Privacy — Norma

Norma is built so the trust question has a concrete answer, not a promise. Below is what the agent can see, what it stores, and the line between your studio's work and the shared industry corpus — stated plainly, with the specifics that back each claim.

Identity-only sign-in

You sign in with Google for identity only. Norma does not request bulk Drive or Gmail scopes — it never asks for read access to your inbox and never crawls your Drive. Sign-in tells Norma who you are; it does not hand over your files.

Work reaches the agent through per-resource shares. You share a specific Sheet, doc, or folder with the agent's address — …@agent.norma.llc — exactly as you would share with a colleague. The agent can only touch what you have explicitly shared with it. Unshare a resource and the access is gone; nothing else in your Drive is ever in scope.

What stays private to your studio

Your firm corpus — every spec book, schedule, and note you share — stays private to your studio. Client names, project labels, internal notes, and dealer-net pricing live against your account and are never pooled into the shared corpus or attached to a shared catalog row.

Dealer-net pricing is the clearest case: the dealer-specific discount is per-quote data. Mark it with the literal [PRIVATE] column marker and it never leaves your studio's side of the line.

The [PRIVATE] column marker

Add the literal substring [PRIVATE] to any column header in a sheet you share — e.g. Client Name [PRIVATE], [PRIVATE] Project Code, or Dealer Net [PRIVATE]. From that point on, Norma's read tool replaces every value in that column with the sentinel [PRIVATE] before the data leaves the server.

The agent literally cannot echo those values — it never sees them. The filter is server-side, not a prompt instruction, so it can't be talked around. Drop the marker the same way to bring a column back into scope.

What enters the shared corpus

When Norma parses a manufacturer URL or a vendor line card you drop in, only the public product facts — SKU, list price, dimensions, designer, certifications, lead time — get written to the shared industry corpus. That is how Norma learns vendors over time, and how every studio benefits from facts confirmed once.

What does not enter the corpus: anything a column marks [PRIVATE]; anything tagged with your name or a client's name; and any dealer-net price. Names, projects, and dealer-net are redacted before pooling — only the public fact about the product survives, never the context of who specified it or what they paid.

What Norma learns

Norma learns from the work — that is the product. It's worth being precise about what that means, because three different things happen and they have different boundaries.

Norma learns your practice: preferred manufacturers, project conventions, client standards, corrections you make. That learning is scoped to your firm and follows the privacy line above — it never pools, and it's what makes your agent yours.

Norma learns the industry: when you correct a parse or confirm a product fact, the system also gets better at reading vendor pages, spec sheets, and quotes in general. That learning is pooled — it's why Norma improves for every studio — and it crosses the line only after client names, project labels, and pricing context are filtered out. What pools is skill at reading documents and public facts about products, never facts about your practice.

Norma's models run on Anthropic's API, which does not train its foundation models on API traffic — per Anthropic's Commercial Terms of Service, Section B ("Customer Content"): "Anthropic may not train models on Customer Content from Services." So the learning described here is Norma's own — deliberate, filtered, and bounded as described — not a side effect of the model provider.

That covers every Norma surface. Work you send by email, the turns Norma's agent runs behind the Claude connector, and the library work behind the API all execute as API traffic under Norma's account — none of it becomes Anthropic training data, whichever door you came through. The one thing this page can't govern is your own Claude account: the chat conversation around the connector runs under your firm's Claude plan. On Claude for Work (Team and Enterprise), Anthropic does not train on your content by default either.

Sourced and reversible

Every fact carries provenance. Tier label, list price, sale price, lead time, dimension, carbon score, certification, designer attribution — each one in Norma's output carries a [source · YYYY-MM-DD] citation pointing at the URL or PDF it was parsed from, with the parse date.

Ask "where did that come from?" on any cited fact and the agent returns the source URL, the source kind, the model used, and what changed since the prior parse. Audits never trust the cache — they re-parse the source and tell you what moved. Because everything is sourced, everything is reversible: you can trace any fact back and roll it back.

Where inference runs

Inference runs on Anthropic's API — Claude Sonnet for chat orchestration, Claude Haiku for cheap structured extraction. Anthropic's API processes requests but does not train its foundation models on API traffic; see "What Norma learns" above for the exact terms citation, and for what Norma's own learning does and does not touch.

Norma retains shared resource contents only as long as you keep the resource shared. Unshare it, or revoke the connection, and Norma's access is wiped from our side.

How long Norma keeps things

Retention follows the same rule as access: Norma holds data for exactly as long as it has a job to do with it, and the trail stays inspectable the whole time.

Your account. The identity Google hands us at sign-in — name and email — lives as long as your account does. Ask us to close the account and the identity record goes with it.
Shared resource contents. Held only while the resource stays shared. Unshare a Sheet or revoke the connection and Norma's copy of that access is wiped from our side, as described above.
Library facts and their evidence. A fact you keep in your library keeps its source citation, and the parsed evidence behind the citation, for as long as the fact is there — that is what makes "where did that come from?" answerable years later. Remove the fact and its evidence loses its reason to stay; the schedule below governs when it goes.
Email threads with your agent. Messages you send to your agent's address are kept as part of the working record — the audit trail that makes every write traceable and reversible.

The formal retention schedule, with numbers per data class, ships alongside the DPA below. Where this page and that schedule ever disagree, the schedule wins.

Subprocessors

Norma runs on a deliberately short list of infrastructure providers. Each one processes firm data only to provide the service described:

Cloudflare — application hosting, database, file storage, and inbound email routing. Everything Norma stores lives here.
Anthropic — model inference, as described above. API traffic; no foundation-model training.
Google — sign-in identity, and the Sheets/Drive APIs for the specific resources you explicitly connect. Nothing beyond what you share.
Resend — outbound email delivery, so your agent's replies reach your inbox.

When billing goes live it will add a payment processor (Stripe) to this list; we will update this page before that happens, and the DPA process below is how you get notified of any subprocessor change.

Data processing agreement

If your firm needs a signed DPA — for client contracts, EU work, or your own vendor review — email hello@norma.llc with "DPA" in the subject. We are finalizing a standard signable agreement covering the posture on this page: identity-only sign-in, per-resource scope, the [PRIVATE] redaction guarantee, the private/shared corpus line, the subprocessor list, and the retention schedule. If your review needs something this page doesn't answer, ask — the honest state of a term is always available.

Open questions, openly

This page is the working privacy posture, stated as plainly as we can. For anything not covered above — or to flag something we got wrong — reach us at hello@norma.llc.