Historical task source. The Eve task board holds current task state.
Phase 181: Phoebe — Brain-Computer Interface, Neuroscience, Clinical Psychology & Autonomous Discovery Platform#
Phoebe (Greek Titaness Φοίβη, "radiant / bright"; grandmother of Apollo and the Titaness associated with the prophetic intellect and the Delphic oracle — "she who sees and foretells") is the comprehensive brain-and-mind science platform of the Oshun ecosystem. Phoebe pursues fundamental breakthroughs across five frontier fields and the connective tissue between them: brain-computer interfaces, computational and systems neuroscience, clinical psychology and psychiatry, the AI-driven design and execution of experiments on brains and behaviour, and the unified knowledge graph that links molecules → cells → circuits → cognition → behaviour → disorders. Every layer is AI-integrated and every layer is built to a single standard: read biological reality honestly, report uncertainty loudly, keep a qualified human in the loop wherever a human brain or a patient is on the other end, and turn passive observation into causal, model-driven discovery.
Domain Assignment: Phoebe Library Prefix: @phoebe/* Goddess:
Phoebe (Φοίβη) — Titaness of radiant intellect, prophetic insight, and the
oracle.
Status: In progress — 183/189 tasks complete. Built & tested: 30 @phoebe/*
packages (incl. the phoebe-neuromodulation Rust crate) + real Python backends
— Brian2 spiking simulation, a SONATA network on two backends (Brian2 + NEST,
matching rasters), Arbor↔TVB multi-scale co-simulation, BMTK/SONATA builds, a
GPU/HPC job scheduler with headroom guards, the TVB virtual-brain twin,
SpikeInterface spike-sorting benchmark, a calcium-imaging pipeline
(motion/segment/dF-F/OME-Zarr), latent-identifiability + RSA/CKA/CCA analysis,
PyNWB HDF5 round-trip, DANDI no-download streaming, LSL sub-ms time-sync + XDF
replay, a pyRiemann-validated motor-imagery benchmark on real EEGBCI data, and
the full @phoebe/connectomics CPU stack — a Biolink synapse-graph store &
connectivity query API validated on a FlyWire-derived olfactory fixture, a
CAVE-style versioned proofreading ChunkedGraph, an ML merge/split boundary-error
detector (from-scratch logistic regression, held-out AUC ≫ chance under a
label-permutation null), Neuroglancer-precomputed/sharded (compressed-Morton +
murmurhash3_x86_128) no-download chunk addressing, and MICrONS-style
function↔structure like-to-like fusion, plus the CPU @phoebe/brain-models
TopoLoss topographic prior (analytic-gradient smoothness objective + topography
diagnostic) and MEI / in-silico-optogenetics synthesis (projected-gradient MEI
that recovers a linear unit's receptive field to its closed-form optimum, and
causal unit-clamp readout), plus the board-agnostic @phoebe/neural-io
acquisition abstraction (BrainFlow-pattern prepare→stream→read lifecycle, a real
SPSC ring buffer that buffers a 384-channel 30 kHz second with zero sample loss,
a deterministic multi-sine synthetic board whose spectral peak recovers the
injected tone, and Iris/Aphrodite bridge adapters that consume their frames via
contracts), and the sEMG/fNIRS/fUS ingestion DSP (surface-EMG rectified
windowed-RMS envelope, the fNIRS modified Beer-Lambert ΔHbO/ΔHbR inversion that
round-trips exactly, and fUS mean-clutter-filtered power-Doppler), a
Neuropixels→NWB adapter (SPSC ring buffer drops 0/30000 samples and a PyNWB
ElectricalSeries passes pynwb.validate + round-trips with a checksum
matching the buffered stream), and a pure-TS Zarr v2 / OME-NGFF writer for
ScanImage/GEVI imaging that round-trips pixel-exact and is read back by
zarr-python. The phoebe PostgreSQL DB, the Neo4j Biolink KG store, and MinIO
object storage are provisioned against the dev docker stack; all five
@phoebe/apps Next.js apps are build-verified. The remaining 6 tasks are under
capability-by-capability audit and are left unchecked rather than represented by
tokenizer-only or fail-loud seams: GPU foundation-model training (spike/EEG/fMRI
foundation models & fMRI-to-image, §181.5.2, §181.9.1.1, §181.9.2.1), the
learned C++/CUDA flood-filling-network on petavoxel EM (§181.7.1.1; its
tiling/resume orchestration is built and tested on CPU). Spike sorting
(§181.10.1.1) now runs MountainSort5 through SpikeInterface with SNR/rate/
presence QC, validates motion estimation and interpolation on a known drifting
recording, exposes a capability-checked Kilosort4 wrapper that refuses to
substitute another sorter when CUDA is absent, and includes a bounded-memory
streaming template sorter for edge use. The latent-model suite (§181.10.2.1) now
runs real CPU-reference GPFA, LFADS/AutoLFADS, pi-VAE, and CEBRA fits, evaluated
on held-out data from a known-generative fixture against label-permutation/null
baselines with explicit identifiability and rotational-degeneracy reports;
production-scale sweeps remain a GPU deployment concern rather than an
implementation gap. Phoebe is a standalone domain that depends on existing
assets via contracts and does not move or absorb their code: it consumes
Iris's BCI device/HID layer (@iris/bci-*, @iris/multimodal-bci), Aphrodite's
biometric/EEG engine (@aphrodite/biometric-engine), Psyche's
affective-computing and conversational stack, the Phase 178 Autonomous-Research
& Agentic-Scientist substrate, and Kalika's autonomous-research-agent and
computational-notebook infrastructure.
The remaining capability gates and their re-entry conditions are recorded in
phase-181-capability-audit.md. That audit
is evidence for keeping the six tasks unchecked, not a substitute for completing
them.
