Three decisions of genuinely different shape, run through the shipping engine, with every figure as the app would display it.
These are not mock-ups and they are not illustrative numbers. Each is a run of the engine that ships inside Orakl, driven by the exact text shown, at the Premium configuration of 600 agents over 32 rounds.
The on-device language model was deliberately declined for all three, so that anyone can reproduce them.This is the same code path a device without Apple Intelligence takes. It has a visible signature: outcome labels are cropped to two words, and the drivers are bare terms lifted out of your sentence rather than phrases someone wrote. On a device with the model, the same run produces the same numbers with better words around them. Run twice in two separate processes, the three runs were byte-identical, convergence curves included.
One property of the model-free path matters before you read the columns below, and it is the least flattering thing on this site: on that path, which side a driver lands on is decided by hashing the word. The weights are real and the simulation is real, but the allocation of a term to “for” or “against” is arbitrary in the absence of the language model. We say more about that under limits.
Should I take the senior role at the smaller company or stay where I am?
The offer is a 14% raise and a title change. My current manager is good and I have been here four years. I am tired, and the commute would be longer.
Final vote share 0.4917 “Smaller company” / 0.5083 “Stay where”. The report agent scored the run at 80% fragility.
The swarm did not merely fail to reach a conclusion; it moved away from one. The convergence curve opens at 0.548 and descends to about 0.504 over the 32 rounds, meaning the leading option commanded less belief at the end than at the start.Full curve: 0.548, 0.543, 0.539, 0.535, 0.532, 0.529, 0.527, 0.524, 0.522, 0.520, 0.518, 0.515, 0.513, 0.512, 0.510, 0.508, 0.506, 0.505, 0.503, 0.502, 0.500, 0.501, 0.507, 0.506, 0.504, 0.503, 0.501, 0.500, 0.508, 0.507, 0.505, 0.504. All eight archetypes finished exactly where they started, and their internal agreement ranged from 0.76 to 1.00 — that is, each camp stayed coherent and no camp persuaded another.
The correct response to this output is not to take the 50% as a recommendation. It is to notice that on the information supplied, the two options are indistinguishable to the model, and that a result this balanced will flip on almost any addition to the context. If a decision matters and comes back looking like this, the missing input is yours: what would have to be true for one side to win?
Should we raise our subscription price from 4.99 to 7.99?
Churn is currently 3.1% monthly. Support volume is low. Two competitors charge 9.99 and neither has our offline mode. Our costs rose with the new hosting contract.
Final vote share 0.9983 Yes / 0.0017 No / 0.0000 Too close to call. Fragility 1%.
Here the question opened with “Should”, so the answer space is three-valued — Yes, No, Too close to call — and four of the six drivers fell on the same side. The swarm started almost evenly divided (belief mass on the leader was 0.487 in round one, below half) and then consolidated steadily to 0.517 by round 32, with five of the eight archetypes abandoning their opening position.
That is what a real consolidation looks like, and it is the one case where the headline number is close to self-explanatory. But note what it is not telling you. It is not telling you that raising the price is correct. It is telling you that within this text, the terms the extractor found pointed overwhelmingly one way and the simulated panel followed them there. Feed it the same decision framed by someone who is against the rise, and it will consolidate the other way with equal enthusiasm.Which is the honest reading of any result from a model whose entire evidence base is the paragraph you typed. Confidence measures the concentration of the vote, not the quality of the input.
Should I move to Lisbon?
(none)
Final vote share 1.0000 Yes / 0.0000 No / 0.0000 Too close to call. Fragility 3%. All eight archetypes were cast from the same single entity and were named identically.
This is the most important of the three, and it is the one a marketing page would never show you. Every figure is at its maximum, and every figure is meaningless.
Four words produced one factor. One factor produced one prior, pointing at one outcome. The archetypes were all cast from that same entity, so although three of them were assigned opening leanings on different outcomes, none of them had anything to defend that position with. Belief mass on the leader climbs monotonically from 0.635 to 0.798 and the vote ends unanimous. Confidence of 100% here does not mean the model is certain you should move to Lisbon. It means there was nothing in the input to disagree about, and the entropy of a unanimous vote is zero by definition.
The correct reading of a maximal, unanimous, zero-split result on a short question is: discard it and type more. The heuristic we would offer is that the result is worth reading only when the drivers listed are things you recognise as the actual considerations. If you look at that list and think “those are not the reasons”, then the number underneath is arithmetic performed on the wrong inputs, however confident it looks.
| Run | Outcomes | Leans | Confidence | Split | Fragility |
|---|---|---|---|---|---|
| Job offer | 2 | 50% | 0% | 98% | 80% |
| Price rise | 3 | 99% | 98% | 0% | 1% |
| Lisbon | 3 | 100% | 100% | 0% | 3% |
Rows two and three look almost the same and mean opposite things. Distinguishing them is what the next section is for.