Frequently Asked Questions

What problem does Constellation solve?

Your firm's most valuable asset is the accumulated knowledge of its best employees. Constellation helps you capture it.

What are firms missing today?

While firms are cautious about preventing outside providers from retaining their raw information, they rarely retain the learning generated by their own employees.

Why not just use a frontier model on our documents?

Atlas supports frontier models. However, frontier models are not cost- or data-sensitivity appropriate for all tasks.

How is this different from Harvey, Hebbia, Rogo, or other applications?

The tools and workflows that we build are customized to each firm.

What do we own at the end of a deployment?

A record of how your people actually work, an evaluation set only you have, locally-trained models where appropriate, and the rules for which model handles which task.

How secure is it?

Runs where you say. Keeps nothing. Learns for no one else.

Where does Atlas run?

On infrastructure you control: your servers or your cloud. Air-gapped if you need it.

What leaves our firm, and what do you keep?

Only prompts sent to a model provider, under zero retention, and only for tasks you choose to send. Your work, rules, and tests stay with you for as long as you run Atlas.

Are you training on our work?

No. Your documents and corrections never train a model for anyone else. Model providers run under zero retention.

Do agents hold credentials?

No. Passwords and keys never enter the model. Access is granted at the connector, one action at a time.

How are privilege and information barriers kept?

The same walls that govern the work govern what Atlas learns. A rule taught on one side of a wall is invisible on the other. Desk-level rules stay on the desk until someone with authority promotes them.

How do we get documents for a security review?

From the Trust Center, or email security@constellationfinance.ai. We turn questionnaires around from there.

How difficult is it to implement?

Local deployments with white-glove service from the Lab.

How long does it take to set up?

The Lab team can work as quickly as you can. Upon signing with Constellation, the Lab allocates full-time FDEs to your project - work with them as long as you'd like.

Do I have to write new prompts or skills?

No. You do your work. Atlas asks when it is unsure. Plus, Atlas imports your existing skills and prompts

Do you support our SSO?

Yes. Identity stays in the system you already run.

Can we keep using Claude, GPT, or Gemini?

Yes. Atlas works with any provider. Switch, re-run your tests, and keep everything you built.

Do we need a firm-wide program to start?

No. Start with one desk or practice group. We deploy, watch live work, and build the first test set from it.

How long until we see something?

Weeks, not quarters. The first test set built on your own work is the first milestone.

How does it learn?

From your existing documents, your chats, and your feedback.

How do I teach Atlas?

You mostly don't have to. Atlas starts by archiving your firm's documents. After that, every query you run teaches it a little more.

What counts as a evaluation?

A real piece of past work with a known good outcome. Pass means the output lands within tolerance, not that it reads well.

Who decides when a local model is good enough?

You do, using the evaluation set built from your own work. We never promote a model because a public leaderboard moved.

How is Atlas different from other self-improving harnesses?

Three ways.

First, Lab builds it for you: Atlas is not a self-serve platform you configure with plain-English runbooks. Our team deploys inside your firm and customizes the harness to your stack.

Second, corrections are governed: Atlas captures not just what was accepted but why, with authority, scope, provenance, and version. A partner rewrite on an IC memo is a different label than a first-year accept on a formatting pass.

Third, local deployment is the default, not an upgrade. The firm model never leaves your perimeter.