Core Pillars:
- Neural Interfaces & Signal Intelligence (BCI) — multimodal neural acquisition with sub-millisecond time-sync (LSL/XDF over EEG/MEG/ECoG/spikes/ sEMG/fNIRS/fUS), real-time streaming signal processing (ICA, CSP, xDAWN, Riemannian/SPD-covariance methods), neural decoders and neural foundation models (POYO/NDT2/LaBraM-class), cross-session/cross-subject adaptation that removes daily recalibration, and closed-loop neuromodulation orchestration (adaptive DBS, RNS, tFUS, temporal-interference) with bidirectional read–write and sensory feedback.
- Computational & Systems Neuroscience — spike sorting and image processing at acquisition scale (SpikeInterface/Kilosort 4, calcium/voltage pipelines), connectomics segmentation & versioned proofreading & petascale graph serving (flood-filling networks, CAVE, Neuroglancer/TensorStore), multi-scale simulation (NEST, NEURON/CoreNEURON, Arbor, Brian2 over the SONATA interchange), brain/neural foundation models, latent-dynamics and manifold analysis (LFADS, CEBRA, RSA≈CKA≈CCA), and circuit/region digital twins with fail-loud uncertainty (whole-brain emulation is treated as infeasible, not a goal).
- Clinical Psychology & Psychiatry — a validated assessment engine (PHQ-9, GAD-7, PCL-5, C-SSRS) with computable DSM-5-TR / ICD-11 coding, measurement-based and collaborative care, digital phenotyping / passive sensing / JITAI, computational-psychiatry models (RDoC, HiTOP, RL/drift- diffusion/active-inference, circuit biotypes), regulated therapeutic content with anti-sycophancy guardrails, a clinician copilot/ambient scribe, and treatment-matching decision support — all human-in-the-loop gated and compliant with the 2025–2026 wave of AI-therapy law.
- Experiment Design & Autonomous Discovery — a hypothesis engine (generate/debate/rank/evolve over the knowledge graph with graph-derived novelty and GNN-derived feasibility), a Bayesian optimal-experimental-design / adaptive-design optimizer (BOED, amortized DAD/iDAD, ADO/QUEST+, Bayesian adaptive trials, contextual-bandit/MRT/SMART JITAIs), a frame-precise experiment authoring & runtime layer with closed-loop control, subject/panel management, an IRB/consent & fail-closed AI-ethics gate & DSMB, BIDS-App and causal-inference analysis pipelines, and a preregistration & reproducibility ledger that makes multiverse analysis the default.
- Brain-Mind Knowledge & Evidence — a Biolink-typed knowledge-graph store and ontology service (ingesting SPOKE/PrimeKG/Hetionet/Monarch/NeuroKG and hosting Cognitive Atlas, NIFSTD, Uberon, NeuroNames, NBO, MFOEM, MeSH, SNOMED CT, DSM-5-TR/ICD-11), the canonical machine-readable RDoC ontology and the maintained DSM↔ICD-11↔MeSH↔SNOMED↔Cognitive-Atlas↔RDoC crosswalk that do not yet exist anywhere (Phoebe's clearest greenfield to own), literature ingestion & claim/evidence extraction, living evidence synthesis, GraphRAG with KG-verified anti-hallucination, and a FAIR dataset catalog.
Two cross-cutting pillars bind these together:
- Neuroethics, Governance & Regulatory Compliance — neural-data governance to the Chile/Colorado/California/Montana/Connecticut neurorights laws & UNESCO 2025 standard, HIPAA/GDPR, federated learning and differential privacy, neuroimaging defacing, W3C-PROV provenance, FDA SaMD lifecycle (510(k)/De Novo/PMA) with Predetermined Change Control Plans, bias/equity testing, and a crisis-safety backbone (988/Samaritans routing, suicidality escalation, jurisdiction-aware feature gating).
- AI Breakthrough Engine — the autonomous closed-loop neuroscientist that ties hypothesis → adaptive design → ethics gate → execution → analysis → theory update, specialized for brains and behaviour, with reliability guardrails (hallucination/fabrication detection) on every autonomous step.
Technology Stack:
- TypeScript for orchestration, APIs, contracts, dashboards, clinician/ researcher consoles, and the knowledge/evidence services.
- Rust for the hard-real-time path that the rest of the platform cannot be:
streaming signal processing and artifact rejection, the decode→stimulate
closed loop (bounded-latency, fail-safe), neuromodulation safety state
machines, and high-throughput connectome graph queries. GPU via
wgpu. - C/C++ for simulation kernels and device DSP (CoreNEURON, Arbor backends, flood-filling-network inference, acquisition drivers).
- Python for ML and neuro/clinical pipelines: PyNWB/MatNWB, MNE-Python, SpikeInterface & Kilosort 4, NEST/NEURON/Arbor/Brian2, fMRIPrep/Nipype, DoWhy/causal-learn, Pyro (variational BOED), ADOpy, PsychoPy, and the neural- foundation-model training stack (POYO/NDT2/CEBRA/LaBraM lineages).
- WASM for the frame-precise browser experiment runtime (PsychoJS/PixiJS/ WebGPU) and for bridging Rust DSP into TS clients.
- PostgreSQL & TimescaleDB for neural/behavioural time-series and EMA; pgvector for embeddings; Neo4j / RDF triplestore for the Biolink knowledge graph; Redis for real-time session state; Kafka for the experiment/clinical/neural event bus.
- Lab Streaming Layer (LSL) & XDF for multimodal time-sync; NWB (HDF5 + Zarr) and BIDS as the canonical data-at-rest formats; Zarr / OME-Zarr + Neuroglancer-precomputed for petascale imaging and connectomes.
