Predictions
A ledger of what managers actually said, graded honestly against what actually happened.
A prediction is a claim someone made, recorded word-for-word. A manager on a call, a newsletter writer, or you in a moment of conviction. The ledger keeps it, and later you rule on how it turned out.
Why this exists
Everyone remembers their good calls. Nobody keeps score of the rest — and the people you follow have no obligation to either. That leaves you allocating trust on the basis of recent vividness, which is the worst available criterion.
A written record fixes that. Not to catch anyone out, but so that your trust weights have something behind them.
The claim and its grading are kept apart
This is the design decision that makes the ledger trustworthy, and it's worth understanding:
- The claim is stored verbatim, in the speaker's own words, exactly as said.
- The falsifiable condition — what would have to be true for it to count as right — is written separately, by you.
Keeping them apart means nobody gets to soften a claim after the fact by rewriting what "really" was meant. The words stay put; the standard is explicit and set in advance.
Only claims someone actually published belong in the ledger. Something an AI recalls a manager saying is a lead to go check, not a record.
Four kinds of claim, never averaged together
A prediction is one of four types, and they are scored separately:
| Type | Example |
|---|---|
| Market timing | "We're early in the AI investment cycle." |
| Stock selection | "This name compounds at 20% a year." |
| Business fundamentals | "Operating margin holds above 11% through the capex cycle." |
| Own behavior | "I'm concentrating into my highest-conviction ideas." |
Pooling them would produce a meaningless number. Someone can be excellent at judging a business and hopeless at calling the market, and you need to know which, because you're only borrowing one of those skills.
Nothing resolves itself
Predictions come due when their horizon arrives — you get a resolution queue rather than a silent expiry — and claims with no fixed date stay open until you rule on them. The system never decides that someone was right. Judging the outcome is yours; the ledger just makes sure the question comes back to you.
Scorecards that refuse to flatter
Per manager and per claim type, the scorecard shows the record. With one deliberate rule: below eight graded claims it shows counts, not percentages. "3 of 4" is honest; "75%" from four data points is a number pretending to be evidence.
Zeitgeist
Alongside the per-manager view, the zeitgeist groups open predictions by theme across everyone you follow — so you can see at a glance where the people you trust agree and where they diverge. It uses the theme tags you assigned, not inferred clustering: it reports what you told it.