How likely, how much, and when.
The instrument that puts a number on things before you act, and the first one that gets graded. A shot that looks risky and a shot that is risky are different objects. Risk has a reality delta too.
- INPUT //a decision-shaped question · an appeal's odds, a plan's chance, a posted price
- RULE //no denominator, no number · no settle condition, no weight
- SPLIT //the instrument quotes odds · conviction and stakes stay with the operator
- SEALED //every weight goes into git with a date · reality grades it when it resolves
01THE CONTRACT
Every estimate is one object. A number has to carry its receipts.
A published rate is a claim, not a fact: easy cases mixed in, abandoned applications never counted, a firm padding its stats. Audit the counting, then take the haircut.
02THE LADDER
Three fields independently rank their evidence and let the rank cap the confidence. This adds the step none of them took: the grade becomes a hard numeric ceiling.
Anecdotes can say possible. They are never allowed to say how often, and no volume of them tightens the band past its ceiling.
03THE BET LAYER
Knowing the odds is half. The other half is when not to shoot, and how hard when you do.
When the consensus sits within friction of your estimate, there is no profitable move even if you're right. Being right isn't enough; you must beat the friction.
The edge estimate is itself a guess; betting big on an inflated guess ruins you faster than betting small on a real one. Under it sits a ruin bound: computed, never vibes.
04THE CHAIN TAX
Multi-step plans fail through hidden couplings: steps that look independent share an actor, an assumption, a deadline, and die together.
NASA's procedure became ours: hunt the shared factors with a checklist before any math. In their own worked example the true failure risk was twenty times worse than the naive multiplication. The naive number is never shown alone.
05THE COURSE
Dead reckoning: project from known position and heading, then correct as real fixes arrive. These are live positions, sealed 2026-07-10. They can miss. That is the point.
Alonso wins the Home Run Derby?
June CPI prints at or above the consensus line?
The GC leader holds to Paris? The recorded line predates the stage that decided it.
- THE BOOK //17 sealed · 10 events resolving 07/11→07/31 · 7 UK planning applications deciding in August
- THE TWIN //every question also answered by a naive twin with no doctrine · the contrast is the experiment
- DEAD-ZONE DEFERS //2 of 10 sat out on purpose · when consensus is close enough, being right isn't enough — you must beat the friction
- STATUS //GRADING PENDING · the ledger is born at the first resolution
Forecasters aren't calibrated because they're brilliant; they're calibrated because they get graded every day. When it says 70%, it should happen 70% of the time. The dots land here as the fixes arrive.
06AGAINST THE AVERAGE MACHINE
A GENERIC MODEL, ASKED THE ODDS
- Says "78%," and cannot say where it came from.
- Confident either way. Never graded.
- Finds a forum post and dresses it up as data.
DEAD RECKONING ///
- Carries its reference class, denominator, tier, and settle date.
- The band width is the honesty; the grade caps it.
- Sealed in git. Scored when reality answers.
If the layer doesn't beat the operator's unaided gut after a real sample of scored calls, it gets called a failure and we say so.
NEXT FILETHE HARNESS — STILL LOADING▸