Distinction from Adjacent Domains:
- vs Iris — Iris owns the multimodal human–computer interface, including
the BCI device/HID layer (
libs/iris/bci/*,libs/iris/multimodal/bci/*— apple-hid, intent-prediction, privacy-framework, neural-signal, bci-device). Phoebe consumes those device primitives and adds neuroscience-grade decoding, neural foundation models, closed-loop neuromodulation, and the clinical/research applications above. Iris is the input device; Phoebe is the science and medicine behind and atop it. - vs Aphrodite — Aphrodite owns the consumer biometric engine, including
libs/aphrodite/biometric-engine/src/eeg/for wellness/affect. Phoebe consumes Aphrodite's biosignal acquisition but owns clinical interpretation, neuroscience analysis, and regulated/research use. Aphrodite = consumer biometrics; Phoebe = clinical and research neuroscience. - vs Psyche — Psyche owns affective computing and digital humans (emotion recognition, avatars, real-time conversational AI). Phoebe consumes Psyche's affect/sentiment signals for digital phenotyping and uses its conversational stack to deliver therapeutic content, but owns the clinical evidence base, validated assessment, and regulated therapeutics. Psyche is the face and voice; Phoebe is the clinical mind that governs what is safe to say.
- vs Phase 178 (Autonomous Research & Agentic Scientist) — Phase 178 is the horizontal autonomous-scientist substrate (Sakana AI Scientist v2, Google Co-Scientist, deep-research agents, physical-lab autonomy, observability/ ledger). Phoebe is the brain/behavioural vertical on top of it: BOED/ADO adaptive design, IRB/consent & fail-closed AI-ethics gating, neuro/behavioural wet-lab and online-experiment execution, BIDS-App analysis, and theory update into a neuroscience knowledge graph. Phoebe does not re-implement the generic engine; it specializes and constrains it for living brains and human subjects.
- vs Kalika — Kalika owns mathematics/physics/materials research agents, the CAS engine, and computational notebooks. Phoebe reuses Kalika's autonomous-research-agent and notebook infrastructure but owns the neuro/psych domain, its data standards, and its human-subjects constraints.
- vs Lilith / Tara — Lilith and Tara own contemplative wellness and
meditation (consumer, non-regulated; see
libs/meditation/*). Phoebe owns clinical psychology and psychiatry (measurement-based, clinician-in-the-loop, FDA/UKCA-gated). The boundary is the clinical line: a wellness practice is Lilith/Tara; a validated diagnostic or therapeutic with a clinician in the loop and a regulatory pathway is Phoebe. - vs Mnemosyne / Nous — Mnemosyne provides memory infrastructure and Nous provides the ML platform; Phoebe consumes both. Phoebe studies biological memory and cognition; it does not own the platform's artificial memory or model-serving substrate.
181.1 Domain Foundation & Infrastructure#
181.1.1 Library Structure Setup#
- 181.1.1.1 Create
libs/phoebe/root with domain README, ownership boundaries, and the planned-package map; add@phoebe/*path mappings totsconfig.base.jsonand register projects with Nx. - 181.1.1.2 Create
libs/phoebe/core/— domain primitives, entities, enums, state machines, and the shared dimensional (RDoC/HiTOP) model. - 181.1.1.3 Create
libs/phoebe/neural-io/— multimodal neural acquisition and LSL/XDF time-sync. - 181.1.1.4 Create
libs/phoebe/signal/— real-time streaming signal processing (Rust hot path & TS/Python bindings). - 181.1.1.5 Create
libs/phoebe/decoding/— decoder runtime, model registry, and neural foundation models. - 181.1.1.6 Create
libs/phoebe/neuromodulation/— closed-loop stimulation orchestration and bidirectional read–write. - 181.1.1.7 Create
libs/phoebe/connectomics/— EM segmentation, versioned proofreading, and petascale graph serving. - 181.1.1.8 Create
libs/phoebe/simulation/— multi-scale neural simulation engines over SONATA. - 181.1.1.9 Create
libs/phoebe/brain-models/— brain/neural foundation models and circuit/region digital twins. - 181.1.1.10 Create
libs/phoebe/neuro-analysis/— spike sorting, imaging pipelines, latent-dynamics and manifold analysis. - 181.1.1.11 Create
libs/phoebe/assessment/— validated psychometric instruments and DSM-5-TR/ICD-11 coding. - 181.1.1.12 Create
libs/phoebe/mbc/— measurement-based and collaborative care. - 181.1.1.13 Create
libs/phoebe/phenotyping/— digital phenotyping, passive sensing, and JITAI. - 181.1.1.14 Create
libs/phoebe/comp-psychiatry/— RDoC/HiTOP models, computational models, and biotype/treatment-matching. - 181.1.1.15 Create
libs/phoebe/therapeutics/— therapeutic content, anti-sycophancy guardrails, and clinician copilot. - 181.1.1.16 Create
libs/phoebe/hypothesis/— KG-grounded hypothesis generation and scoring. - 181.1.1.17 Create
libs/phoebe/design/— optimal experimental design and adaptive-trial/JITAI optimizers. - 181.1.1.18 Create
libs/phoebe/experiment-runtime/— experiment authoring, frame-precise runtime, and closed-loop control. - 181.1.1.19 Create
libs/phoebe/subjects/— subject/panel management and data-quality firewall. - 181.1.1.20 Create
libs/phoebe/analysis/— BIDS-App pipelines and causal inference. - 181.1.1.21 Create
libs/phoebe/reproducibility/— preregistration, multiverse, and the provenance ledger. - 181.1.1.22 Create
libs/phoebe/knowledge-graph/— Biolink KG store and ontology service. - 181.1.1.23 Create
libs/phoebe/literature/— literature ingestion and claim/evidence extraction. - 181.1.1.24 Create
libs/phoebe/evidence-synthesis/— living systematic reviews and GraphRAG. - 181.1.1.25 Create
libs/phoebe/datasets/— FAIR dataset catalog and NWB/BIDS I/O connectors. - 181.1.1.26 Create
libs/phoebe/governance/— neural-data governance, neurorights, privacy, and provenance. - 181.1.1.27 Create
libs/phoebe/regulatory/— FDA SaMD lifecycle and validation harness. - 181.1.1.28 Create
libs/phoebe/safety/— crisis-safety and compliance backbone. - 181.1.1.29 Create
libs/phoebe/ai/— the autonomous closed-loop neuroscientist (AI breakthrough engine). - 181.1.1.30 Create
libs/phoebe/apps/— researcher/clinician/control- room/participant application suite. - 181.1.1.31 Create
libs/phoebe/integration/— cross-domain hub. - 181.1.1.32 Create
libs/contracts/phoebe/(@phoebe/contracts) — Zod schemas and cross-domain DTOs for every contract Phoebe exposes/consumes.
