How a trivia answer gets judged.
Every answer runs down the same ladder until something decides it. Most answers never reach the language model: the answer key settles them. The model only sees what the key can’t, and it has to say why.
The ladder.
Steps 1 to 6 are plain code working from the answer key. Step 7, in yellow, is the only one that uses the model.
- 1Clean it up
Strip filler (“um, I think it's”), read numbers and units, fix capitals and spacing.
- 2Exact or alias
Compare against the answer and every accepted alias in the key.
- 3Sounds like
Spelling and pronunciation matching, allowed only when exactly one answer in the key is close.
- 4Specific enough?
If it fits more than one thing, ask for more by voice, or apply the house rule on slips.
- 5Known wrong
Check the list of common wrong answers, so a near miss is ruled wrong with a reason.
- 6Hedges
Two answers given? The house rule decides: first, last, or neither.
- 7Ask the model
Only if nothing above decided it: the model rules on meaning, with the key as context, and gives a reason.
- 8Put it on record
The ruling, the step that decided it, and the answer-key version, for any challenge.
Anatomy of a question.
- Question
- Who wrote the novel Frankenstein?
- Answer
- Mary Shelley
- Also accepted
- Mary Wollstonecraft Shelley; Mary W. Shelley; M. Shelley
- Ask for more
- “Shelley” could mean her husband, Percy
- Known wrong
- Percy Bysshe Shelley (wrote the 1818 preface); Bram Stoker (Dracula); Victor Frankenstein (the character)
- Source
- [CITATION]
- As of
- Stable fact, no expiry
- Difficulty
- b = −0.4, a little easier than average [EXAMPLE VALUE]
- Sharpness
- a = 1.3, separates stronger and weaker players well [EXAMPLE VALUE]
- Bias check
- No age or region skew detected [EXAMPLE]
- Exposure
- Seen by [N] tables; leaves ranked play at [LIMIT]
- Key version
- 3, after a winning challenge added “M. Shelley”
Answers come in three lists.
Accepted, ask-for-more, and known wrong. The last list is what stops a confident near miss from being waved through.
Every fact has a source.
If a question can't point to where its answer comes from, it doesn't go into ranked play.
Difficulty is measured.
The two numbers come from a standard test-scoring model and are updated from real play, not set by the writer.
The key has a version number.
A challenge that wins becomes a proposed alias. An editor approves it, and the version goes up.
Every pub quiz argument, settled.
These are the disputes that eat quiz nights. Each has a default ruling and a house-rule dial, set before play and shown on screen.
| The argument | Example | How Zebra rules |
|---|---|---|
| Spelling | “Chaikovsky” for Tchaikovsky | Accepted when it sounds the same and only one answer in the key is close. |
| Too broad | “Roosevelt” | By voice, it asks for a first name. On slips, the house rule decides; the default is wrong when two answers fit. |
| Surname only | “Lincoln” | Accepted when only one well-known person fits the question. |
| Right name, wrong person | “Franklin Roosevelt Jr.” | Wrong. Extra detail that changes who you mean is still wrong. |
| Two answers | “Paris or Lyon” | House rule: first, last, or neither. The default is first on slips and last-before-lock by voice. |
| Units | “About a thousand feet” for an answer in meters | Converted first, then scored against the tolerance written into the question. |
| Old names | “Peking” for Beijing | Accepted, unless the question asks for the current name. |
| Facts that changed | Moon counts, records, office holders | Every question carries an as-of date. If the fact changed since, both answers count and the question goes back to an editor. |
| A bad question | “Tallest mountain”: Everest or Mauna Kea? | Flagged as ambiguous. Both answers count, and the question leaves ranked play until it's rewritten. |
Where the questions come from.
- 1Source
Facts from licensed reference data, public knowledge bases, and staff writers. Every fact keeps its citation.
- 2Draft
The model writes the question and a first answer key: aliases, answers that are too broad, likely wrong answers.
- 3Verify
A separate check against the source. Published packs also go past a human editor.
- 4Calibrate
Starting difficulty is the model's estimate. Real difficulty replaces it as tables play.
- 5Play
With a household's opt-in, only counts come back: right or wrong per question. Never audio or transcripts.
- 6Learn
Winning challenges become proposed aliases. Questions that split by age or region at equal skill get flagged.
- 7Retire
Questions seen too often leave ranked play, so memorizing the bank doesn't win a league.
- The one rule
Facts never come from the model’s memory. The model writes and grades; the fact bank decides what’s true.