What this report is

This is a readiness + instrumentation report for TP bridge experiments. It is not a final scientific claim. It shows that the measurement pipeline is now operational.

Dashboard

TP Bridge Dashboard

Axis labels (how to read)

X-axis (all plots): training step t
Representation Drift Y-axis: relative drift from initialization (t=0)
NTK Drift Y-axis: relative Frobenius drift from initialization
Layer Update Norms Y-axis: L2 norm of per-step parameter update

What was implemented

Trace generation (real training runs):
- experiments/src/run_tp_bridge_toy_mlp_purepy.py (baseline)
- experiments/src/run_tp_bridge_residual_mlp_trace.py (residual 2-block MLP)
- experiments/src/run_tp_bridge_seq_mixer_trace.py (residual token/channel mixer)
Metrics computation (schema-preserving):
- experiments/src/tp_bridge_metrics.py
Width/LR sweep:
- experiments/src/run_tp_bridge_width_sweep_purepy.py
Chart rendering:
- experiments/src/render_tp_bridge_images.py

Key outputs

Baseline trace/metrics:
- experiments/results/tp_bridge/toy_trace_real.json
- experiments/results/tp_bridge/toy_metrics_real.json
Architecture-realistic traces/metrics:
- experiments/results/tp_bridge/arch_residual_trace_real.json
- experiments/results/tp_bridge/arch_residual_metrics_real.json
- experiments/results/tp_bridge/seq_mixer_trace_real.json
- experiments/results/tp_bridge/seq_mixer_metrics_real.json
Architecture image bundles:
- experiments/results/tp_bridge/arch_residual_images/
- experiments/results/tp_bridge/seq_mixer_images/
Sweep table: experiments/results/tp_bridge/sweep/sweep_results.md

Initial interpretation

Residual MLP (architecture-realistic): high NTK drift final (0.6203) with substantial representation drift (0.2745) and favorable transfer regret (-0.4817) indicates strong feature-learning dynamics with good small→large transfer in this run.
Sequence mixer (architecture-realistic): lower representation/NTK drift (0.0644 / 0.2914) plus small positive transfer regret (+0.0279) suggests a more stable but slightly less transfer-friendly regime.
Across both architecture runs, layer update norms remain nontrivial (not frozen), so the bridge captures genuine parameter motion beyond toy behavior.

Schema compatibility (preserved)

All new runners keep the same tp_bridge_metrics contract and summary keys:

representation_drift_*
ntk_drift_*
layer_update_norm_*
transfer_regret

No schema drift was introduced in metrics JSON.

How this fits Epimechanics

These metrics are operational proxies for representational dynamics:

representation drift → representation state motion proxy,
NTK drift → frozen-vs-restructuring regime signal,
layer updates → where representational work occurs,
transfer regret → scale-transfer efficiency signal.

This enables better falsifiable experiments before broader theory claims.

Current limitations

Current live runs are toy-level and pure Python.
Not yet architecture-realistic (no large-model framework runs yet).
Spectral diagnostics in live runs still optional.

Next step (required)

Move the same instrumentation to architecture-realistic training traces while preserving the exact metric schema and visual reporting format.

TP Bridge Report: Instrumentation, Results, and Readiness

Contribute to Epiphysics