181.1.2 Core Data Model & Persistence#
- 181.1.2.1 Provision the
phoebePostgreSQL database (per domain-isolation convention) with TimescaleDB hypertables for neural and behavioural time-series andpgvectorfor embeddings. - 181.1.2.2 Stand up the knowledge-graph store (Neo4j or RDF triplestore) with the Biolink model as the schema layer.
- 181.1.2.3 Provision object storage (MinIO/S3) with Zarr/OME-Zarr + Neuroglancer-precomputed layouts for petascale imaging/connectomes and NWB (HDF5 & Zarr) for electrophysiology.
- 181.1.2.4 Define the canonical entity set in
@phoebe/core: Subject/ Participant, Study, Protocol, Session, Recording, Signal/Channel, Stimulus, Trial, ConsentRecord, Assessment, Measurement, Hypothesis, Experiment, Claim, Dataset, OntologyTermRef — with explicit state machines. - 181.1.2.5 Implement the dimensional clinical model: RDoC matrix (6 domains × 7 units of analysis) and HiTOP spectra & p-factor, as first-class, queryable structures (not free text).
- 181.1.2.6 Define the LSL/XDF stream descriptor and NWB/BIDS object mappings so every recording is born standards-conformant.
181.1.3 API Gateway & Service Topology#
- 181.1.3.1 Implement the Phoebe API gateway (TypeScript) with role-scoped access (researcher, clinician, participant, admin, autonomous agent) and per-role capability tokens.
- 181.1.3.2 Define the service topology: real-time (Rust) neural/decode/ stim services, batch (Python) pipeline services, knowledge/evidence services, and clinical services — with the Kafka event bus between them.
- 181.1.3.3 Wire OpenAPI &
@phoebe/contractsso every endpoint is contract-validated; publish the Phoebe contracts other domains import. - 181.1.3.4 Implement a global fail-loud uncertainty convention:
every inference/decode/prediction response carries calibrated confidence
and an explicit
not_configured/insufficient_evidenceseam rather than fabricating a result.
181.2 @phoebe/core — Domain Primitives#
181.2.1 Neural & Signal Types#
- 181.2.1.1 Define typed neural-signal primitives spanning modality, sampling rate, channel geometry, and reference scheme for EEG, MEG, ECoG, intracortical spikes, sEMG, fNIRS, and fUS.
- 181.2.1.2 Define the device-capability descriptor (channel count, bandwidth, latency budget, invasiveness class) and a modality-to-bandwidth ceiling model (e.g., ~50–63 bits/s non-invasive vs intracortical 8–10 bits/s cursor control) used to validate use-case feasibility.
- 181.2.1.3 Define provenance and lineage types that attach to every sample, derivation, and claim (W3C PROV-aligned).
181.2.2 Clinical & Experimental Types#
- 181.2.2.1 Define the assessment/measurement types with instrument versioning, scoring bands, and a hard-trigger flag for crisis items.
- 181.2.2.2 Define Study/Protocol/Hypothesis/Experiment types with a preregistration lock and an IRB-approval state machine.
- 181.2.2.3 Define ConsentRecord with broad/dynamic consent variants, withdrawal, and neural-data-class tagging.
- 181.2.2.4 Write domain-correctness tests for every state machine (legal/illegal transitions for consent, IRB approval, preregistration lock, and crisis escalation).
181.3 @phoebe/neural-io — Multimodal Acquisition & Time-Sync#
181.3.1 Acquisition & Device Abstraction#
- 181.3.1.1 Implement a board-agnostic acquisition abstraction
(BrainFlow pattern) and adapters that consume Iris's
@iris/bci-*device layer and Aphrodite's biometric engine rather than re-implementing drivers. - 181.3.1.2 Implement Neuropixels 1.0/2.0/Ultra and high-channel (Argo/
Connexus-class) ingestion adapters writing directly to NWB, with 30 kHz
streaming buffers.
- File:
libs/phoebe/neural-io/src/neuropixels.ts(+SampleRingBufferinacquisition.ts, PyNWB backendpython/neuropixels_nwb.py) - Acceptance: a 384-channel 30 kHz stream is buffered without sample loss and
persisted as a valid NWB
ElectricalSeries. Verified: ring buffer drops 0/30000 samples; PyNWB writes a NeuropixelsElectricalSeries(int16, 2-column staggered geometry) that passespynwb.validate, round-trips, and whose checksum matches the buffered stream bit-for-bit.
- File:
- 181.3.1.3 Implement ScanImage/light-beads-microscopy and GEVI voltage
adapters emitting OME-Zarr.
- File:
libs/phoebe/neural-io/src/imaging.ts - Verified: a pure-TS Zarr v2 + OME-NGFF 0.4 writer round-trips ScanImage (TZYX uint16) and GEVI (TYX float32) stacks pixel-exact and its output is read back pixel-exact by zarr-python 3.2.1 (external spec-compliance cross-check).
