Executive summary

This is the public progress report for the TP bridge validation track.

Goal: demonstrate corrected computational regimes with reproducible instrumentation, then escalate from toy runs to architecture-like runs while preserving metric semantics.

Current status:

✅ Metrics stack operational end-to-end
✅ Real-run traces generated (not synthetic-only)
✅ Architecture-like runs added (residual + sequence mixer variants)
⚠️ Still pure-Python and small-scale; framework-level large-model validation is next

What was built

Core pipeline artifacts:

Trace generation:
- experiments/src/run_tp_bridge_toy_mlp_purepy.py
- experiments/src/run_tp_bridge_width_sweep_purepy.py
- experiments/src/run_tp_bridge_residual_mlp_trace.py
Metrics engine:
- experiments/src/tp_bridge_metrics.py
Visualization:
- experiments/src/render_tp_bridge_images.py

Metric contract (preserved across runs)

Every valid run reports:

representation_drift
ntk_drift
layer_update_norms
transfer_regret

This contract is the backbone of cross-run comparability.

Result snapshots

Run	Model	rep_drift_final	ntk_drift_final	layer_update_norm_mean	transfer_regret	Regime read
Toy real run	tiny_mlp_purepy	1.2599	0.1307	0.0186	-0.0321	feature-learning / non-frozen
Architecture run	residual_mlp_2block_purepy	0.2745	0.6203	0.0147	-0.4817	strong non-frozen dynamics
Architecture run	seq_mixer_2block_purepy	0.0644	0.2914	0.0036	+0.0279	mixed to non-frozen, weaker transfer

Interpretation:

Non-frozen dynamics are consistently detectable.
Transfer quality is configuration-sensitive (not uniformly positive once architectures vary).
This is exactly why the regime diagnostics are necessary before strong claims.

Public Validation Report: TP Bridge Experimental Progress

Contribute to Epiphysics