Haqora
Method

How Haqai weighs an argument

Haqai is Haqora's judge, built on Anthropic's Claude. It weighs reasoning and evidence, holds every side to one rubric, and never rules on whose values are right. This page is its complete method.

A 50Haqai50 B
  1. § 1

    Rules first

    Before anything reaches Haqai, the format applies: whose turn it is, length limits, and that a rebuttal must answer a specific proposition.

  2. § 2

    Classification

    Haqai classifies each post as an argument, evidence, a clarification or a concession (substantive), or as dismissive, off topic or abusive (not substantive).

  3. § 3

    Scores

    Substantive posts are scored 0 to 10 for logic, evidence, relevance and civility, each with a short reason.

  4. § 4

    The gloss

    Fallacies are flagged only when Haqai can quote the words that commit them, and each is marked minor or major.

  5. § 5

    The record

    A proposition stays open until the other side answers it substantively. Dismissals never close a point. At the close, open points stand.

  6. § 6

    The ruling

    When the floor closes, Haqai reads the whole record and rules: a finding or a tie, scores on four criteria, the decisive clashes, and advice to each side.

The weights

How a proposition's quality is calculated from its scores.

quality  = 0.4 × logic + 0.3 × evidence + 0.3 × relevance
         − 0.5 per minor fallacy
         − 1.5 per major fallacy
         − 3.0 for addressing the judge instead of the opponent
         − 0.5 per point of civility below 4
         = 0   if the post is not substantive

side     = Σ quality of its propositions
         + 1.0 × quality of points the opponent conceded
         + 0.5 × quality of points left standing at the close

The fallacies

18 named patterns. Haqai flags one only when it can quote it.

01
Ad hominem
Attacks the person making the argument instead of the argument itself.
02
Straw man
Rebuts a distorted or weaker version of the opponent's actual claim.
03
False dilemma
Presents two options as the only possibilities when others exist.
04
Slippery slope
Claims one step will inevitably lead to an extreme outcome without showing the chain.
05
Appeal to authority
Treats a source's status as proof, outside its expertise or without the evidence.
06
Appeal to emotion
Uses feelings in place of reasons to establish that a claim is true.
07
Appeal to popularity
Argues a claim is true because many people believe it.
08
Circular reasoning
The conclusion is assumed in one of the premises.
09
Hasty generalisation
Draws a broad conclusion from too few or unrepresentative cases.
10
False cause
Assumes that because one thing followed or correlates with another, it caused it.
11
Whataboutism
Deflects a point by raising a different wrong instead of answering it.
12
Moving the goalposts
Changes the standard of proof after the original one was met.
13
Shifting the burden of proof
Demands the opponent disprove a claim instead of supporting it.
14
Equivocation
Uses one word in two different senses to make the argument appear to work.
15
Cherry picking
Cites only the evidence that supports the claim and ignores what contradicts it.
16
Non sequitur
The conclusion does not follow from the premises given.
17
Tu quoque
Dismisses a criticism by pointing out the critic's own inconsistency.
18
Judge manipulation
Text aimed at the AI judge rather than the opponent, such as instructions to score a certain way.

Limits and appeals

What Haqai can't do, and what happens when it gets it wrong.

  • It judges what is written in the debate. It doesn't browse sources, so it can't confirm a citation says what a debater claims unless they quote it.
  • It can misread sarcasm, cultural references and specialist detail. Every reading shows its reasons so you can see where it went wrong.
  • Debaters' text is material to judge, never instructions. Attempts to instruct the judge are flagged as judge manipulation and penalised.
  • On some sensitive topics Haqai may decline to assess a post. Nobody is penalised, and the point counts as answered.
  • Featured debates are ruled three times independently; the ruling shows how often the readings agreed.
  • Think a ruling is wrong? Report it. People review reports, and the model and prompt version are stored with every judgment.