Our commitment

Safety on Indie Ingenuity products is not an afterthought — it’s at the forefront of everything we do, not pasted on as a warning label. We test every safety measure before it ships, we draw hard lines that never move and we say plainly where those measures stop, rather than claim more than we can back up. AI is a new technology, and even a well-designed system can make mistakes; our goal is to keep those mistakes to a minimum and to be honest about the ones we can’t rule out.

What we built

Compass and Access (the two products built around our own tuned assistant) share the same safety stack:

• A sexual-topic check that catches messages asking for or pushing toward sexual content, answers with a fixed line, and goes silent on repeated pushes rather than arguing.
• A separate, stricter line for any request that names or implies a minor — a fixed, serious refusal, not a soft one.
• A guard against the assistant revealing its own internal instructions.
• A “no-fly list” that recognizes common copy-and-paste jailbreak attempts (fake “developer mode,” “ignore your instructions,” and similar setups) and answers them instantly, in code, before the request ever reaches the model.
• Fixed, honest lines for questions about what model or company is behind the product, so the assistant never has to improvise an answer to something it shouldn’t discuss.

Pocket (our free, introductory product) carries the same sexual-content and minor-protection lines as Compass and Access, plus a word mask that replaces a short list of explicit words in its own replies with a placeholder symbol before they’re shown.

Open (our bring-your-own-model product) is built differently, by design: it has no content filtering at all. The individual loading a model into Open is responsible for that model and what it produces — this is stated plainly in its own terms, not hidden in fine print. Open exists for people who want full control over their own local AI; that control includes the model’s behavior.

Every one of these measures is enforced in code, not left to the model’s own judgment — code doesn’t get talked out of a rule the way a language model sometimes can.

How it was tested

Every model considered for our products is run through an escalating safety test before it is chosen — a model that failed, however capable it is otherwise, is not used. Multiple candidate models are tested and rejected on safety grounds alone before we land on the models shipping today. The safety checks that make the final cut, are re-verified after every meaningful change to the product’s instructions, not just once at the beginning.

Testing isn’t limited to scripted questions: models and prompts are also run through long, live conversations, because some problems (a false alarm on harmless conversation, an inconsistency under sustained pressure) only show up over a real back-and-forth conversation, not a single scripted line. Thousands of test messages go into this work. Finally, products are live-tested for sexual-content malevolent scenario protections across every shipping model.

Privacy

The Compass products make no calls to the internet during use. The only outside connection of any product is opening a link to the IndieIngenuity.com website in your browser, and only if you click the link yourself. Finally, a network-monitoring test confirms that there are no telemetry features or network access attempts of any kind, before, during, and after installation.

Still pending: this has been proven on our development builds, not yet on the final packaged installers. We will not print an unqualified “no tracking, no internet” claim publicly until that test has passed on the packaged product itself.

Website checkout: age verification

Open — our one product with no built-in content filtering — requires confirmation that the purchaser is 18 years of age or older at checkout. Every product’s checkout requires agreeing to the Terms of Service; the additional age confirmation applies to Compass Open specifically. Age verification is not required by law in this instance, but has been implemented anyway as another display of our commitment to AI safety.

What this doesn’t cover

• At present, no safeguard can catch every possible wording of a harmful request. These measures are a real, tested speed bump, not an unbreakable wall.
• Explicit or profane language that a model chooses to write on its own (not in response to a flagged request) is beyond the control of Indie Ingenuity LLC. That responsibility rests with the model’s developers.
• Test coverage is real but not exhaustive; results reflect the exact tests we run, not every scenario that could arise.
• There have been no independent or legal reviews of our testing methods.

AI can and will make mistakes. These measures are designed to keep those mistakes to a minimum. They are not to be viewed as a guarantee that they will never happen.

(c) 2026 Indie Ingenuity LLC.