References

Existing methodology notes

• methodology/cvae-disentangling CVAE-style learning setup for latent strategies.
• methodology/concepts/token_weighted_excess_reconstruction Teacher-forcing dilution and token-weighted reconstruction argument.
• methodology/cvae-objective-and-losses Central loss decomposition and links to the active methodology variants.

Existing theory task

• path:.tasks/doing/2026-03-15-theory-latent-space-exploration/task.md Related theory thread on latent-space exploration gains.
• path:.tasks/doing/2026-03-15-theory-latent-space-exploration/spec.md More formal treatment of latent-strategy exploration and learned-factorization robustness.

Relevant ideation notes

• ideation/concepts/abstract-strategy-vs-solution Distinguishes latent strategy variables from concrete realized solutions.
• ideation/concepts/analogy-as-strategy-transfer Connects abstract strategy reuse to cross-problem analogy.

Nearby implementation context

• path:core/disentanglement/shared_discrete_cvae.py Loss composition and latent-routing implementation.
• path:experiments/synthetic_sequences/disentangled/eval_utils.py Controlled and sampled latent evaluation utilities.
• path:experiments/synthetic_sequences/disentangled/summary_plots.py Alignment metrics and summary plotting.