- File:
- 181.3.1.4 Implement sEMG (Meta-Neural-Band-class, 16-ch @ 2 kHz) and fNIRS/fUS ingestion adapters.
181.3.2 Lab Streaming Layer & Synchronization#
- 181.3.2.1 Integrate Lab Streaming Layer (pylsl & a Rust client) as
the multimodal transport, with NTP-style offset/RTT clock filtering.
- Acceptance: cross-stream synchronization jitter is sub-millisecond (≈≤200 µs SD) across EEG & eye-tracking & IMU & marker streams.
- 181.3.2.2 Implement XDF recording and replay for deterministic offline re-runs and decoder regression testing.
- 181.3.2.3 Implement signal-quality and impedance telemetry with a per-channel health model feeding the monitoring dashboards.
181.4 @phoebe/signal — Real-Time Streaming Signal Processing#
181.4.1 Filtering & Artifact Rejection#
- 181.4.1.1 Implement the streaming DSP DAG in Rust (Timeflux/ NeuroPype pattern): re-referencing, band-pass/notch, resampling, epoching — with a bounded per-window latency budget exposed as an SLA.
- 181.4.1.2 Implement ICA artifact removal (Picard/Infomax) with ICLabel-style component classification for EOG/EMG/ECG rejection.
- 181.4.1.3 Implement online drift/impedance-change detection that flags when recalibration or the cross-subject adapter is required.
181.4.2 Feature Extraction#
- 181.4.2.1 Implement Common Spatial Patterns (generalized eigendecomposition of class covariances) and xDAWN ERP spatial filters.
- 181.4.2.2 Implement Riemannian / SPD-covariance featurization
(pyRiemann-equivalent in Rust): tangent-space projection, MDM classifier,
and Riemannian Procrustes alignment.
- Acceptance: on a public motor-imagery set, the Riemannian MDM pipeline reproduces published accuracy within tolerance and beats a CSP+LDA baseline on cross-session transfer.
- 181.4.2.3 Implement streaming spike binning and band-power features for intracortical decoders.
181.5 @phoebe/decoding — Decoders & Neural Foundation Models#
181.5.1 Decoder Runtime & Model Registry#
- 181.5.1.1 Implement a hot-swappable decoder runtime supporting Kalman, LSTM/RNN, and transformer decoders behind a uniform interface with a bounded inference-latency contract.
- 181.5.1.2 Implement the model registry with versioning, per-subject calibration artifacts, and edge/on-device serving.
- 181.5.1.3 Implement an LLM-coupled language-model decoding stage for Brain-to-Text (phoneme/word logits → LLM rescoring), reproducing the Brain-to-Text-benchmark ensemble-merge pattern.
181.5.2 Neural Foundation Models#
- 181.5.2.1 Spike-level foundation models (POYO/POYO+ tokenization,
Perceiver backbone) with cross-session stitching. Split 2026-09-18: the
code/serving half is actionable; the training half is budget-gated.
- 181.5.2.1.a Implement the tokenizer, the Perceiver-style backbone architecture and the inference/serving path in the Phoebe libraries, exercised on synthetic spike trains. Verify: a spec tokenizes a fixture multi-unit spike recording, runs a forward pass and asserts output shape, determinism, and that cross-session stitching aligns two fixture sessions.
- 181.5.2.1.b Train the spike foundation model on real neural
recordings. Needs an owner decision on a compute and data budget (stated
2026-09-18).
blocked:governance
- 181.5.2.2 Masked-autoencoding population models (NDT2/NEDS) and
EEG/SEEG foundation models (LaBraM neural-spectrum tokenizer). Split
2026-09-18: the code/serving half is actionable; the training half is
budget-gated.
- 181.5.2.2.a Implement the LaBraM neural-spectrum tokenizer, the masked-autoencoder model architecture and its inference path, tested on synthetic EEG/SEEG. Verify: a spec tokenizes a fixture EEG segment, runs masked reconstruction and asserts the tokenizer round-trips and the model output shape.
- 181.5.2.2.b Train the population/EEG foundation models on real
recordings. Needs an owner decision on a compute and data budget (stated
2026-09-18).
blocked:governance
- 181.5.2.3 fMRI/EEG-to-image path (MindEye2-style shared-subject latent → CLIP → diffusion) as a research capability: assemble the pipeline from PUBLISHED pretrained components (a pretrained CLIP encoder and an open diffusion model, run on CPU; install-first, record versions, park only with a named failed install attempt) plus a small trainable brain→CLIP ridge/MLP adapter. Verify: a spec runs the pipeline end-to-end on a fixture (open NSD-style sample) and asserts a reconstructed image is produced and that the adapter improves CLIP-space cosine similarity over a random baseline. Training the adapter at scale on the full dataset is a budget decision, but the pipeline and a small-fit adapter are actionable here.
- 181.5.2.4 Establish a benchmarking harness (WER, bits/s, co-bps, ITR;
NLB/FALCON/Brain-to-Text style) with held-out cross-session/subject
splits.
- Acceptance: leaderboards are reproducible and a decoder that returns random or fixed outputs scores at chance — i.e., the metrics actually test decoding, not data flow.
181.5.3 Cross-Session / Cross-Subject Adaptation#
- 181.5.3.1 Implement Euclidean/Riemannian alignment and manifold alignment for training-free cross-subject transfer.
- 181.5.3.2 Implement online recalibration and meta-learning to attack session-to-session drift (the central usability blocker).
181.6 @phoebe/neuromodulation — Closed-Loop Stimulation & Read–Write#
181.6.1 Adaptive Stimulation Orchestration#
- 181.6.1.1 Implement a hard-real-time decode→stimulate loop in Rust with explicit deadlines, watchdogs, and fail-safe stop.
