Baselines#

Local implemented baselines:

  • Linear/ridge sanity baseline with finite-output checks and stable least-squares fallback.

  • MLP baseline.

  • TCN/Conv1D baseline.

  • Transformer baseline.

  • SSM fallback baseline.

  • NeuroTwin translator.

  • NeuroTwin NFC synthetic suite baselines and ablations.

  • Pair-Operator baseline/ablation for low-rank relational updates inside the NFC question.

Prepared benchmark coverage:

  • Future-state forecasting: local baselines and NeuroTwin are ranked on identical prepared windows.

  • Masked neural reconstruction: local baselines and NeuroTwin are ranked on identical prepared masks.

  • Cross-modal translation: local baselines and NeuroTwin are ranked where paired modalities exist.

  • Stimulus-to-fMRI response: tribe_style is a local toy clean-room baseline for task plumbing until real pretrained stimulus features are integrated.

  • Supervised prepared-task artifacts include bootstrap confidence intervals for MSE and MAE.

  • Few-shot subject adaptation: reported as an auxiliary support/query sanity metric, not a full adaptation claim.

  • Adaptation reports fixed support sizes when enough held-out query windows exist; query/test labels are never used for adapter fitting.

  • Dataset/site generalization: reported as an auxiliary source-to-target sanity metric, not a full external validation claim.

  • Failed baselines are recorded in baseline_failures.json and excluded from rankings instead of contaminating metrics with NaN/Inf values.

Competitor references:

  • TRIBE v2: stimulus-to-fMRI competitor. Exact reproduction status remains unavailable; local tribe_style is a toy clean-room approximation and does not use TRIBE code, weights, notebooks, configs, or heavy runtime dependencies.

  • BrainVista: fMRI forecasting/stimulus-to-brain competitor. Exact reproduction status: unavailable until upstream code/data/protocol are integrated.

  • Brain-OF: fMRI/EEG/MEG shared foundation competitor. Exact reproduction status: unavailable until upstream code/data/protocol are integrated.

  • BrainOmni: EEG/MEG tokenizer/model competitor. Exact reproduction status: unavailable until upstream code/data/protocol are integrated.

  • Brain Harmony: structure + function token competitor. Exact reproduction status: unavailable until upstream code/data/protocol are integrated.

  • Braindecode and CEBRA: optional wrapper slots. Status remains unavailable unless dependencies and compatible protocols are installed.

Prepared reports distinguish exact, local_baseline, clean_room_approximation, approximation, and unavailable. Approximation rows are lanes for fair comparison, not claims of exact reproduction.

Pair-Operator is no longer the main architecture. It remains a local baseline and ablation lane for testing whether low-rank pair state helps the Neural Field Compiler.

The TribeStyleModel.from_pretrained and get_events_dataframe names are temporary public compatibility aliases for baseline-plumbing smoke tests only. Prefer get_event_rows or build_events for row-returning local event fixtures so callers do not infer a pandas DataFrame contract. Do not extend the TRIBE-shaped shim surface; sunset those aliases when exact upstream TRIBE reproduction is integrated or before a stable v1 model API is declared.

Do not claim full competitor reproduction unless exact code, data, and protocol are actually used.