How we know

Follow one Fact

1. 1

   Evidence

   A video is titled “Pablo Rodríguez y Corina Herrera Lisboa Tango Festival 2017”. The Matcher reads a city in the title: Lisbon.
2. 2

   The rule

   “The city named in a video's title” has 528 decided random checks, 95.5% of them right: past the 95% line for places, so it applies, marked unverified.
3. 3

   A check

   Picked at random, a reviewer, a person or an agent working from the video alone looks at the video and the place, and answers right, wrong or unsure. This one was right.Pablo Rodríguez y Corina Herrera Lisboa Tango Festival 2017 → Lisbon · rightWho made the 6,011 random checks so far:AI agents 6,011Reviewers 0Other people 0
4. 4

   It counts

   That answer joins the method's count. If its precision falls under the line, the method goes back to proposing, and its Facts wait for support.

From Evidence to Facts

Every Evidence row we keep, by where it came from, what the rules did with it, and the Facts it became. Widths are to scale; kinds of Fact under 0.1% of all Facts are counted in Other.

• Waiting
• Replaced by a newer reading
• Applied, unverified
• Verified
• Rejected

28,469 of 895,018 agent decisions were undone later and their earlier state restored.

Known gap: the Matcher writes most dancer credits straight onto the video, so their Evidence rows still wait here, and only 131 dancer credits show as Facts.

The precision ladder

Each way in is checked on a random sample. After 200 decided checks its precision counts: from 90% it can back up another source, from 98% it applies on its own (places and venues from 95%). Below 200 it can only propose. The line around each dot is where its true precision most likely lies.

• What kind of video, from its shapeWould apply on its own once in force99.7%
• Who sang, read by agents from catalogue pagesApplies on its own99.7%
• Who composed it, read by agents from catalogue pagesApplies on its own99.4%
• Who danced, a near-spelling of a known nameWould apply on its own once in force98.1%
• The venue named in a video's titleApplies on its own97.4%
• Which recording plays, from the audioBacks up another source97.3%
• Who danced, read from the titleWould back up another source once in force96.9%
• The city named in a video's titleApplies on its own95.5%
• What kind of video, presumed from a named songCan only propose89%
• What kind of video, from the channel's own labelCan only propose85.8%
• What kind of video, by a method nobody recordedCan only propose85.5%
• What kind of video, from a word in the titleCan only propose85.2%
• Who danced, a name found inside the titleCan only propose84.3%
• The city named in a video's descriptionCan only propose82.7%
• What kind of video, decided in agent sessionsCan only propose78.9%
• The Mundial round, from the titleStill counting · 40 of 20040/200
• Birth dates entered by an editorStill counting · 25 of 20025/200
• Chapters of a talk, from its transcriptStill counting · 17 of 20017/200
• Not tango, decided in agent sessionsStill counting · 17 of 20017/200
• Who danced with whom, from our own recordsStill counting · 16 of 20016/200
• Catalogue numbers, taken from catalogue pagesStill counting · 13 of 20013/200
• Which song, read from the titleStill counting · 12 of 20012/200
• Places sung in the lyricsStill counting · 20 of 20020/200
• Which dance of the set, from the titleStill counting · 19 of 20019/200
• Which period, from research notesStill counting · 21 of 20021/200
• Not tango, from a word in the titleStill counting · 18 of 20018/200
• Places mentioned in a textStill counting · 16 of 20016/200
• Tango, vals or milonga, from a word in the titleStill counting · 17 of 20017/200
• Tango, vals or milonga, decided in agent sessionsStill counting · 13 of 20013/200
• Which song, an exact title matchStill counting · 10 of 20010/200

“Once in force”: credits, kinds and roles are measured, but what writes them does not consult the ladder yet.

Only random checks count: 6,011 so far. Another 16,663 checks of rows chosen on purpose are kept and never counted.