How precision is counted
The method note for the accuracy ledger.
Items are drawn at random, and no percentage is
published until 50 samples have resolved verdicts.
Right now: 0 of 50 resolved random samples; 0 drawn, 0 unresolved. No percentage is shown yet.
Right now: 0 of 50 resolved random samples; 0 drawn, 0 unresolved. No percentage is shown yet.
HOW AN ITEM IS CHOSEN
The review queue draws uniformly at random from every real item
published in the trailing 30 days that has not already been randomly
reviewed. Sample output is excluded. Nobody picks which item comes next — not the reviewer,
not the builder.
This is the whole point. Before it, all seven reviews on the record were of items somebody already suspected were wrong, because those are the ones you go looking at. A percentage over those measures which items a human chose to open, not how often RADAR is right — and the opposite bias is just as easy to produce by reviewing the items that look solid.
This is the whole point. Before it, all seven reviews on the record were of items somebody already suspected were wrong, because those are the ones you go looking at. A percentage over those measures which items a human chose to open, not how often RADAR is right — and the opposite bias is just as easy to produce by reviewing the items that look solid.
WHAT COUNTS AS CONFIRMED
Confirmedcounts in the numerator
A named human opened the item, followed it to the primary source, and found the change real and
described accurately.
False positivecounts in the denominator only
The change did not happen, did not matter, or the write-up went beyond the evidence.
Unresolvedcounted in neither
The reviewer could not tell from the available evidence. Excluded from both halves of the
fraction rather than quietly scored as a success.
WHY TARGETED REVIEWS ARE EXCLUDED
Targeted review still happens and is still recorded — it is how a suspected error gets
confirmed before a correction is written. Those verdicts stay in the ledger and they are how
corrections get made. They are simply not evidence about
accuracy, because the item was selected precisely for being doubtful.
Every verdict records which kind it was, so the two can never be quietly blended.
Every verdict records which kind it was, so the two can never be quietly blended.
WHAT THIS NUMBER IS NOT
It is reviewed by one person, who also built the system. There
is no independent reviewer. A builder grading their own work is the weakest form of this
evidence, and random sampling fixes the selection problem without fixing that one.
It covers the trailing 30 days, so it moves as items age out. It says nothing about what RADAR missed — precision is not recall, and nothing here measures recall.
Treat it as an honest count kept in the open, not as an audit.
It covers the trailing 30 days, so it moves as items age out. It says nothing about what RADAR missed — precision is not recall, and nothing here measures recall.
Treat it as an honest count kept in the open, not as an audit.
CORRECTIONS
A corrected item keeps its headline and its original text, with the correction stamped on it,
and every correction records a cause and how long the claim
stood before it was fixed. Removing the claim would leave a reader who acted on it no way to
discover they were wrong.