#05–08 cross-model replication — bootstrap Δ(A4 − A1)
label = is_correct (bool). Perturbation features (10) vs no-pert baseline.
NOTE: Cohere/Mistral/NIM-slow datasets lack verbal_logprob_gap and verbalized_confidence.

model                             n    pos      A4      A1      A2  solo_frag         Δ                  95% CI       p
Cerebras llama-3.1-8b           382  0.393  0.5507  0.6001  0.4576     0.5334   -0.0495  [-0.0924,-0.0096]   0.012
NIM llama-3.1-8b                100  0.380  0.6545  0.6621  0.5225     0.4346   -0.0076  [-0.1226,+0.1056]   0.889
Cohere Command-A                100  0.210  0.3448  0.3460  0.3731     0.5172   -0.0012  [-0.1360,+0.1327]   0.978
Mistral-Small                   100  0.370  0.5903  0.5663  0.4702     0.5624   +0.0240  [-0.0798,+0.1304]   0.648

== Aggregate across models ==
  Δ  mean = -0.0086   std = 0.0264   range [-0.0495, +0.0240]
  solo fragility AUC  mean = 0.5119   std = 0.0475
  models where Δ < 0 (pert hurts): 3 / 4
  models where Δ > 0 (pert helps): 1 / 4
  models where p < 0.05:          1 / 4

== Prior-doubt reminders ==
  * is_correct judgement source differs by model (single LLM vs string match).
  * Cohere/Mistral/NIM n=100 each; CIs are wide.
  * No verbal_logprob_gap on 3/4 datasets ⇒ A1 is weaker ⇒ Δ appears larger.
  * Broken data files were excluded; see audit_power_and_dataquality_report.txt.
