CCDD combines rubric-based adjudication, transcript processing, structural alignment and provenance-preserving organisation, so that each debate contribution is connected to an identifiable scoring target and to individual judgments.
Three design objectives guide the resource: preserve interaction context, identify the target of each assessment, and retain individual judgments with their provenance. Ordered match text preserves the context of a contribution; side and role associations identify what is assessed; individual records permit analysis of variation among judges.
A match is one complete debate between affirmative and negative sides (4v4 or 2v2). A stage is a scheduled segment with a procedural type; separate constructive speeches are distinct stages, a shared head-to-head segment is one stage. A stage–side instance associates a stage with one evaluated side and its three original judge scores: a speech stage contributes one, an interactive stage with assessments of both sides contributes two. A turn is one continuous contribution from one side; transfer of the floor starts another.
Senior debate experts developed the rubric before the competitions. Its three criteria are Task Performance (fulfilment of the obligation of the evaluated stage and role), Battlefield Judgment (appropriateness and importance of the point of contention engaged) and Degree of Advancement (progress achieved at that point of contention). Together they guide one overall integer score from 1 to 10 per evaluated side and judge; the full profiles, including two profiles for score 6, are on the judges and rubric page.
Three judges from a pool of 157 registered adjudicators (129 of whom adjudicated the retained matches) assessed each match and submitted scores and votes independently after the match and before public commentary. Competition-time adjudication denotes judgments collected as part of this procedure; individual submissions preserve variation within each panel.
Each judge casts an impression vote, a stage-based vote (following that judge's comparison of the two sides' stage scores, which may be split equally) and a decisive vote.
The components are consolidated within the judge: the side with the larger total receives the decision, and the decisive vote resolves a 1.5–1.5 tie.
The majority of the three judge decisions determines the winner. Each judge separately allocates three best-debater votes, giving nine across the panel.
| Step | Operation |
|---|---|
| Collection | The organising committee collected recordings, ASR transcripts, scoring forms and supplementary notes from 182 matches in six competitions (May 2 – August 20, 2026) and authorised research use and redistribution. |
| Transcript correction | Tencent Meeting ASR output corrected under a constrained prompt (recognition errors and sentence segmentation only; no rewriting). Corrected units keep src line references to the raw ASR; debate speech is separated from judge commentary. |
| Stage and turn recovery | Moderator cues and timestamps recover stage boundaries; stage types record the procedural function; changes of the speaking floor determine turns. |
| Score alignment | Scoring-form entries are aligned with their scheduled stage and evaluated side, so shared interactive text supports distinct score vectors for the two sides. |
| Manual review | All retained text entries were reviewed in a dedicated interface for transcription and stage segmentation: 304 transcription errors corrected and boundaries adjusted in 71 stages. No residual-error estimate against an independent transcription is claimed. |
| Filtering | 34 matches excluded (forfeits, late-arrival losses, incomplete score sheets, recording failures). Validation removed eight invalid scoring rows, including opposing-side entries for single-side speech stages, from 2,698 rows: 2,690 valid stage–side instances remain. |
| Appeals | 22 appeal reassessments exist in the source; 20 are linked to retained matches and stored separately from the original judgments without increasing the number of unique matches. |
| Competition | Matches | Dates (2026) | Format | Teams | Debaters | Judges |
|---|---|---|---|---|---|---|
| Competition A | 36 | May 2–31 | D–P | 40 | 214 | 50 |
| Competition B | 9 | May 12–Jun 13 | D–P | 13 | 59 | 23 |
| Competition C | 24 | May 16–Jun 16 | S–Z | 26 | 152 | 51 |
| Competition D | 22 | Jun 13–Jul 5 | S–Z | 24 | 133 | 47 |
| Competition E | 32 | Jul 10–Aug 12 | 2v2 | 36 | 68 | 42 |
| Competition F | 25 | Jul 18–Aug 20 | S–Z | 30 | 152 | 47 |
| Sum across competitions | 148 | May 2 – Aug 20 | — | 169 | 778 | 260 |
| Distinct in retained records | 148 | — | — | 158 | 629 | 129 |
Competition names are withheld for anonymous review (the public package labels them C1–C6, in a different order). S–Z denotes 4v4 Shenlun–Zhixun (13 stages, no head-to-head); D–P denotes 4v4 Duibian–Panwen (14 stages, one head-to-head). Column sums include repeat appearances across competitions: 150 teams occur in one competition, five in two and three in three; 510 debaters occur in one, 94 in two, 20 in three and five in four. Teams appear in the package as T-identifiers.
Generated summaries and reasoning accounts are separately identified supplements produced from the recorded commentary (the commentary is the source; debate text only clarifies references; the outcome preference is supplied as input). Thirty-one experts screened them for severely unreasonable content; two records were rejected and regenerated. They are judgment-conditioned supplements, not observations of a judge's reasoning, and are kept apart from the human records.