Three figures come back from every run. They are easy to misread in opposite directions, so here is precisely what each one is a function of.
Every figure Orakl reports is computed from a single object: the distribution, which is the fraction of agents whose strongest belief, after the last round, sits on each outcome.
All four are displayed truncated rather than rounded, so a vote share of 0.5083 appears as 50%, not 51%. If a bar looks slightly ahead of what the number says, that is why.
Let d be the gap between the first and second outcome. Split is defined as 1 − d. With exactly two outcomes, the distribution is fully determined by d as well — the shares must be (1+d)/2 and (1−d)/2 — so Confidence is also a function of d and of nothing else.This is a property of the definitions, not a bug. But it means that on a two-option question, “Confidence 2%, Split 80%” is not two pieces of evidence agreeing with each other. It is one piece of evidence printed twice in different units.
| Two-outcome vote | Confidence | Split |
|---|---|---|
| 51 / 49 | 0% | 98% |
| 60 / 40 | 2% | 80% |
| 75 / 25 | 18% | 50% |
| 90 / 10 | 53% | 19% |
| 99 / 1 | 91% | 2% |
Two consequences follow. First, on a two-way question, Confidence is brutal: a 60/40 result — a twenty-point margin — scores 2%. That is not the model being defeatist. Entropy is simply unforgiving about near-even binary votes. Second, the interesting combinations only exist when there are three or more outcomes on the table.
Impossible with two outcomes; routine with three. A vote of 50 / 48 / 2 gives a split of 98% and a confidence of 29% — far above the 0% that the equally-tied 51/49 two-way vote scores.
| Three-outcome vote | Confidence | Split |
|---|---|---|
| 50 / 48 / 2 | 29% | 98% |
| 45 / 45 / 10 | 13% | 100% |
| 60 / 20 / 20 | 13% | 60% |
| 70 / 15 / 15 | 25% | 45% |
| 80 / 10 / 10 | 41% | 29% |
What it means. The swarm has done something genuinely useful: it has eliminated an option. The mass collected on two of the three and abandoned the third. The remaining two are tied.
What to do. Treat the run as having answered a different question than you asked. It has told you which option is out, and it has told you that the other two are a real choice that this model cannot make for you. Re-ask with the two survivors and see whether adding context separates them.
A vote of 60 / 20 / 20 has a comfortable forty-point lead — split 60% — and a confidence of 13%. The leader is clear and the reading is still weak.
What it means. The opposition did not consolidate. Nothing beat the leader, but the swarm also never formed a coherent second camp: the objections were scattered across the remaining options rather than pointed at one alternative.
What to do. This is the shape to be most careful with, because the headline percentage looks respectable. A leader that wins against a fragmented field is not the same as a leader that survived a real challenge. Look at the drivers listed underneath and at what the report agent says would flip it. If nothing on that list is something you actually believe, the lead is an artefact of the vocabulary the extractor happened to pick out.
Leans 100%, Confidence 100%, Split 0%. This should raise your guard rather than lower it, and it is common on short inputs. A four-word question can produce a single factor, one prior and nothing for anyone to disagree with; the swarm then agrees perfectly and the entropy of a unanimous vote is zero by definition. Example three is exactly this case, shown in full.
The app also draws “How consensus formed”: the mean belief mass on the leading outcome at the end of each round. It is worth more attention than the three headline figures, because it shows direction rather than a final state.
Two runs can end on the same three numbers and have completely different curves. If a result matters to you, the curve is where the difference is visible.
None of these figures is calibrated against anything. There is no historical outcome data behind them, no base rate, no backtest. Confidence is not a probability that the leading outcome will occur, and we do not know what the relationship between the two would be, because we have not measured it and would need real resolved decisions at scale to do so.
The app does let you record what actually happened afterwards, so you can build your own record of when it was right. That record is yours, on your device, and we never see it. We cannot therefore tell you Orakl’s hit rate, and you should be sceptical of anyone in this category who quotes one.