=== STATISTICAL ANALYSIS SUMMARY ===

1. SELF-PREDICTION WILLINGNESS:
   100-word corpus: 5/10 models (50.0%) ever predicted themselves
   500-word corpus: 4/10 models (40.0%) ever predicted themselves
   p-value (100w): 6.23e-01 (not significant)
   p-value (500w): 3.77e-01 (not significant)

2. EXACT MODEL PREDICTION ACCURACY:
   100-word corpus: 10.3% vs 10.0% baseline
   500-word corpus: 10.9% vs 10.0% baseline
   p-value (100w): 0.752 (not significantly different)
   p-value (500w): 0.343 (not significantly different)

3. BINARY SELF-IDENTIFICATION ACCURACY:
   100-word corpus: 82.1% (significantly below 90.0% threshold)
   500-word corpus: 72.3% (significantly below 90.0% threshold)
   p-value (100w): 1.00e-02
   p-value (500w): 3.48e-07

4. GPT/CLAUDE FAMILY PREDICTION BIAS:
   Observed rate: 97.7% of all predictions
   Expected rate: 40.0% (based on uniform distribution)
   Chi-square statistic: 1387.2
   p-value: 1.27e-303 (highly significant bias)

5. 95% CONFIDENCE INTERVALS:
   Exact accuracy (100w): 10.3% (95% CI: 8.6%-12.3%)
   Exact accuracy (500w): 10.9% (95% CI: 9.1%-13.0%)
   Binary accuracy (100w): 82.1% (95% CI: 73.3%-88.3%)
   Binary accuracy (500w): 72.3% (95% CI: 62.5%-79.9%)

=== CONCLUSIONS ===
• Self-prediction rates are significantly below expected levels
• Exact model prediction accuracy is near random baseline
• Binary self-identification accuracy is significantly below 90% threshold
• Extreme and statistically significant bias toward GPT/Claude families
• All findings provide strong statistical evidence for systematic self-recognition failures