- 181.6.1.2 Implement adaptive-DBS policy primitives (beta-band LFP
sensing → amplitude adjustment, BrainSense-style) and responsive-
neurostimulation detect-and-stimulate state machines (RNS-style).
- Acceptance: a beta-driven aDBS policy modulates a simulated stimulator within the latency deadline and logs every state transition to an append-only ledger.
- 181.6.1.3 Implement non-invasive neuromodulation control interfaces for tFUS/LIFU and temporal-interference (envelope-frequency) stimulation, including target/beam parameterization.
181.6.2 Bidirectional Read–Write & Sensory Feedback#
- 181.6.2.1 Implement biomimetic ICMS sensory-feedback encoding (edge/orientation/motion patterns) for the read–write loop.
- 181.6.2.2 Implement a thermal/power/charge-density safety governor (≤0.5 °C cortical-heat constraint, charge-per-phase limits) that hard-caps stimulation regardless of policy output.
- 181.6.2.3 Implement the append-only stimulation incident/audit log with manual-override and safety-stop verification tests.
181.7 @phoebe/connectomics — Segmentation, Proofreading & Graph Serving#
181.7.1 Segmentation & Proofreading#
- 181.7.1.1 Flood-filling-network (FFN) inference for EM neurite segmentation with tiled, resumable processing: this is INFERENCE, not training — use a published FFN model. Install the runtime (record the version; park only with a named failed install attempt) and run tiled, resumable segmentation on the CPU (GPU on the Mac/lavapipe where available). Verify: a spec segments a fixture EM sub-volume, asserts the tiling covers the volume with the configured overlap, that a resumed run reproduces the un-interrupted result, and reports the segmentation against a small labelled fixture.
- 181.7.1.2 Implement ML-driven merge/split error detection to automate the proofreading cost center.
- 181.7.1.3 Implement CAVE-style versioned, distributed annotation (continuous versioning, edit history, multi-user proofreading).
181.7.2 Petascale Serving & Query#
- 181.7.2.1 Implement Neuroglancer-precomputed / TensorStore serving for petabyte-class volumes with streaming, no-download access.
- 181.7.2.2 Implement a synapse-graph store and query API scaling toward
10⁹–10¹² edges, with a cross-dataset schema spanning FlyWire/MICrONS/H01.
- Acceptance: a connectivity query (e.g., all post-synaptic partners of a cell class within a region) returns correct results against a FlyWire-derived fixture.
- 181.7.2.3 Implement function↔structure fusion linking activity recordings to the connectome (the MICrONS pattern).
181.8 @phoebe/simulation — Multi-Scale Neural Simulation#
181.8.1 Engine Backends & Interchange#
- 181.8.1.1 Implement managed backends for NEST (point/spiking), NEURON/CoreNEURON (biophysical), Arbor (GPU morphological), and Brian2/GeNN, behind one job interface.
- 181.8.1.2 Adopt SONATA as the model-interchange format so a model
can be ported across backends; implement a BMTK-style build/simulate/
analyze flow.
- Acceptance: a SONATA network runs on at least two backends and produces matching spike rasters within numerical tolerance.
- 181.8.1.3 Implement Arbor↔TVB multi-scale co-simulation (detailed circuit embedded in a whole-brain neural-mass model).
181.8.2 Scheduling & Scale#
- 181.8.2.1 Implement GPU/HPC job scheduling with memory checkpointing and headroom guards for large simulations.
- 181.8.2.2 Integrate Open Brain Institute / Allen (BMTK, SONATA) open reconstructions as starting assets rather than building circuits cold.
181.9 @phoebe/brain-models — Foundation Models & Digital Twins#
181.9.1 Brain Foundation Models#
- 181.9.1.1 fMRI foundation models (BrainLM-style MAE) and multi-modal
neural foundation models, with a hosted model zoo. Split 2026-09-18: the
code/serving half is actionable; the training half is budget-gated.
- 181.9.1.1.a Build the model-zoo serving scaffold in the Phoebe libraries: a registry of model cards (architecture, inputs, checksum), a loader, and an inference endpoint that serves a registered model on CPU. Verify: a spec registers a fixture (small) model, loads it and serves an inference, and rejects a card whose checksum does not match.
- 181.9.1.1.b Train the BrainLM-style fMRI MAE and the multi-modal
models. Needs an owner decision on a compute and data budget (stated
2026-09-18).
blocked:governance
- 181.9.1.2 Implement topographic priors (TopoLoss) as an optional training objective for biologically-plausible representations.
181.9.2 Circuit/Region Digital Twins#
- 181.9.2.1 Function→structure digital twins (mouse-visual-cortex-FM /
MICrONS pattern) predicting cell type and connectivity from function.
Needs an owner decision on a compute budget to train the twin, and the
paired functional+structural data (the MICrONS dataset is openly available
but must be fetched and stored) (stated 2026-09-18).
blocked:governanceblocked:corpus - 181.9.2.2 Implement in-silico optogenetics / most-exciting-input (MEI, EGG-diffusion) stimulus synthesis for closed-loop in-silico→in-vivo design.
- 181.9.2.3 Implement patient-specific virtual-brain-twin pipelines
(TVB-style) that report calibrated uncertainty and refuse to
overstate; whole-brain emulation is explicitly out of scope and
documented as infeasible.
- Acceptance: twin outputs always carry uncertainty bounds; a request beyond
validated scope returns an explicit
out_of_validated_scoperather than a fabricated prediction.
