Specific questions with specific answers. Where the answer is “we do not know”, it says so.
Is Orakl a chatbot, or a language model?
No. The part that produces the numbers is a numerical simulation with no language model in it: several hundred agents holding probability distributions and averaging them with their neighbours. On devices that have Apple’s on-device model, that model is used to name the outcomes, name the drivers, cast the archetypes and write the prose — and it is explicitly prevented from supplying any figure. The method page has the detail.
Does it need an internet connection?
No. The default path performs no network requests at all, and the app works in airplane mode. There is one exception and it is opt-in: a Premium setting called “Ground in real sources”, off by default, which when switched on sends your question and context text to a lookup service so the simulation can cite articles. Nothing else is ever sent. See privacy.
If I ask the same question twice, do I get the same answer?
Yes, exactly the same answer, down to the per-round convergence values. The simulation is seeded with a stable hash of your question and context, so identical text produces an identical run on any device. We verified this by running the shipping engine twice in separate processes over five scenarios and comparing the output byte for byte.The seed is an FNV-1a hash of your question and context joined by a pipe character — not a timestamp and not a random draw. Swift’s own hashValue is randomised per process and could not have been used for this.
Change a single character and the seed changes completely, so the run changes completely. That means rephrasing a question is not a way to check whether an answer is robust.
How many agents are there, and how many rounds?
Free runs 40 agents for 12 rounds. Premium — labelled “Deep Simulation” in the app — runs 600 agents for 32 rounds. The graph topology and all the behavioural coefficients are identical in both; only the population and the number of rounds differ. A larger swarm gives a finer-grained vote count, not a more accurate one.40 agents over 12 rounds still runs the full pipeline: same graph, same eight archetypes, same bounded-confidence update. Nothing about the analysis is withheld from the free tier except population and depth.
How many archetypes argue?
Eight, by default, clamped to between three and twelve. Each one becomes a group of agents rather than a single agent, which is how several hundred distinct positions are maintained without several hundred model calls.
I have an iPhone and a Mac. Is my history shared?
No, and it cannot be. Orakl ships as four separate applications — iPhone, Mac, Apple TV and Apple Vision Pro — built from one codebase. None of them syncs with the others, because none of them talks to a server. Each holds its own history in its own private storage on that device. If you run the same question on two of them you will get the same result, because the run is deterministic; you will not see the same history.
Are they one purchase or four?
They are separate products on the App Store and are purchased separately. Restoring inside a given app uses Settings → Restore Purchases, signed in to the Apple Account that made the purchase.
What happens on a device without Apple Intelligence?
Everything still runs. Each of the four model-assisted stages has a deterministic fallback, and the simulation itself never needed a model. What you lose is the quality of the words: outcome labels become cropped two-word phrases, and the drivers become bare terms lifted from your sentence, assigned to a side by hashing the word. That last point is a genuine weakness and we describe it properly under limits.
You can tell which path a report came from: the line “Generated privately on your device” appears only when the on-device model wrote it.
What is in the free tier?
Five simulations in total — a lifetime allowance, not a monthly one — at 40 agents and 12 rounds, with the most recent analysis kept in history. Premium lifts the run limit, raises the swarm to 600 agents over 32 rounds, and keeps the full archive.
Why does the leading figure say 50% when the chart clearly shows more than half?
The figures are truncated rather than rounded. A vote share of 0.5083 is displayed as 50%. The prose report alongside applies the same truncation independently, so the two can differ by one point.
Why is Confidence 0% when the app still picked a side?
Confidence is one minus the normalised entropy of the vote. On a two-option question a near-even vote has almost maximal entropy, so Confidence collapses toward zero even though one side is nominally ahead. It is telling you the margin is worthless, not that the app malfunctioned. Reading the numbers works through the combinations.
Can it tell me who will win an election, or a match?
No, and the app no longer offers to try. Orakl models decisions you own — career, business, money, personal, product, team. The spectator categories were removed during 1.0 because the app had no basis for them. If you type such a question anyway you will still get output; it will be a simulation of an argument about your sentence, not information about the event.
How accurate is it?
We do not know, and we cannot find out. The only record of what actually happened lives on the user’s own device, because nothing is uploaded. The app lets you mark afterwards whether the leading outcome came true, and shows you your track record; we never see it. We would rather be unable to publish a hit rate than be able to.Which is also why you should treat a published accuracy figure from any app in this category with suspicion: to compute one honestly, someone has to be collecting your decisions and their outcomes.
What languages does it support?
The interface is localised into 34 languages. Note that the analysis reasons over your own words, so a question typed in one language and a question typed in another are different inputs and will produce different runs.
Is anything I type stored or sent anywhere?
Stored: yes, on that device, in the app’s private storage, and deleting the app deletes it. Sent: no, unless you have switched on the optional grounding setting described above. There is no account, no analytics and no third-party SDK collecting anything.
Can I delete my data?
Delete an entry from History, or delete the app. There is no account to close and no server-side copy to request, because there is no server-side copy.