phase2_homeostasis_Claude_analysis.md

Phase 2 Experiment Log: phase2_homeostasis (parallel)

Block 1, Batch 1: Baseline Initialization

Strategy

Baseline comparison to isolate the effect of rollout length (time_step). All 4 slots use identical training parameters.

Configuration

Slot	time_step	lr_node_homeo	lr_emb_homeo	data_aug_loop	batch_size
0	4	0.01	0.001	1000	8
1	16	0.01	0.001	1000	8
2	32	0.01	0.001	1000	8
3	64	0.01	0.001	1000	8

Rationale

• lr_node_homeo=0.01 is 10x higher than default (0.001) to help MLP_node escape zero initialization
• lr_emb_homeo=0.001 is lower because MLP_node must learn first before embeddings can separate
• Same params across all slots isolates the effect of time_step (rollout length)
• Longer rollouts accumulate more homeostatic signal but have noisier gradients

Key Questions for This Batch

1. Which rollout length (4, 16, 32, 64) produces the best phase2_score?
2. Does lr_node_homeo=0.01 help MLP_node escape zero initialization (phase2_node_magnitude > 0)?
3. Do any slopes move toward GT values (-0.001, -0.002)?

---