- Acceptance: twin outputs always carry uncertainty bounds; a request beyond
validated scope returns an explicit
181.10 @phoebe/neuro-analysis — Sorting, Imaging & Latent Dynamics#
181.10.1 Spike Sorting & Image Processing#
- 181.10.1.1 Integrate SpikeInterface as the orchestration core wrapping Kilosort 4 / MountainSort5 with drift correction and QC metrics; add a streaming/edge sorter for closed-loop use.
- 181.10.1.2 Implement calcium/voltage image pipelines (motion correction, segmentation, dF/F) at acquisition speed with OME-Zarr output.
- 181.10.1.3 Build a realistic-simulator & ground-truth benchmark harness for sorter accuracy.
181.10.2 Latent Dynamics & Representation#
- 181.10.2.1 Implement latent-variable models: LFADS/AutoLFADS, GPFA, pi-VAE, and CEBRA (consistent cross-subject/session embeddings).
- 181.10.2.2 Implement manifold-native APIs (project/align/decode across sessions) and a unified RSA≈CKA≈CCA representation-comparison module (bias-corrected).
- 181.10.2.3 Report identifiability/degeneracy as a first-class
quantity (per the 3,400-RNN solution-degeneracy result) on any
mechanism-recovery analysis.
- Acceptance: latent-model fits include a degeneracy/identifiability report; decoding accuracy is validated against a known-generative-model fixture, not asserted by truthiness.
181.11 @phoebe/assessment — Validated Psychometric Instruments#
181.11.1 Instruments & Scoring#
- 181.11.1.1 Implement PHQ-9 (0–27, bands 0-4/5-9/10-14/15-19/20-27, ≥10 cutoff) with exact scoring and versioning.
- 181.11.1.2 Implement GAD-7 (0–21, ≥10 cutoff) and PCL-5 (20×0–4, 0–80, DSM-5 diagnostic algorithm ≥1 B / ≥1 C / ≥2 D / ≥2 E rated ≥2, cutoff 31–33).
- 181.11.1.3 Implement C-SSRS (ideation severity 1–5, intensity,
behaviour, lethality) and PHQ-9 item-9 handling.
- Acceptance: a C-SSRS or PHQ-9 item-9 positive hard-triggers the
@phoebe/safetycrisis module and a clinician alert — never an autonomous reassurance response. Scoring is unit-tested against published worked examples.
- Acceptance: a C-SSRS or PHQ-9 item-9 positive hard-triggers the
181.11.2 Computable Nosology#
- 181.11.2.1 Implement DSM-5-TR and ICD-11 Chapter 06 computable coding via the ICD-11 Foundation API & post-coordination.
- 181.11.2.2 Map every instrument result to the dimensional RDoC/HiTOP
model and to nosology codes through the
@phoebe/knowledge-graphcrosswalk.
181.12 @phoebe/mbc — Measurement-Based & Collaborative Care#
181.12.1 Measurement-Based Care#
- 181.12.1.1 Implement scheduled repeat administration, symptom- trajectory dashboards, and treat-to-target alerting (Joint Commission CTS.03.01.09 aligned).
- 181.12.1.2 Implement the Collaborative Care Model caseload
registry and stepped-care review (IMPACT-trial pattern).
- Acceptance: a registry review surfaces patients not improving on schedule and computes correct response/remission status from instrument trajectories.
181.13 @phoebe/phenotyping — Digital Phenotyping & JITAI#
181.13.1 Passive Sensing & EMA#
- 181.13.1.1 Implement an EMA engine (configurable cadence, compliance tracking) and a passive-sensing pipeline (sleep, mobility/GPS entropy, typing, voice acoustics) on the Beiwe/mindLAMP pattern.
- 181.13.1.2 Integrate Psyche affect/sentiment signals and implement speech/voice biomarkers (acoustic & semantic-coherence) as decision-support features.
- 181.13.1.3 Implement relapse/anomaly detection (CrossCheck/SHARP pattern) that emits clinician alerts with explicit, calibrated uncertainty (effect sizes are small; surface that honestly).
181.13.2 Just-In-Time Adaptive Interventions#
- 181.13.2.1 Implement a JITAI engine using contextual bandits / Thompson sampling (HeartSteps II pattern) with nightly re-fit.
- 181.13.2.2 Implement micro-randomized-trial (MRT) instrumentation so
JITAI components are causally evaluable.
- Acceptance: phenotyping outputs are typed as decision-support signal, never autonomous diagnosis; the privacy path anonymizes on-device by default.
181.14 @phoebe/comp-psychiatry — Computational Psychiatry & Treatment Matching#
181.14.1 Computational Models#
- 181.14.1.1 Implement reinforcement-learning (reward-sensitivity / RPE), drift-diffusion, and model-based/model-free control fits as transdiagnostic markers.
- 181.14.1.2 Implement active-inference / predictive-coding analyses as a lens (documented as contested, not a validated mechanism).
181.14.2 Biomarkers & Treatment Matching#
- 181.14.2.1 Implement ingestion of circuit biotypes (Stanford six- biotype pattern) and EEG antidepressant-response markers (rACC theta).
- 181.14.2.2 Implement treatment-matching decision support routing to
pharmacotherapy / esketamine / TMS-SAINT / neuromodulation /
psychedelic-AT trials.
- Acceptance: outputs are research-use-only and advisory to a clinician; the module refuses autonomous prescribing and records that ML suicide prediction (e.g., VA REACH VET PPV ≈0.05%) is too inaccurate to drive automated decisions.
181.15 @phoebe/therapeutics — Therapeutic Content & Clinician Copilot#
181.15.1 Therapeutic Content & Guardrails#
- 181.15.1.1 Implement modality-specific (CBT/DBT/ACT/MI) retrieval- grounded therapeutic content delivered through the Psyche conversational stack.
- 181.15.1.2 Implement anti-sycophancy guardrails that challenge rather than confirm delusional/false beliefs, with an adversarial test suite for the documented failure mode.
