{
  "p1-dimensions": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. Right of zero favors the skeptical side. These are area scores before weighting, not additive contributions.",
  "p1-checks": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. Parentheses count comparisons. Rows overlap and are not independent replications.",
  "p1-decomposition": "Bar lengths show additive weighted contributions, not independent causes. The small rounding residual and any burden adjustment reconcile their sum with the overall gap; both are reported in the text.",
  "p2-topics": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. Counts are in parentheses. These stable research groups are not the current browsing categories.",
  "p2-moves": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. Each side’s qualifying moves are averaged before comparing sides. Rows overlap; these are not official weighted overall scores.",
  "p2-order": "Each bar counts calculations in which the topic has the highest gap. This is a stability exercise, not the probability that it is truly the hardest topic in the wider world.",
  "p3-direct-rates": "Rust is theist; teal is non-theist. Each debate receives equal weight after allowing for speech length. Both panels start at zero on the same scale. Protected slogans are a subset of unsupported slogans; do not add them. These dated counts have not been extended to new transcripts.",
  "p3-direct-uncertainty": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. The unsupported-slogan range crosses zero; the protected-slogan range does not. Both use the September direct annotations, not new scorecard judgments.",
  "p3-proxy": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. A flag requires all three score thresholds on the same move. A low score does not demonstrate a criticism-blocking slogan. Never combine these percentages with the transcript word rates.",
  "p4-estimates": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population. Right of zero favors CON. The weighted same-speaker range is now above zero, but no comparison randomly assigns roles, opponents or topics.",
  "p4-strata": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population.",
  "p5-cohorts": "Fractions count lower-scoring sides without a fallacy label over all decisive results in each group. The dark total includes the other groups. No tag does not mean no error. Counts are assessments, including the disclosed shared-video pair.",
  "p5-dimensions": "Each bar counts losses with no named-fallacy label that trail in exactly the stated number of weighted scoring areas. Bars add to the eligible total; these are not independent deficiencies.",
  "p5-labels": "A move may carry several labels, so label instances exceed tagged moves. The six names are the accepted inventory in this snapshot, not an exhaustive catalogue of reasoning errors.",
  "p6-bridge": "The horizontal coordinate is the earlier average and the vertical coordinate the later average. Below the diagonal means lower later. Both axes enlarge the observed score range. These are different performances, not the same debate judged twice.",
  "p6-levels": "Dots show averages; lines contain the middle 95% from 20,000 repeated draws of the stated records. Zero means no average difference. These ranges describe sensitivity to the saved observations, not all sources of judging error or a representative population.",
  "p6-tags": "A move counts once if it has any fallacy or bias tag. Bar length is the percentage of all moves in that process. Different annotation rates can reflect procedure as well as different arguments.",
  "p6-score-use": "Teal is earlier; rust is later. Each bar shows the share of moves assigned that exact area score. Both panels use identical axes. More distinct marks mean finer numerical resolution, not necessarily better judgments.",
  "p7-ranges": "Dark dots are displayed places. Teal lines contain the middle 95% of ranks obtained by drawing again from recorded scores. Dashed rust lines use a model allowing variation estimated across the group. Parentheses count appearances. Neither line is a future-win probability.",
  "p7-samples": "Gray bars are below the three-appearance research threshold; teal bars enter the fixed field. A small sample is not a poor score.",
  "p7-counts-width": "Each dot is a person. Height is the upper rank endpoint minus the lower: 2–9 has width 7. Lower means a narrower calculated range, not a lower performance score. Rank density and score variation matter as well as sample size."
}