Synthetic Sequences Experiment Lines

The synthetic-sequences setting is not one single run family. It is a collection of related experiment lines that probe different parts of the overall disentanglement question.

Main Branches

The major branches are:

• entangled baseline training, which establishes the mixed strategy source model,
• direct-pretrained disentangling, which asks whether a pretrained model can be factorized without first creating a synthetic entangled source checkpoint,
• entangled-initialized disentangling, which asks whether starting from a synthetic entangled checkpoint yields cleaner latent strategy structure than direct-pretrained initialization.

Latent-Type Branches

Within disentangling, the main methodological split is between:

• discrete latent experiments, which emphasize explicit latent-token control and controlled generation metrics,
• continuous latent experiments, which emphasize latent geometry, sampled-latent diagnostics, and posterior-vs-router comparisons.

Experimental Progression

The overall progression of the setting has followed a recurring pattern:

1. establish a baseline training pipeline for the task family,
2. run initial disentangling sweeps to identify obvious failure modes,
3. add targeted objective terms or diagnostics to isolate those failures,
4. compare direct-pretrained and entangled-initialized variants,
5. refine discrete and continuous lines separately as their failure modes diverge.

Boundary With Report Collections

This note tracks the stable experimental branches of the setting, not the per-collection narrative. Collection-specific design notes, figures, and result interpretation remain in the separate experiment-reports/ area.