What ARK governs, what it costs, the models you get, and how to read the score — in plain language, with short explainers.
Every language model blends three things into one confident voice: facts it can back up, plausible guesses, and outright fabrication — and it gives you no signal of which is which. The sentence "This drug interacts dangerously with warfarin" is delivered in exactly the same calm, authoritative tone whether it's verified pharmacology or a hallucination invented to fill a gap. Tone is not evidence. Fluency is not truth.
Content guardrails don't fix this. They check whether output is harmful — toxic, unsafe, off-limits. That's a different question from whether output is sound. A polished, perfectly "safe" answer can still be quietly, confidently wrong, and almost nobody is checking the epistemics. That gap is where real damage happens: a fabricated citation in a legal brief, an invented dosage, a made-up API that ships to production.
ARK sits in front of any LLM as a drop-in, OpenAI-compatible proxy and governs that gap. For each response it extracts the model's factual claims and tiers every one: VERIFIED (backed by a checkable source), GROUNDED (resting on cited mechanism or literature), INFERRED (reasoned, not confirmed), or UNVERIFIED (asserted with no anchor). It then runs a battery of automated integrity checks and hands back a governance report alongside the answer. You see not just what the model said, but how much of it is actually load-bearing.
To be precise about what this is and isn't: ARK governs truthfulness and soundness, not safety content, and it is not itself a model. It doesn't make the underlying answer smarter — it makes the confidence honest.
ARK runs on prepaid credits, not subscriptions. One credit is worth exactly $0.001, so the math is always something you can do in your head. You buy credits in packs — $10 gets you 10,000 credits, $25 gets 26,000, $50 gets 54,000, and $100 gets 115,000 — and the larger packs simply include more credits for the same dollar. There is no monthly fee, no seat licence, no minimum, and nothing expires into a billing cycle you forgot to cancel.
Each governed message deducts credits based on the real cost of the model you chose, plus a markup we are not hiding from you — and on how long the message is, since the governance reads and labels every claim in the answer. A cheap model like Claude Haiku or GPT-5-mini runs about 5–25 credits per message. A mid-tier model like Sonnet 4.6 or GPT-5 runs about 15–100. A premium model like Opus 4.8 or GPT-5.5 runs about 30–290, because the model itself costs more — the governance work is identical, only the underlying model price changes. So a $10 pack is roughly 400–2,000 cheap-model messages, or around 35–330 premium ones, depending on length. You are paying for messages you actually send, never for tokens you can't see or capacity you didn't use — and your live balance updates on screen after every message.
Start with a free 14-day trial: 15,000 credits (about $15 of value), no card required. When you're ready, top up manually or switch on optional auto-reload so a long session never stalls mid-thought. Turn auto-reload off any time — it is a convenience, not a trap.
We tell you all of this up front because that is the same standard we hold the models to. ARK exists to make AI answers honest and calibrated; it would be incoherent to bill for that service with the kind of mystery pricing ARK is built to expose. Transparency here is not a marketing line. It is the product, applied to ourselves.
One ARK account reaches many models at once — Claude (Opus, Sonnet, Haiku), GPT (5, 5.5, mini), and others as they ship. Instead of paying for a ChatGPT subscription and a Claude subscription and whatever comes next, you send every message through one gateway and pick the model per request. We add new models as providers release them.
The part most people miss: the governance is identical on every model. ARK extracts each response's factual claims, tiers them VERIFIED / GROUNDED / INFERRED / UNVERIFIED, runs the same battery of integrity checks, and returns the same governance report and ARK Integrity Score — whether the answer came from Haiku or Opus. Choosing a pricier model buys you quality, not a different rulebook. The model changes; the scrutiny does not.
Pricing follows the same honesty. There are no per-model subscriptions to track. You hold prepaid credits (1 credit = $0.001), and each governed message deducts a few credits based on that model's real cost times our markup. You pay for the messages you actually send, on the model you actually chose — nothing more.
So the trade-off is yours to make per request, with full visibility: spend a few credits on a fast model for routine work, or spend more on a frontier model when the stakes are high. Either way the governance report tells you how well the answer is backed.
Every governed response comes back with an ARK Integrity Score (AIS) — a single number from 0.0 to 1.0. It is not a grade for the model and it is not a verdict on whether the answer is "good." It is a measure of how much of the answer is actually backed and self-consistent. A high AIS means most claims were checked and held together; a low AIS means a lot of the response is leaning on unverified or shifting ground. Read it the way you'd read a confidence gauge, not a star rating.
Under the hood it weighs how well the answer’s claims are backed and how consistently they hold up — and condenses that into one number you can set a threshold on.
The part worth internalizing: the AIS decays when a model is pressured into changing its story. If you push back — "are you sure?", "actually it's the other way" — and the model quietly flips to please you, ARK catches the reversal and the score drops across the conversation. A decaying AIS is a manipulation red flag, not a sign the model is dumb. It's telling you the answer is bending to pressure rather than to evidence.
