The checks that guard your money and your name.
The parts of Jinn that protect your budget and your brand are deterministic code, not another AI’s opinion. Paid spend is metered and ceilinged; every public input crosses a fixed security perimeter; and each product’s publish gate is a hard rule a model cannot talk its way past.
Deterministic means you can reason about it. These checks do not ask a second model to judge whether your money is safe or your brand is intact - they decide it in code.
What it guards
Two things a language model must never be trusted to police on its own: the money you spend and the name you put on the work. Both are held by deterministic checks.
How spend is held
Paid AI calls pass one gateway that reserves budget before the call and settles it after, against per-brand and global caps. If the meter is unreachable, the call is refused.
How your name is held
Each product’s publish gate is code, not a second model’s opinion: banned-word checks hard-fail, a pixel check outranks the visual judge, and unbacked claims are held back.
How your money is guarded
Paid AI spend is checked and reserved against per-brand-daily, global-daily, and global-monthly caps before a call runs, then settled to the real cost or released. The check-and-reserve is a single atomic operation, so two calls racing for the last of the budget cannot both slip through.
If the meter itself is unreachable, the call is refused rather than allowed. A broken meter stops spending; it never allows untracked spending. That is what "fail closed" means, and it is the rule on every surface that gates paid spend.
Spend is reserved, then settled - not refunded to the cent the instant a call ends. An interrupted call can hold its reservation until it expires, which keeps the ceiling conservative by design rather than risking an overrun.
What every public input crosses
Every public, URL-accepting demo route passes one composable perimeter in a fixed order. Each check rejects before the next one runs.
- human check
- A challenge that fails closed on every failure mode - unconfigured, network error, replay, or invalid - so an automated caller never reaches the work behind it.
- address check
- A synchronous probe that rejects private-network and malformed URLs before any request is made.
- rate limits
- Per-IP and per-email windows that DENY when their backend is unreachable, because these routes gate paid spend - the safe default is to refuse, not admit.
- pinned fetch
- The fetch resolves DNS and pins the address, closing the rebind window a static check cannot, and concurrent requests for the same domain share one in-flight fetch.
- replay guard
- Each human-check token carries a one-time key and a five-minute replay guard, so the same token cannot be used twice even inside the provider’s own lag window.
This posture - refuse when unsure - is scoped to surfaces that gate paid spend. It is a deliberate choice per surface, not a blanket setting on everything.
That is the machinery around the work. The work itself is your brand, done to your rules. See your brand read
How your name is guarded
The same discipline repeats inside each product, and always as deterministic code rather than a second model’s judgment. In the writing product, a publish gate hard-fails on a banned word - the model does not get to argue it back in. In the ad product, a pixel-level contrast check is authoritative for legibility: if the math says the text is unreadable, the ad is rejected no matter what the AI judge thought. In research, a claim that cannot be backed by a quality source is downgraded and held back rather than published as fact.
The frame is simple: the checks that guard your money and your name are code you can reason about, not another AI’s opinion you have to trust.
Keep going
This page is the guardrail layer. The full tour shows how it wraps everything Jinn learns, wears, and works for a brand.
- The full tourThe mechanism end to end: how Jinn learns, wears, and works a brand.
- The recordThe 346-signal Brand DNA record, group by group.
- Models & routingWhich model does which job, and the gateway rules underneath.
- Measurement disciplineThe floors our own measurements must clear before we publish them.
- Check it yourselfWhat “verified” means here: human approvals, recorded provenance, and facts that age on purpose.
Guardrails, answered
- What happens if your cost meter goes down?
- The paid call is refused. A broken meter stops spending; it never allows an untracked charge. Fail closed is the rule on every surface that gates paid spend.
- Can a bot abuse the free tools?
- Every public, URL-accepting route passes a fixed perimeter first: a human check that fails closed, a private-address reject, rate limits that deny when their backend is down, and a DNS-pinned fetch. Each rejects before the next runs.
- Do you refund spend instantly if a call fails?
- Spend is reserved before a call and settled to the real cost after, or released on failure. An interrupted call can hold its reservation until it expires - conservative by design, so the ceiling is never overrun.
- Is a model deciding what is safe to publish?
- No. The checks that guard your name are deterministic code: banned-word gates hard-fail, a pixel check outranks the visual judge, and unbacked claims are held back. A model cannot talk its way past a hard rule.
Guarded by code, not by an opinion.
The parts that protect your budget and your brand are deterministic checks you can reason about. Run the free brand read and see the work they guard.