MOVICI Article 3 reproducibility package
Data audit

Canonical sources
- Participant data derived from MOVICI_Canonical_Dataset_OSF_2026-09-06_FINAL/data/MOVICI_Canonical_Dataset_96.csv.
- Dyad data derived from the CANONICAL_ONLY_DYADS sheet of MOVICI_Evaluation_Matching.xlsx.

Participant table
- 96 participant rows.
- 34 debate clusters.
- 47 Competition and 49 Cooperation participants.
- Missing family SES: 1.
- Missing mean Web-source expertise: 4.
- Missing mean Web-source benevolence: 4.
- T1/T2/T3 factual and subjective outcomes are complete.
- Canonical attitude total equals SSA + VAHC + VASA + EP at T1, T2 and T3.
- Canonical theoretical stored-score ranges used for QC/diagnostics: SSA -18 to -2; VAHC +2 to +18; VASA -18 to -2; EP -8 to +8; total -42 to +22.

Dyad table
- 212 ordered evaluator-to-evaluatee rows.
- 91 evaluators across 29 debate clusters.
- No self-evaluation dyads.
- Missing perceived expertise: 16.
- Missing perceived benevolence: 0.
- Evaluator/evaluatee T2 factual scores reproduce the linked participant-table scores.

Family SES provenance
- For each participant, parental education is averaged across available parents.
- Parental social position is averaged across available parents.
- Parental economic resources are averaged across available parents.
- Each of these three components is z-standardized across the canonical 96-person sample.
- The family SES score is the mean of available standardized components when at least two are present.
- Its distribution exactly reproduces the Article 1 public-package family_ses distribution (1 missing; min -1.459656; max 1.777368).

Data minimization
- Released variables are restricted to Article 3 H1-H3, complementary C1, and RQ1.
- Exact dates/times, workstation identifiers, raw demographic variables, raw transcripts, URLs, free-text data, SDO, homophily variables, dispositional profiles, and debate-behavior variables are excluded.
- Article 3 uses an independent public pseudonymization to prevent cross-package linkage with Articles 1 and 2.

Analysis safeguards
- The R script stops before inference if row counts, cluster counts, condition counts, ID linkage, knowledge ranges, subjective-knowledge ranges, canonical attitude total identities/ranges, or dyad linkages fail.
- Participant-level inference uses CR2 cluster-robust standard errors by pseudonymized debate cluster.
- Paired variance changes use the Pitman-Morgan test.
- Direct variance/convergence contrasts use debate-cluster bootstrap resampling.
- Attitude-subscale and Web-credit-component decompositions use Benjamini-Hochberg correction within their stated families.

Equivalence-sensitivity additions (v2)
- RQ1 partner-expertise recognition is additionally estimated on a standardized outcome and evaluated with CR2/Satterthwaite TOST sensitivity bounds of +/-0.10, +/-0.20 and +/-0.30 SD.
- RQ1 mean-partner-knowledge -> T3 is additionally estimated on a standardized T3 outcome and evaluated with the same bounds.
- The +/-0.20 bound is a focal small-effect benchmark only; equivalence testing was added after initial analyses and is not preregistered.
- H2 cross-outcome change couplings receive the same standardized equivalence-sensitivity grid as a secondary diagnostic.
- T2->T3 is documented as the full debate -> partner evaluation -> argument recall/evaluation -> T3 sequence, not debate alone.
