NeuroTwin Project State: NFC Pivot and Evidence-Gated Plan#
One-Paragraph Summary#
NeuroTwin is a Python research package and CLI for leakage-audited neural translation experiments. The repo began as a benchmark scaffold for missing-modality reconstruction, future-state forecasting, cross-modal translation, few-shot subject adaptation, and dataset/site generalization under held-out splits. The current research direction keeps that infrastructure but makes NeuroTwin NFC, the Neural Field Compiler, the experimental model primitive: recordings are treated as partial, lossy observations of one evolving latent neural field rather than separate modalities to fuse directly. The repo is currently an evidence-gated research harness, not a completed model-superiority result.
Architecture Hierarchy#
Layer |
Role |
Status |
|---|---|---|
NeuroTwin NFC / Neural Field Compiler |
Core experimental architecture |
implemented experimental path |
NeuralStateSpaceTranslator |
Legacy/current NeuroTwin baseline |
retained baseline |
NeuroTwinPairOperator |
Historical pairwise operator |
retained baseline and ablation |
Ridge/autoregressive ridge/direct MLP/TCN/Transformer/SSM |
Local sanity and sequence baselines |
implemented baselines |
BrainVista-style/TRIBE-style |
Clean-room approximation lanes |
approximations only |
Manifest, leakage, claim gates, reports |
Evidence infrastructure |
trusted but must stay claim-hygienic |
fNIRS/TurboVec/TurboQuant |
Optional theory/infrastructure notes |
docs only or deferred |
Track A and Track B#
Track A is the reproducibility and model-gates paper lane. It uses MOABB/EEG and synthetic smoke paths to validate leakage audits, identity probes, model cards, paper-mode reports, evidence bundles, and honest claim boundaries. Track A does not require NeuroTwin to beat ridge.
Track B is the model architecture paper lane. It uses NFC. The first proof point is strict synthetic latent-field recovery. Only after that gate passes should the repo move to Algonauts/CNeuroMod fMRI with verified real stimulus features, then later to 1x A100 debug and 6x A100 DDP.
What Old NeuroTwin Is Now#
NeuralStateSpaceTranslator remains the legacy/current NeuroTwin baseline. It is useful because it shares the task API, prepared-manifest path, and evaluation surface. It is not the future claim by itself.
What Pair-Operator Is Now#
NeuroTwinPairOperator is no longer the main primitive. It is retained as a historical baseline and as an ablation lane for low-rank relational updates inside the NFC question.
What NFC Is Now#
NFC is the experimental architecture path. It models a subject-specific latent field F_s(x,t,omega) and learns observation operators for signals such as fMRI, EEG, behavior, and stimulus-conditioned responses.
What fNIRS Contributes#
fNIRS supports the observation-operator worldview: physiology, optics, hemodynamics, and artifacts make measurement a lossy operator over latent activity. This repo does not claim fNIRS support, MDD support, clinical diagnosis, or private-data results.
What TurboQuant/TurboVec Contributes#
TurboQuant/TurboVec is optional future retrieval, compression, and semantic-audit infrastructure around NFC. It is not the core model contribution and must not become a required dependency.
What Is Implemented#
ntCLI for doctor, data prepare/smoke/audit, split audit, estimate, train, eval, and report commands.Event manifests, split manifests, leakage audits, prepared-task generation, and claim gates.
Synthetic and MOABB preparation lanes.
Prepared-manifest training with checkpointing, resume, metrics, reports, and DDP hooks.
Local baseline runners and competitor registry.
NFC model modules under
src/neurotwin/models/nfc/.NFC synthetic suite and falsification report.
A100 packaging and handoff scripts.
What Is Scaffolded#
BIDS/OpenNeuro derivative scanning without raw preprocessing.
A100/H100 cluster templates.
BrainVista/TRIBE-style approximation lanes.
Paper-mode artifact contracts.
What Is Theory Only#
Full fNIRS observation support.
TurboVec/TurboQuant integration.
Spike/calcium modality implementations.
Exact upstream TRIBE v2, BrainVista, Brain-OF, BrainOmni, or Brain Harmony reproduction.
What Is Archived/Deferred#
Pair-Operator as the main architecture.
Clinical digital-twin language.
Depression/MDD diagnostic claims.
Raw public neural data in git.
Full Algonauts/CNeuroMod model claims before strict gates.
Next Experiment Ladder#
Correct and trust the NFC synthetic falsification gate.
Run strict 1x A100 NFC synthetic diagnostic:
synthetic50.If needed and documented, escalate to
synthetic100.Only after pass, run Algonauts/CNeuroMod 1x debug with verified stimulus hashes.
Only after debug pass, consider 6x A100 DDP.
Draft model-paper claims only after held-out evidence and baselines pass.
Claim Boundaries#
Allowed: leakage-audited benchmark infrastructure, experimental NFC path, synthetic NFC smoke/falsification suite, MOABB Track A reproducibility evidence path, A100 handoff packaging from clean committed HEAD.
Not allowed: model superiority, proven NFC, NeurIPS-quality result, clinical diagnosis, fNIRS support, exact TRIBE v2 or BrainVista reproduction, first brain foundation model, or A100 success without evidence bundles.
Kill Criteria#
NFC cannot beat direct baselines on true field-grounded synthetic tasks.
NFC equals no-observation or no-pair ablations within the configured threshold.
Metrics are NaN or missing.
Shape mismatch is padded or broadcast instead of failing.
Real stimulus hashes are unavailable for stimulus-to-fMRI claims.
Held-out splits fail leakage audit.
Evidence gates and final reports disagree.