P07_reflection.md

P07 — Reflection

Field	Value
ID	P07
Name	Reflection
Category	Error Identification
Origin	White et al. (2023)
PromptPrism Level	Semantic / Pragmatic
Volatility Class	II* — Context-Sensitive (Claude-Divergent)
Transfer	0.66 [.59,.72]
Restricted Transfer	0.75 [.68,.81]
PSI	0.20
Core Mechanism	Self-introspective error detection

Per-Model PTS

Model	PTS	±CI
GPT-4o (M1)	0.79	±.04
Claude 3 Opus (M2)	0.82	±.04
Gemini 1.5 Pro (M3)	0.77	±.04
LLaMA-3-70B-Instruct (M4)	0.68	±.05
Mistral-7B-Instruct (M5)	0.59	±.07
MedLLaMA-2-7B (M6)	0.52	±.08

Canonical Specification (V-A)

Answer [TASK]. Then reflect critically on your own response: What assumptions did you make? Where might you be wrong? What additional information would change your answer? Present this reflection clearly separated from your main response.

Deployment Guidance

Claude 3 Opus shows anomalously HIGH PTS (0.82) on this pattern—above GPT-4o (0.79)—due to the reflective mechanism aligning with its alignment training. However, refusal-adjacent variants trigger divergence. Apply G4 mitigation: task-specific system prompt improves Mistral-7B by +0.09. Temperature-sensitive on open-weight models (Δ=0.18 at T=1.0 vs T=0.0 on Mistral-7B).