The maths
Seven of the ideas the system is built on, with toy numbers so you can check them yourself. The production versions combine several of these and use parameters we do not publish.
01 · Uncertainty and sizing
The size of every position is tied to how certain the system is. When uncertainty rises, exposure falls automatically, without anyone having to notice.
What this means for you: when the system is unsure, you are automatically less exposed.
02 · Correlated agreement
If your analysts all read the same news, their agreement is worth less than it looks. The effective number of independent opinions falls as their errors become correlated. The system measures how correlated its own AI analysts are and discounts their agreement accordingly.
where N is the number of analysts and ρ̄ is the average correlation between their errors.
What this means for you: a room full of AI agents agreeing does not fool the system into betting bigger.
Read the note: Five agreeing analysts are worth 1.47 opinions
03 · Calibration
A forecast is only useful if its confidence is honest. Group every forecast by the confidence it was made with, then check how often each group was right. A well-calibrated system sits on the diagonal.
Illustrative, not results: 60% stated, 57% observed; 70%, 71%; 80%, 78%; 90%, 86%.
What this means for you: you can see whether the system’s confidence deserves your trust, instead of taking it on faith.
Used in Assay
04 · Honest scoring
A good forecast is right and appropriately unsure. The continuous ranked probability score (CRPS) grades both at once. Lower is better, and it punishes a confident wrong forecast hardest.
where F is the forecast distribution and y is what actually happened.
Worked example: two forecasts both centred on +1%, one narrow (spread 1%) and one wide (spread 3%).
| What happened | Narrow | Wide | Scored better |
|---|---|---|---|
| +1.2% | 0.250 | 0.706 | Narrow: it was right and said so |
| +4.0% | 2.437 | 1.807 | Wide: it admitted it might be wrong |
| −3.0% | 3.436 | 2.562 | Wide |
What this means for you: the system is scored for being honest, not just for being bold, so it has no reason to overstate its confidence.
Used in Assay
05 · Sourced evidence
This one is a rule rather than a formula. When the AI reads a filing or a news article and extracts a fact, it must point to the exact characters it read. Ordinary code, with no AI involved, checks that those characters really say what the AI claims.
A second check confirms the passage actually supports the claim, not just that the words exist.
What this means for you: every fact behind a decision can be traced to the sentence it came from, and a made-up quote cannot get through.
Used in Scribe
06 · Human oversight
“A human approves every decision” means nothing if the human has stopped reading. So we measure whether human review is actually working, and treat any decline as a problem to fix.
What this means for you: human oversight becomes something you can check, not a sentence in a brochure.
07 · Honest uncertainty
Every automated check in the system reports two numbers, not one: how often it declines to decide, and how often it is wrong when it does. For example: “declines on 12% of cases, and is wrong on 4% of the rest, against a 5% target.” This lets you compare every safeguard on the same scale.
What this means for you: you always know how reliable each safeguard is, in plain terms.
Request access and a founder will walk you through it. Design partners get more of the maths, and the production versions are discussed in conversation.