Phoebe Domain - Feature Reference

Partially implemented (TODO Phase 181, in progress). Phoebe is the brain-and-mind science platform of the Oshun ecosystem, pursuing fundamental breakthroughs across brain-computer interfaces, neuroscience, clinical psychology/psychiatry, the AI-driven design and running of experiments, and a unified brain-mind knowledge graph. Most capabilities below describe the planned envelope, but an initial slice already ships under libs/phoebe/ — core, assessment, mbc, and comp-psychiatry (~1,992 non-test LOC). In particular the validated-assessment engine (PHQ-9 / GAD-7 / PCL-5 / C-SSRS with canonical scoring bands and crisis hard-triggers) is real, shipping code.

Phoebe's mission is to turn passive observation of brains and behaviour into causal, model-driven discovery and care — with AI integrated at every layer, honest uncertainty everywhere, and a qualified human in the loop wherever a living brain or a patient is on the other end.

Pillar 1 — Neural Interfaces & Signal Intelligence (BCI)#

• Multimodal acquisition across EEG, MEG (incl. OPM-MEG), ECoG, intracortical spikes (Neuropixels 1.0/2.0/Ultra, Connexus-class), sEMG (Meta-Neural-Band class), fNIRS, and functional ultrasound — with sub-millisecond time-sync over Lab Streaming Layer and XDF recording/replay.
• Real-time signal processing: filtering, ICA artifact removal, Common Spatial Patterns, xDAWN, and Riemannian/SPD-covariance featurization on a bounded-latency Rust hot path.
• Decoding & neural foundation models: hot-swappable Kalman/RNN/transformer decoders, spike foundation models (POYO/POYO+), masked-population models (NDT2/NEDS), EEG/SEEG models (LaBraM), LLM-coupled Brain-to-Text, and fMRI/EEG-to-image — with a benchmarking harness (WER, bits/s, co-bps, ITR).
• Cross-session / cross-subject adaptation (Riemannian/manifold alignment, meta-learning) that targets the daily-recalibration problem.
• Closed-loop neuromodulation: adaptive DBS (beta-LFP-driven), responsive neurostimulation, transcranial focused ultrasound and temporal-interference control, biomimetic ICMS sensory feedback, and a thermal/charge safety governor.

Pillar 2 — Computational & Systems Neuroscience#

• Spike sorting & imaging via SpikeInterface & Kilosort 4 with drift correction and QC; calcium/voltage pipelines at acquisition speed.
• Connectomics: flood-filling-network segmentation, ML-driven proofreading, CAVE-style versioned annotation, and Neuroglancer/TensorStore petascale serving with a synapse-graph query API and function↔structure fusion.
• Multi-scale simulation: managed NEST/NEURON/CoreNEURON/Arbor/Brian2 backends over the SONATA interchange, Arbor↔TVB co-simulation, and GPU/HPC scheduling — seeded from Open Brain Institute / Allen reconstructions.
• Brain foundation models & digital twins: fMRI/multimodal foundation models, topographic priors, function→structure twins, in-silico optogenetics / MEI stimulus synthesis, and patient-specific virtual-brain twins that report calibrated uncertainty and refuse to overstate.
• Latent dynamics & representation: LFADS/AutoLFADS/CEBRA/pi-VAE, manifold-native APIs, a unified RSA≈CKA≈CCA comparison module, and first-class identifiability/degeneracy reporting.

Pillar 3 — Clinical Psychology & Psychiatry#

• Validated assessment engine: PHQ-9, GAD-7, PCL-5, C-SSRS with exact scoring/banding and computable DSM-5-TR / ICD-11 coding; crisis items hard-trigger escalation.
• Measurement-based & collaborative care: scheduled re-administration, symptom-trajectory dashboards, treat-to-target alerting, and a caseload registry (IMPACT/CoCM pattern).
• Digital phenotyping & JITAI: EMA, passive sensing (sleep, mobility, typing, voice acoustics), relapse/anomaly detection with honest uncertainty, and a contextual-bandit just-in-time adaptive-intervention engine (MRT-evaluable).
• Computational psychiatry & treatment matching: RDoC/HiTOP dimensional models, RL/drift-diffusion/active-inference analyses, circuit-biotype and EEG-biomarker ingestion, and treatment-matching decision support (research-use, advisory to a clinician — never autonomous prescribing).
• Regulated therapeutics & clinician copilot: CBT/DBT/ACT/MI content with anti-sycophancy guardrails delivered through Psyche, crisis detection, and a behavioural-health ambient scribe (SOAP/GIRP/BIRP/DAP) — all clinician-signed.

Pillar 4 — Experiment Design & Autonomous Discovery#

• Hypothesis engine: multi-agent generate/debate/rank/evolve over the knowledge graph (Co-Scientist pattern, on the Phase 178 & Kalika substrate), with graph-derived novelty and GNN-derived feasibility scoring.
• Optimal experimental design: Bayesian OED (expected information gain), amortized DAD/iDAD policies, ADO/QUEST+ for psychophysics, Bayesian adaptive trials (response-adaptive randomization, group-sequential boundaries, BOIN/CRM, master protocols), and SMART/offline-RL dynamic-treatment-regime design.
• Experiment runtime: a frame-precise authoring tool compiling to a WebGPU browser runtime and a native lab runtime, a closed-loop layer over LSL for brain-state-dependent stimulation, and robotic/wet-lab execution backends (Opentrons, Emerald Cloud Lab/Strateos, automated patch-clamp, DeepLabCut/ SLEAP).
• Subject & panel management: Prolific/MTurk/Gorilla/SONA connectors, a data-quality firewall (incl. LLM-agent contamination detection), and eligibility/patient↔trial matching.
• Analysis & reproducibility: containerized BIDS-App/MNE pipelines with enforced multiple-comparison correction, a DoWhy/causal-learn discover→ identify→estimate→refute flow, machine-readable auto-preregistration, multiverse analysis by default, and an end-to-end provenance ledger.

Pillar 5 — Brain-Mind Knowledge & Evidence#

• Knowledge graph & ontology service: a Biolink-typed store ingesting SPOKE/PrimeKG/Hetionet/Monarch/NeuroKG, hosting Cognitive Atlas, NIFSTD, Uberon, NeuroNames, NBO, MFOEM, MeSH, SNOMED CT, DSM-5-TR, and ICD-11 — plus the canonical machine-readable RDoC ontology and the DSM↔ICD-11↔MeSH↔SNOMED ↔Cognitive-Atlas↔RDoC crosswalk that do not exist anywhere today.
• Literature & evidence: PubMed/OpenAlex/Semantic-Scholar/Europe-PMC/preprint ingestion, SciFact-style claim/evidence extraction, smart-citation typing, living systematic reviews, and GraphRAG with KGARevion-style KG verification to suppress hallucination.
• FAIR dataset catalog: UK Biobank, HCP, ABCD, ENIGMA, OpenNeuro, IBL, Allen ABC Atlas, ABIDE, ADNI, PPMI, AnswerALS, NDA, DANDI, BIL, EBRAINS — with modality, scale, and access-tier metadata and native NWB/BIDS/DICOM I/O.

Cross-Cutting — Neuroethics, Governance, Safety & the AI Engine#

• Governance: neural-data governance to the Chile/Colorado/California/Montana/ Connecticut neurorights laws & UNESCO 2025 standard, HIPAA/GDPR, mandatory neuroimaging defacing, federated learning and differential privacy, and dbGaP-style controlled access with immutable audit.
• Regulatory: FDA SaMD lifecycle (510(k)/De Novo/PMA), Predetermined Change Control Plans, bias/equity and model-drift testing, post-market monitoring, and a clinical-evidence/RCT & registration pipeline.
• Crisis-safety backbone: 988/Samaritans routing, suicidality escalation, age-gating, AI-disclosure, and jurisdiction-aware feature gating (Illinois WOPR / Nevada AB 406 / Utah HB 452).
• Autonomous closed-loop neuroscientist: hypothesis → adaptive design → fail-closed AI-IRB ethics gate → execution → analysis → theory update → re-rank, with hallucination/fabrication guardrails and human checkpoints on every human-subjects action.

Application Suite#

• Researcher Workbench — hypothesis → design → run → analyze → preregister/ reproduce.
• Clinician Console — assessment, MBC dashboards, treatment-matching, and the copilot/scribe, all with human sign-off gates.
• Neural-Interface Control Room — live BCI decode/stim, signal-quality and device-health telemetry, latency SLAs.
• Knowledge Explorer — KG / literature / evidence browsing with verified-citation answers.
• Participant/Patient app — consent, EMA, phenotyping opt-ins, and therapeutic content with crisis safety.