TLAgent-Pro

Inner loop (≈ gradient descent)

1. Try Criteria → 2. LOO Verify → 3. Inspect GT (always) → threshold? → yes: Done / no: ↓

4. Reflect → three exploration paths:

• A. Fix/Refine — debug, tweak thresholds, switch model. Different strategies on the same data source (e.g. different computation or model).
• B. Explore Explicit — exhaust hand-crafted features. Different data sources = different workflows.
• C. Try Implicit — latent features (PCA, PLS). Check if correlated with GT.

→ loop back to step 1

Plateau escape (outer loop ≈ architecture change)

• When agent loops without improvement for multiple rounds
• Human intervention — identify agent blind spots: what has the agent never tried?
• This is not a gradient update, but a structural change ≈ change optimizer / architecture
• → restart with new insight

Experiment trajectory

R1-R2: Try most intuitive criteria (gland_in_tumor) → all zeros, broken.

R3: Switch to implicit features (PCA embeddings) → r jumps 0.12 → 0.41 (breakthrough).

R4-R7: On PCA basis, exhaust explicit features (nuclear morphology, lumen, spatial overlap).

R8: Model improvement (PLS2 + threshold optimization) → r 0.44 → 0.55.

R9-R14: Exhaust ensemble, ROI features, spatial features → stuck at 21/33.

R15: Human intervention — identified agent blind spots: always used PCA compression (unsupervised), never tried L1 for supervised selection; spent 4 rounds building hand-crafted heterogeneity features, never realized embedding std IS heterogeneity.