation stays outside the tool.

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 node per office region. The firm’s tenant, not a vendor account.

    Buyer effort after the runtime is up: 2–4 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

    Grae SI provides

    Qwen 2.5 7B. Clause-split long drafts before considering a 14B step-up.

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

  5. Layer 5

    Evaluation and validation

    Shared

    A tier-matched suite, thresholds, and an exportable report. The labeled set, the acceptance call, and any statistics your quality system requires on top.

    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 comparison view in the document system. No email integration in the first release.

    Buyer effort after the runtime is up: 4–8 person-weeks

  10. Layer 10

    Organizational validation and sign-off

    You own

    Knowledge partner and general counsel sign the rule of use, including the outbound ban.

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

Buyer effort for this profile

10–16 person-weeks

The gold set of agreements is the schedule. The model pull is not.

  • Knowledge partner
  • General counsel’s office
  • Firm IT
  • An associate team to label the 50-agreement set

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

8–12 weeks

Three offices, one playbook, no client-facing product.

Cost versus a cloud API

About $12k–$30k a month avoided

Illustrative: heavy daily use across several hundred lawyers and long agreements. Privilege and client terms rule out the API even where the price is close.

Compute

$2k–$5k a month

One GPU node in each of three regions, scaled to office hours.

First-pass review

1–2 hours saved on a standard BAA or CTA

Illustrative when playbook coverage holds. Unusual deals are not in this number.

Accuracy target

90% clause coverage, under 10% false differences

Gates on the 50-agreement set.

Assumptions

  • Illustrative estimate for this sandbox. It is not a measured customer result.
  • Internal work product only. The firm is not selling this as a product to clients in this profile.
  • Playbook quality limits the tool. A vague playbook produces a vague comparison.
  • Health, pharma, and insurance forms are the corpus. General corporate M&A is out of scope until the playbook exists.

All sandbox profiles · The full stack