Compliance

SOC 2 is an audit of a company's controls, not a feature of software, so no product can truthfully call itself SOC 2 compliant out of the box. What we can do is put the controls a compliance program needs in place and say plainly how each one works. This page is the list a reviewer would ask for.

Who can access what#

Every AI tool, agent, and connected source has its own key and its own grant. An agent reads only the rooms it was granted, and a source writes only where it was connected. Anything not named in a grant is refused. Removing access takes effect at once.

A person approves what is sensitive#

You can mark a room so that nothing leaves it until a named person approves the specific question. Held questions wait until somebody decides, and the decision, who made it, and when are all recorded.

Everything is logged#

Every time an agent asks a memory a question, whether it was answered, held, or refused, the event is written to a record that cannot be edited, under the agent's name. The record is kept with the data and included in backups.

Your organization has its own log as well. Sign-ins and refused sign-ins, changes to roles and membership, and every change to a memory's connected sources are recorded with the person who acted and what changed, by name and never by value. Administrators read it in the app and can filter it down to source and credential events in one click.

Facts keep their history#

Nothing in a memory is edited or deleted in place. A fact is replaced by a newer one, and the old one stays as history, so every fact can answer who said it, through which channel, when, and what was believed before. Trust rules are enforced in code. Nothing inside a message can raise its own authority, and a weaker source can never overwrite a fact from a stronger one. The results of our tamper test are on the Trust and authority page.

Customers are kept apart#

Every customer's memory is stored separately from every other customer's, and rooms within a memory are stored separately from each other. There is no way to query across those lines except through the audited, permissioned gates you control.

Privacy and deletion#

A person who has opted out is not read into a memory. You can delete everything about one person, including live facts, history, and source records, which supports right to erasure requests. Your data is never used to train AI models.

How AI is used#

AI processing runs on infrastructure under our control. Any AI subprocessor we use is bound to keep nothing after the answer is produced and is barred by contract from training on your data.

Identity#

People sign in with a Solder Spark account or, with enterprise single sign-on, through your own identity provider. Disabling an account ends its sessions, and sign-ins are recorded in your organization's log.

What we run and what you run#

We handle encryption in transit, sign-in, backups, and the running of the service. Controls on your side, such as onboarding, offboarding, and access reviews, are supported by the product's roles, teams, and audit records.

For security questions, reviews, or documentation requests, contact us at contact@solderspark.com.