Daily Decision Stack

Hypothesis: if metric contract is preserved across complexity steps, we can detect regime transitions reproducibly.
Success criteria:
- all four metrics reported in each run,
- at least one interpretable regime shift,
- public-facing report artifact produced.
Failure criteria:
- missing metrics,
- contradictory/unstable signals with no interpretable cause,
- no publishable artifact.

Progression executed today:

Core metric contract was preserved in all runs:

Metrics snapshot

Run	rep_drift_final	ntk_drift_final	layer_update_norm_mean	transfer_regret
Toy real run	1.2599	0.1307	0.0186	-0.0321
Residual architecture run	0.2745	0.6203	0.0147	-0.4817
Seq-mixer architecture run	0.0644	0.2914	0.0036	+0.0279

Mixed (productive).

Positive: instrumentation stayed consistent while model complexity increased.
Positive: non-frozen dynamics are detectable across settings.
Caution: transfer quality is not uniformly positive once architecture changes.
Decision: promote instrumentation pipeline, iterate on architecture-realistic validation.

Commits: d62ad21, 193afab, 92a872c, 31dbd99, f95abd9
Report pages:
- docs/experiments/tp_bridge_report.md
- docs/experiments/tp_bridge_public_validation_report.md

We now have a reliable way to test optimization behavior day by day:

That’s exactly the kind of compounding understanding loop we want.

TP Bridge Dashboard

Architecture Representation Drift

Architecture NTK Drift

Run a framework-backed experiment (non-purepy) with the same metric schema and compare against today’s architecture-like purepy baseline.

Success criterion: