Domain · Architecture

Phoebe Domain - Architecture

Planned-only domain.

5sections5 minread
Markdown source
On this page

Planned Brain-Computer Interface, Neuroscience, Clinical Psychology & Autonomous Discovery Platform (TODO Phase 181)

Phoebe is the Oshun domain for the science and medicine of the brain and mind. It spans five frontier fields and the connective tissue between them: brain-computer interfaces (BCI), computational and systems neuroscience, clinical psychology and psychiatry, the AI-driven design and execution of experiments on brains and behaviour, and a unified knowledge graph linking molecules → cells → circuits → cognition → behaviour → disorders. Every layer is AI-integrated, and every layer is built to one standard: read biological reality honestly, report uncertainty loudly, keep a qualified human in the loop wherever a living brain or a patient is involved, and turn passive observation into causal, model-driven discovery.

The domain is planned-only. No libs/phoebe/* or apps/phoebe/* packages exist in the monorepo yet. All design detail here belongs to TODO Phase 181 (§181.1.1), which creates the workspace and implements the platform from scratch. This document captures the intended topology, layer structure, and ownership boundaries so implementation can begin against a defined contract rather than an empty namespace.

Status#

Planned-only domain. No workspace packages exist yet. Phoebe is a standalone domain that depends on existing assets (Iris, Aphrodite, Psyche, the Phase 178 autonomous-research substrate, Kalika) via contracts; it does not move or absorb their code.

Planned Workspace Shape#

Phase 181 (§181.1.1) creates thirty libs/phoebe/* libraries plus a libs/contracts/phoebe/ contracts package. The application suite (Researcher Workbench, Clinician Console, Neural-Interface Control Room, Knowledge Explorer, Participant/Patient app) lives inside libs/phoebe/apps/ (§181.30); there are no standalone apps/phoebe/* projects in the planned scope.

text
libs/phoebe/
  core/                  # domain primitives, RDoC/HiTOP dimensional model, state machines
  neural-io/             # multimodal acquisition & LSL/XDF sub-ms time-sync
  signal/                # real-time DSP (Rust): ICA/CSP/xDAWN/Riemannian
  decoding/              # decoder runtime, neural foundation models, cross-subject adaptation
  neuromodulation/       # closed-loop aDBS/RNS/tFUS/TI, read–write, safety governor
  connectomics/          # FFN segmentation, CAVE proofreading, petascale graph serving
  simulation/            # NEST/NEURON/Arbor/Brian2 over SONATA, multi-scale co-sim
  brain-models/          # brain foundation models & circuit/region digital twins
  neuro-analysis/        # SpikeInterface/Kilosort, LFADS/CEBRA, manifolds
  assessment/            # PHQ-9/GAD-7/PCL-5/C-SSRS & DSM-5-TR/ICD-11 coding
  mbc/                   # measurement-based & collaborative care
  phenotyping/           # passive sensing, EMA, voice biomarkers, JITAI
  comp-psychiatry/       # RDoC/HiTOP models, biotypes, treatment matching
  therapeutics/          # CBT/DBT/ACT/MI content, anti-sycophancy, clinician copilot
  hypothesis/            # KG-grounded generate/debate/rank & novelty/feasibility
  design/                # BOED/DAD/ADO, Bayesian adaptive trials, SMART/JITAI
  experiment-runtime/    # frame-precise authoring/runtime & closed-loop/robotic execution
  subjects/              # panel connectors, data-quality firewall, eligibility matching
  analysis/              # BIDS-App/MNE pipelines, causal inference
  reproducibility/       # auto-preregistration, multiverse, provenance ledger
  knowledge-graph/       # Biolink KG, ontology service, RDoC OWL & DSM↔ICD crosswalk
  literature/            # PubMed/OpenAlex/S2 ingestion, SciFact claim extraction
  evidence-synthesis/    # living systematic reviews, GraphRAG & KGARevion
  datasets/              # FAIR catalog, NWB/BIDS/DICOM I/O
  governance/            # neurorights/HIPAA/GDPR, defacing, federated & DP compute
  regulatory/            # FDA SaMD & PCCP, bias/drift testing, RCT evidence
  safety/                # 988 crisis routing, age-gating, jurisdiction-aware HITL
  ai/                    # autonomous closed-loop neuroscientist & fail-closed AI-IRB
  apps/                  # researcher/clinician/control-room/explorer/patient surfaces
  integration/           # cross-domain hub

libs/contracts/phoebe/

Architecture#

The domain is structured in seven layers. Each layer builds on the one below it; no layer reaches across a layer boundary without going through the contracts package.

  • Core — the foundation: subject/participant, study, protocol, signal, recording, stimulus, trial, consent, assessment, hypothesis, experiment, claim, dataset, and ontology-term primitives, plus the RDoC/HiTOP dimensional model and the state machines (consent, IRB approval, preregistration lock, crisis escalation) that every other package depends on.
  • Neural-interface layer — acquisition, real-time signal processing, decoding/foundation models, and closed-loop neuromodulation. The hard real-time path is Rust; it owns the decode→stimulate loop and the stimulation safety governor.
  • Neuroscience layer — connectomics, multi-scale simulation, brain foundation models / digital twins, and spike-sorting & latent-dynamics analysis. Governed by the fail-loud uncertainty hard requirement: digital twins always carry calibrated uncertainty and refuse to overstate; whole-brain emulation is explicitly out of scope.
  • Clinical layer — assessment, measurement-based care, digital phenotyping, computational psychiatry, and regulated therapeutics. Governed by the human-in-the-loop hard requirement: no autonomous diagnosis, prescribing, or unsupervised therapy.
  • Discovery layer — hypothesis generation, optimal experimental design, experiment runtime, subject management, analysis, and reproducibility. The AI breakthrough engine orchestrates these into a closed loop.
  • Knowledge layer — the knowledge graph & ontology service, literature ingestion, evidence synthesis, and the dataset catalog. The single source of truth for "what is known" and the substrate every reasoning step is grounded and verified against.
  • Governance & experience layer — neuroethics/neurorights/privacy governance, the FDA SaMD regulatory harness, the crisis-safety backbone, the application suite, and the cross-domain integration hub. Governed by the append-only audit and jurisdiction-aware compliance hard requirements.

Boundaries#

Phoebe's scope is the science and medicine of the brain and mind. The boundaries below define where Phoebe ends and an adjacent domain begins, so the same fact is never owned twice.

  • Phoebe owns neuroscience-grade decoding and analysis, neural foundation models, closed-loop neuromodulation, clinical-psychology evidence and regulated therapeutics, the brain/behaviour experiment lifecycle, and the brain-mind knowledge graph.
  • Iris owns the multimodal human–computer interface, including the BCI device/HID layer (libs/iris/bci/*, libs/iris/multimodal/bci/*). Phoebe consumes those device primitives; it does not re-implement drivers, intent HID, or the device-level privacy framework.
  • Aphrodite owns the consumer biometric engine (libs/aphrodite/biometric-engine, incl. eeg/). Phoebe consumes biosignal acquisition and adds clinical/research interpretation.
  • Psyche owns affective computing and digital humans. Phoebe consumes affect/sentiment signals for phenotyping and uses the conversational stack to deliver therapeutic content, but owns the clinical guardrails on what is safe to say.
  • Phase 178 (Autonomous Research & Agentic Scientist) / Nous own the generic autonomous-scientist substrate and the ML platform. Phoebe specializes and constrains them for living brains and human subjects (BOED/ADO, IRB gating, neuro/behavioural execution, BIDS-App analysis).
  • Kalika owns mathematics/physics/materials research agents and notebooks. Phoebe reuses that infrastructure for the neuro/psych domain.
  • Lilith / Tara own contemplative wellness and meditation (consumer, non-regulated; libs/meditation/*). The boundary is the clinical line: a wellness practice is Lilith/Tara; a validated diagnostic/therapeutic with a clinician in the loop and a regulatory pathway is Phoebe.
  • Mnemosyne owns memory infrastructure; Phoebe consumes it and studies biological memory rather than owning the platform's artificial memory.

Verification Expectations#

When the packages are implemented, each area must be covered by a dedicated test suite asserting domain correctness, not data flow. The acceptance gates:

  • Time-sync tests — multimodal LSL streams synchronized to sub-millisecond jitter (≈≤200 µs SD).
  • Decoder benchmark tests — WER / bits-per-second / co-bps / ITR on held-out cross-session and cross-subject splits, where a random or fixed-output decoder scores at chance (the metric tests decoding, not plumbing).
  • Stimulation safety tests — latency-deadline adherence, thermal/charge hard-caps (≤0.5 °C), manual-override, and append-only incident logging.
  • Simulation interchange tests — a SONATA model produces matching spike rasters across at least two backends within numerical tolerance.
  • Digital-twin uncertainty tests — twins always emit calibrated bounds; an out-of-validated-scope request returns an explicit refusal, not a fabrication.
  • Assessment scoring tests — PHQ-9/GAD-7/PCL-5/C-SSRS scored against published worked examples; a C-SSRS or PHQ-9 item-9 positive hard-triggers the crisis backbone.
  • Adaptive-design tests — ADO reproduces the delay-discounting efficiency result (≥0.95 reliability in 10–20 trials) on a simulated participant; adaptive trials enforce Type-I control.
  • Causal-inference tests — the refutation step runs before any effect is reported; a confounded fixture is flagged by a placebo/refutation test.
  • Reproducibility tests — preregistration-vs-execution auto-diff flags an undisclosed analysis deviation; provenance graphs recompute with one command.
  • Crosswalk tests — a DSM-5-TR disorder resolves to its ICD-11/MeSH/SNOMED codes and related RDoC constructs through the SSSOM crosswalk, with provenance.
  • Verification tests — a GraphRAG triple absent from / contradicted by the KG is filtered before reaching an answer (KGARevion pattern).
  • Governance tests — a neural-data export for an unconsented purpose is hard-blocked; every access is immutably logged.
  • AI-IRB tests — the autonomous engine cannot enroll a human subject without passing the fail-closed ethics gate plus a human sign-off; a protocol violation halts the loop.