refactor(module-3): honest diagnostic framework - validated claims only, limitations explicit

CRITICAL HONESTY IMPROVEMENTS (following Modules 1-2 pattern):

1. RELATED WORK EXPANDED (Section 5) - Added 18 citations (9→27 total):
   - 5.1 Assumption Diagnostics (Stuart, Austin, Ho, Petersen, VanderWeele, Hainmueller)
   - 5.2 Federated Causal Inference (FACE, FLAME, FedCI, Duan, Chang)
   - 5.3 Adaptive Inference Methods (TMLE, meta-learners, quasi-oracle, Manski)
   - 5.4 Robustness and Sensitivity (Rosenbaum, VanderWeele, Cinelli, Carnegie)

2. THRESHOLD JUSTIFICATION HONEST (Section 2.1):
   - Changed: "Mode Selection Rules" → "Proposed Mode Selection Rules (Heuristic, Requiring Validation)"
   - Added honest caveat: "0.8 cutoff builds on established heuristics (SMD<0.1, tail mass<5%)"
   - Explicit: "Combination into overall scores lacks formal validation"
   - Stuart (2010): SMD<0.1 is 'rule of thumb' without power analysis
   - Requirements: "Controlled simulation studies with known ground truth violations"

3. PRIVACY CLAIMS DOWNGRADED (Section 3.5):
   - Changed: "0% covariate disclosure" → "Partial covariate privacy"
   - Added: "Information leakage caveat" - scores reveal data quality characteristics
   - Example: Unconf=0.70 suggests imbalance → indirect information
   - Honest: "Variable identity privacy is partial, not absolute"
   - Formal DP (Lap(Δ/ε)): "Future work" requiring utility-privacy calibration

4. NATURAL ACADEMIC PROSE:
   - Removed: "Key Observations:", "Key Findings:", "Key contributions:"
   - Converted numbered lists to flowing paragraphs
   - Changed from bullet points to "First... Second... Third..." structure

5. LIMITATIONS COMPLETELY REWRITTEN (Section 4.4):
   Before: 5 bullet points, defensive tone
   After: 6 detailed paragraphs with honest assessment:
   - Threshold calibration: "We do not know diagnostic power (sens/spec)"
   - Synthetic data: "Cannot establish diagnostic accuracy for real violations"
   - Ground truth missing: "Critical validation absent - controlled violation injection needed"
   - Information leakage: "Not quantified" - adversarial inference studies required
   - Three-site limitation: "Limited evidence for large network scalability"
   - Retrospective only: "Provides no guidance for prospective study design"

6. ABSTRACT REWRITTEN:
   - Removed: "complete covariate privacy", "explicit safeguards validated"
   - Changed to: "partial covariate privacy", "heuristic safeguards proposed"
   - Added: "Critical limitations: (1) thresholds lack calibration, (2) synthetic data cannot validate real violations, (3) information leakage unquantified"
   - Final: "Operational tooling requiring substantial validation before high-stakes deployment"

7. CONCLUSIONS COMPLETELY REWRITTEN (Section 6):
   Before: "First diagnostic framework", "explicit safeguards prevent overconfidence", "complete privacy"
   After: "Proposes framework", "heuristic safeguards (not validated)", "partial privacy"

   Honest assessment paragraph:
   - "Contribution = systematizing existing diagnostics into federated workflow"
   - "NOT developing novel diagnostic theory or proving formal guarantees"
   - "Provides heuristic guidance, requires substantial validation"
   - "Cannot deploy in high-stakes clinical/regulatory contexts without validation"

8. CRITICAL VALIDATION NEEDS EXPLICIT:
   Future work must:
   - Controlled simulation with known violations (bias factors 1.5, 2.0, 3.0)
   - Assess diagnostic sensitivity/specificity
   - Calibrate thresholds to achieve desired performance
   - Quantify false positive/negative rates
   - Measure information leakage rigorously
   - Validate on real multi-site data (not just Synthea)

Changes prioritize academic honesty over claims:
- No "validated safeguards" where validation is absent
- No "0% privacy" where information leakage exists
- No "formal thresholds" where heuristics exist
- Honest about synthetic data limitations (cannot validate real violations)
- Honest about contribution scope (operational tooling, not theory)

Author: watilde