Update GUIDE.md

GUIDE.md

@@ -59,6 +59,9 @@ In QML, the **structure ("shape") of a circuit directly impacts performance**.
   The most critical complexity metric.  
   On NISQ devices, CNOT gates are **10x–100x noisier** than single-qubit gates.
 
+**Note on Feature Correlation:** While structural metrics (like `gate_entropy` or `depth`) describe the complexity of the circuit, they do not encode the specific rotation angles of individual gates. 
+Therefore, predicting the exact expectation value using only structural features is an **extremely challenging task** (Non-Trivial Mapping).
+
 ---
 
 ## 🎯 3. Multi-Target Regression (The Bloch Vector)
@@ -67,6 +70,15 @@ Unlike traditional benchmarks that focus on a single observable, QSBench targets
 
 [⟨X⟩global, ⟨Y⟩global, ⟨Z⟩global]
 
+```text
+       | +Z (0)
+       |
+  -----|---- +Y
+      /|
+     / | -Z (1)
+   +X
+```
+
 ---
 
 ### Why predict all three?
@@ -109,6 +121,16 @@ The Hub uses a **Random Forest Regressor** to establish a baseline of predictabi
      - X/Y depend more on circuit structure  
      - reveals architectural biases (HEA vs QFT, etc.)
 
+📉 **How to Interpret "Bad" Metrics?**
+
+If you see a **negative** R2 or clustering around zero, don't panic. This is the expected behavior for standard regression on quantum data:
+
+- **Mean Predictor Baseline:** In complex circuits (n=8, depth=6), expectation values naturally concentrate around 0. A model that simply predicts "0" for everything will have a low MAE but a zero/negative R2.
+
+- **The Complexity Gap:** A negative R2 proves that the relationship between circuit shape and quantum output is highly non-linear.
+
+- **Research Challenge:** Use these baseline results to justify the need for more advanced architectures like **Graph Neural Networks (GNNs)** or **Recursive Quantum Filters** that can process the gate sequence itself.
+
 ---
 
 ## 🔗 5. Project Resources