Not a disclaimer at the bottom of a sales page. This is part of the description of what the product is.
The app states this on every report, and we will state it here in the same words it uses:
Orakl is an analytical simulation. Its output is a model, not a fact — it is not financial, medical, legal or professional advice.
Nothing in a run is prepared by a person qualified to advise you, and nothing in it is specific to your jurisdiction, your circumstances, or your obligations. If a decision has legal, medical, tax or safety consequences, the output of a simulation is not a substitute for someone accountable for the answer.
There is no database, no index, no memory of previous questions, and on the default path no network. The entire evidence base of a run is the question and the context box.The single exception is the optional “Ground in real sources” setting, off by default, which fetches a capped digest from a lookup service. Even then, the swarm is reasoning over a summary of a handful of pages, not over knowledge.
This has a direct and testable consequence: the model cannot tell you that you have left out the decisive fact, because it has no way of knowing that the fact exists. It will happily produce a clean-looking result from a paragraph that omits the most important thing about your situation. A thin input produces a thin analysis dressed in the same three confident-looking figures as a thorough one.
This is the least flattering fact about the implementation, and it belongs here rather than in a footnote.
When the device has Apple’s on-device model available, the outcomes and the drivers are chosen by that model, and each driver is assigned to the outcome it actually supports. When the model is not available, a deterministic parser stands in. It picks salient terms out of your text — that part is reasonable — and then decides which outcome each term favours by taking a hash of the word and reducing it modulo the number of outcomes. The weight is derived the same way.
So on a device without Apple Intelligence, the for-and-against columns are not an analysis of your situation. They are a stable, reproducible, arbitrary partition of the vocabulary of your sentence. The simulation that runs on top of them is a real simulation, faithfully executed, over a prior that was assigned by a hash function.Stable and reproducible matters here: the same word always lands on the same side, so the result does not wander. But it is not evidence about your decision, and we would rather say so than let the two-column layout imply otherwise.
We think this is still worth shipping, because the structure — name the options, name the drivers, see how a divided group moves — has value even when the allocation is arbitrary, and because the alternative was to leave a large fraction of devices with no product at all. But you should know which path your device is on. The report says “Generated privately on your device” when the on-device model wrote it; if that line is absent, you are on the deterministic path.
Orakl models decisions you own: a job, a price, a hire, a move, a product cycle. It does not model events you merely watch. The spectator categories — sport results, world events, markets — were removed from the app in the ninth build of 1.0, deliberately, because the app had no business offering them.
If you type a question about an event you do not control, you will still get a result, because the engine does not know the difference. It will be a simulation of how a divided synthetic panel would talk about the sentence you wrote. It will not be information about the future.
The parameters of the model were chosen to make it argue, not fitted to data. Specifically, and so that you can weigh them yourself: the social graph connects each agent to four neighbours on each side with 10% of edges rewired at random; the drift toward the extracted prior is 0.05 scaled by the agent’s expertise; the contrarian pull removes 25% of the majority’s belief mass scaled by the agent’s contrarianism; belief updates are a bounded-confidence average with a listening threshold of twice the agent’s openness.
We cannot show you that those are the right values, because we have not validated them against resolved real-world decisions. We do not have such a dataset and we are not going to imply that we do. Change the rewiring probability and the same question can converge differently. That is true of every opinion-dynamics model in the literature, and it is true here.
Running 600 agents instead of 40 does not repair any of this. It samples the same dynamics more finely.
We do not know Orakl’s accuracy, because there is no way for us to find out: the history that would be needed to measure it never leaves the user’s device, which is the same property that makes the app safe to type a real decision into. That trade is deliberate, and it cuts both ways. We would rather be unable to quote a hit rate than be able to.
We also do not know how the simulation’s behaviour compares with how real groups of people converge. The model is drawn from a well-studied family in the opinion-dynamics literature; the specific configuration here has not been compared against human data by us or, as far as we know, by anyone.