Credibility leaderboard
Two scores, kept apart: Prediction measures forecasting — resolved, falsifiable calls only — and Integrity measures factual honesty — fact-checked claims about the past. They are never combined. Every receipt is archived at time of publication; select a row for the full record.
| Tracked person | Prediction score | Trend | Integrity score | Resolved | Open | Falsifiable | Repeated |
|---|
Prediction score — forecasting
Prediction score history
Integrity score — factual honesty
Falsifiability
Queue a video
Recently queued
Subscribe to a channel
Subscribed
Processing queue
Transcripts are fetched and archived locally. Extraction is a separate, unhurried step — nothing expires while an item waits. Two lanes share one worker: live (subscriptions and on-demand requests) is always extracted first, and backfill (bulk channel history) drains with whatever is left.
Scheduled resolutions
Open predictions are re-checked daily against market data; price targets resolve early the moment they hit.
| Prediction | Person | Due | Days left | Last seen | Resolver |
|---|
What counts as a prediction
The extractor reads every transcript and keeps only statements that name a subject, a measurable outcome, and a deadline. Each one is stored with a machine-checkable resolution rule — for example S&P 500 ≥ 7,700 by 2026-08-15, intraday touch, source: Yahoo Finance.
Statements with no falsifiable content are kept on record but never scored. The share of a person's statements that can be checked is reported as their falsifiability rate — chronic vagueness is itself a signal.
Repeats are tracked, not double-counted. The same call made again in a different video raises a repetition count instead of earning fresh score weight, and a recycled claim or talking point is linked to the first video that made it — checked once, at its origin. The share of a person's output that recycles their own earlier videos is published per video and per person, so a viewer can judge how much of a new upload is actually new.
A near-certainty is not a prediction. When the difficulty model prices a statement at 95% or higher implied probability at the moment it was made — for example "it'll dip below today's price at some point this year" — it is recorded as unfalsifiable rather than scheduled, with the computation stored as the reason. Scoring such a statement would hand out a Hit for noise.
How predictions resolve
- A daily worker checks every open prediction against its data source. All sources used are free and keyless: CoinGecko, Yahoo Finance, FRED.
- Price targets resolve as a hit the first day the target is met — not only at the deadline.
- A prediction that reaches its deadline unmet is a miss.
- Every check writes an audit row, so any resolution can be re-examined later.
Two scores, kept apart
Forecasting accuracy and factual accuracy answer different questions, so they are never combined into one number. Someone can be a careful forecaster who is loose with history, or scrupulous about facts and bad at calling markets — and a reader should be able to see which.
Prediction score — does this person know their game? Resolved, falsifiable forecasts only.
w = difficulty × specificity × recency
k = 3 (prior weight, shrinks small samples toward 50)
Calling a momentum continuation in a bull market carries low difficulty and barely moves a score even when right. A dated target against consensus carries high difficulty and moves it a lot. Repeated versions of the same call count once; repetition is tracked as conviction, not extra credit.
Integrity score — does this person bluff? Fact-checked claims about the past only. A verified claim counts 1, a misleading one 0.35, a false one 0.
Claims that could not be verified are excluded entirely, never counted against anyone — an absence of evidence is not evidence of dishonesty. Many statements are unverifiable by construction, because they rest on a speaker's own proprietary indicator that no third party can reproduce.
Each score is rated on its own evidence: with fewer than five resolved predictions, or fewer than five checked claims, that side reads Not yet rated rather than being given a verdict a thin record cannot support.
Receipts, not opinions
Every scored item links the original quote, the source video, the timestamp, and the resolution evidence. Transcripts are archived verbatim at ingestion — deleting a video does not delete the record. The index publishes outcomes and math; it does not editorialize about people.
Anyone tracked can flag an extraction as misread; flagged items are excluded from scoring until re-reviewed.