Imported from alex-zhilkin/DL-course-project (AGENTS.md). Install upstream with npx skills add alex-zhilkin/DL-course-project. Copyright stays with the author.

Repository instructions

• In notebooks, display plots inline by default. Do not save figures as PNG, PDF, SVG, or any other image/file format unless the user explicitly asks for saved figure files.

• Keep notebook Markdown minimal and paper-clean. A concise title or necessary section heading is fine, but do not add speculative interpretations, narrative explanations, result claims, or instructional prose unless the user explicitly requests that text.

• Keep the plot stsyle coherent, using "Editorial" colours from the common theme of the repo

• When new feature is added to training or inference or what ever, don't automateically opt in to do it as an conditioned on option. Just implement it as is, unless it's already part of something conditional or you think it's really critical, at that point just ask me.

• For work on 07b_mixed_reid_depablo_lj_context_ablation.ipynb, read and update notebooks/latent_space/07b_experiment_log.md. Record completed experiments with their exact recipe and source-wise held-out rollout results; do not rely on pooled metrics alone.

• Prefer broader, hypothesis-driven experiment matrices and multiple seeds when they can answer an open research question. The user wants to use the available cluster capacity for many useful tests; do not default to a minimal run merely to conserve compute. Avoid duplicating completed runs unless replication or a controlled comparison warrants it.

• Submit independent single-node jobs to mendels_q with shared placement (place=free:shared), no host pins, and measured CPU/memory requests. Do not reserve entire nodes exclusively for ordinary experiments. Available cluster capacity does not require multi-agent delegation.

• Treat latent dimensionality as a hypothesis to test. Include matched 2D versus 4D comparisons when relevant; do not assume 4D is necessary. Prefer the smaller representation when source-wise validation performance supports it.

• Persist every experiment's exact recipe, split identities, seeds, code version/hashes, job ID, completion/failure status, and source-wise metrics with valid/total counts. Preserve negative results and interrupted attempts; distinguish validation from final test and avoid pooled-only conclusions. Save machine-readable results and concise experiment-log entries so future work can reuse them.

• Before proposing or launching research experiments, consult notebooks/latent_space/experiment_results_index.md and the relevant experiment logs/results. Keep the original scientific objective, successful historical recipes, failures, confounds, and pending runs in view. Update the index and logs after collecting results; use accumulated evidence to choose the next tests rather than restarting the search from memory.

• Scientific requirement: learn network dynamics from observed states/trajectories and network structure alone. Do not inject p-ratio, strain, or other expert-derived observables, slow-mode labels, physics-informed losses, or handcrafted expert targets into model inputs, training objectives, or checkpoint/model selection. P-ratio is a candidate emergent slow observable to evaluate after fitting, not a quantity to teach the model. Keep such observables strictly as post-training diagnostics. Historical response-selected results must be identified as incompatible with this requirement, not promoted as compliant baselines.

• Test whether noisy-LJ needs more latent capacity, including matched 6D and 8D comparisons against 2D/4D. Do not assume all sources require the same minimal dimension or that expert supervision is needed when a small AE struggles.