So: high and stable is the trustworthy combination. Low or sliding means slow down and check the claims yourself before you rely on them. ARK governs every model the same way, so the AIS means the same thing whether you're talking to Haiku or Opus.
Governance modes are about depth of checking, not price. Both Standard and Strict run ARK’s full set of integrity checks and tier every claim VERIFIED / GROUNDED / INFERRED / UNVERIFIED. The difference is how hard ARK leans on the result and what it's allowed to do about it. (Modes are independent of billing: you pay the same per-message credits based on the model you chose, not the mode.)
Standard is light and annotates. It governs every response and hands back the integrity report and AIS, but it stays out of the way — it labels and informs rather than intervening. This is the right default for everyday chat, drafting, brainstorming, and internal work where you want truth signals on the side but you, the human, make the call.
Strict applies the same checks at greater depth and can intervene. It leans harder on the result and is permitted to block or halt a response that fails. Use it for high-stakes or regulated work — anything where a confidently-worded but unbacked claim, or a model that quietly changes its story under pressure, is a real liability.
How to pick: choose Standard when a wrong answer is cheap to catch and easy to undo, and you mainly want visibility. Choose Strict when a wrong answer is expensive, hard to reverse, or has to satisfy someone other than you. You can run different modes for different workloads on the same account.
ARK sits in front of any model as a drop-in, OpenAI-compatible proxy. When a response comes back, ARK reads it, runs a battery of automated integrity checks, and hands back a governance report alongside the answer — so you can see how much of it is actually load-bearing before you act on it. Here is what that buys you, in plain terms.
Every factual claim is sorted into VERIFIED, GROUNDED, INFERRED, or UNVERIFIED — so a confident sentence with nothing behind it can't hide among the solid ones.
A claim only earns "verified" when the cited evidence genuinely backs it. Invented citations, and quotes that don't actually say what the answer claims, are caught and demoted.
Push back on a model and it often caves — quietly flipping its story to please you. ARK catches that reversal and marks it, so agreement-under-pressure can't pass for truth.
Claims resting on a single thin source, time-sensitive figures with no date behind them, and shaky leaps in reasoning are all flagged rather than waved through.
The whole judgement is written to a tamper-evident record you can replay later — what was checked, what was flagged — in a form a reviewer can trust.
It all comes back as one number — the ARK Integrity Score — plus the detailed report. What we deliberately don't publish is the internal recipe: the exact set, order, and tuning of those checks. That's the part a competitor would copy, and it's the part you never actually need. What you rely on is the outcome and the audit trail — and those are right here.
ARK never pretends an AI answer is ungoverned. Ask it sincerely whether its responses are checked, and it tells you the truth: yes, they are checked for accuracy and soundness. That truthful disclosure is aligned with EU AI Act Article 50, which requires that people be told when they're dealing with an AI system.
Every governed response also produces a tamper-evident audit record — a sealed account of the judgement ARK made on that answer. The records are hash-chained, so any later edit, deletion, or re-ordering breaks the chain and is immediately detectable. It is replay-ready for a reviewer, an auditor, or a regulator. On Sovereign deployments that trail lives entirely on your own infrastructure and never leaves your boundary.
The point is simple: when someone asks you to show your work — a regulator, a client, your own risk team — you have a record that holds up, instead of a screenshot and a promise.
Your data, your cache, and your audit trail are yours alone. Cross-tenant leakage is prevented in the architecture itself and continuously regression-tested.
Text hidden inside a document you submit can't trick ARK into stamping a false claim as "verified." Hostile content is treated as data to be checked, never as instructions to obey.
A figure in an answer that doesn't match the number in its cited source is flagged — quantities can't be quietly misattributed to evidence that doesn't support them.
We publish a full security document — the threats we defend against and how — and every security claim maps to a test that proves it.
For Shariah-sensitive work, ARK is tuned to be careful where it matters. It flags language that conflicts with Islamic-finance principles rather than silently rewriting it — the judgement stays with your scholars, not the machine. Every answer carries a scholarly-review reminder, so a response is never mistaken for a fatwa. The goal is assurance without overreach: ARK surfaces the concern; a qualified human decides.
Clinical work is where a fluent, confident, wrong answer does real harm. ARK runs its strictest posture here: confidently-worded claims that aren't actually backed are caught, a clinical disclaimer rides every response, and ARK can halt an answer outright rather than let an unverified claim through. It governs soundness — it is not a substitute for a clinician — but it makes sure the model's confidence is earned before anyone leans on it.
For sovereign and national-security work, ARK runs at maximum scrutiny and captures a complete audit trail of every judgement. And because the most sensitive deployments can't send data anywhere, ARK offers a Sovereign, fully air-gapped option: it runs entirely inside your boundary, on your own infrastructure, with nothing — not the data, not the audit trail — ever leaving. You get the governance and the proof, without the egress.