- 181.15.1.3 Implement crisis-detection on every turn, escalating to
@phoebe/safety(988/Samaritans) and a human.- Acceptance: autonomous therapeutic conversation is gated behind clinician supervision and jurisdiction-aware feature flags (Illinois WOPR / Nevada AB 406 / Utah HB 452); the system never makes therapeutic decisions unsupervised and surfaces mandatory AI-disclosure.
181.15.2 Clinician Copilot & Ambient Scribe#
- 181.15.2.1 Implement a behavioural-health-tuned ambient scribe (SOAP/GIRP/BIRP/DAP) with DSM-5-TR/ICD-11 auto-coding and risk summarization.
- 181.15.2.2 Require clinician review/sign-off on every note, plan, and risk flag (unsupervised AI treatment plans are prohibited).
181.16 @phoebe/hypothesis — KG-Grounded Hypothesis Generation#
181.16.1 Multi-Agent Generation#
- 181.16.1.1 Implement a generate/debate/rank/evolve loop (Google
Co-Scientist pattern) over the
@phoebe/knowledge-graph, built on the Phase 178 substrate and Kalika research agents rather than re-implementing them. - 181.16.1.2 Implement an Elo tournament for pairwise hypothesis ranking with test-time-compute scaling.
181.16.2 Novelty & Feasibility Scoring#
- 181.16.2.1 Implement graph-derived novelty = KG distance + SemMedDB predication absence (not subjective LLM rating).
- 181.16.2.2 Implement GNN-derived feasibility (TxGNN/PROTON-style
link-prediction confidence) emitting explainable multi-hop paths.
- Acceptance: a known repurposing result (e.g., a documented compound→disease link) scores high feasibility with a correct, inspectable evidence path; a nonsensical link scores low.
181.17 @phoebe/design — Optimal Experimental Design & Adaptive Trials#
181.17.1 Bayesian Optimal Experimental Design#
- 181.17.1.1 Implement a BOED core maximizing Expected Information Gain (variational/ACE estimators, Pyro-backed).
- 181.17.1.2 Implement amortized DAD/iDAD/Step-DAD policies (train
offline, design in milliseconds) for real-time adaptive psychophysics.
- Acceptance: an amortized policy produces a next-design decision in sub-second wall-clock and beats a fixed/random design on cumulative EIG.
181.17.2 Adaptive Design in Psychology & Psychophysics#
- 181.17.2.1 Implement ADO (ADOpy pattern: psychometric function, delay discounting, choice-under-risk) and QUEST+/Psi adaptive staircases.
- 181.17.2.2 Reproduce the ADO efficiency result (delay-discounting ≥0.95 reliability in 10–20 trials) on a simulated participant as a regression test.
181.17.3 Adaptive Clinical Trials & Bandit Interventions#
- 181.17.3.1 Implement Bayesian adaptive-trial primitives: response- adaptive randomization, group-sequential boundaries (O'Brien-Fleming/ Pocock/Haybittle-Peto), BOIN/CRM dose-finding, and predictive-probability futility.
- 181.17.3.2 Implement master-protocol support (platform/basket/umbrella, I-SPY 2 / GBM AGILE graduation thresholds) with operating-characteristic simulation (FACTS/East pattern, Type-I control enforced).
- 181.17.3.3 Implement SMART designs and offline-RL DTR pre-training (EpiCare benchmark) for dynamic treatment regimes.
181.18 @phoebe/experiment-runtime — Authoring, Runtime & Closed-Loop Control#
181.18.1 Authoring & Frame-Precise Runtime#
- 181.18.1.1 Implement a Builder GUI (PsychoPy/lab.js pattern) compiling to a frame-precise browser runtime (PsychoJS/PixiJS/WebGPU, WASM) and a native lab runtime, with a plugin architecture (jsPsych pattern).
- 181.18.1.2 Implement timing instrumentation and calibration targeting
sub-millisecond lab precision and characterizing the macOS ~1-frame visual
onset delay.
- Acceptance: a timing test reports measured stimulus-onset/response variance and flags configurations that exceed the precision budget.
181.18.2 Closed-Loop & Robotic Execution#
- 181.18.2.1 Implement a closed-loop layer over LSL for brain-state- dependent stimulation (e.g., mu-rhythm-phase-triggered TMS, real-time fMRI neurofeedback).
- 181.18.2.2 Implement robotic/wet-lab execution backends: Opentrons Python Protocol API, Emerald Cloud Lab / Strateos APIs, automated patch-clamp, and DeepLabCut/SLEAP/Keypoint-MoSeq behavioural rigs.
181.19 @phoebe/subjects — Subject & Panel Management#
181.19.1 Recruitment & Data Quality#
- 181.19.1.1 Implement connectors to Prolific / CloudResearch / MTurk / Gorilla / SONA with per-panel quality metadata.
- 181.19.1.2 Implement the data-quality firewall: attention/IMC checks,
IP/geo uniqueness, bot honeypots, and LLM-agent contamination
detection with human-verified attestation.
- Acceptance: a simulated bot/LLM respondent that passes naive attention checks is still flagged by the adversarial layer.
181.19.2 Allocation & Eligibility#
- 181.19.2.1 Implement randomization/allocation services and longitudinal subject tracking with consent-state gating.
- 181.19.2.2 Implement eligibility/patient↔trial matching (TrialGPT pattern) as clinician-reviewed decision support.
181.20 @phoebe/analysis — Pipelines & Causal Inference#
181.20.1 Neuro/Behavioural Pipelines#
- 181.20.1.1 Implement containerized BIDS-App runtimes (fMRIPrep/ Nipype; FSL/SPM/AFNI) and MNE-Python/EEGLAB pipelines, orchestrated via Nextflow/Snakemake/CWL.
- 181.20.1.2 Enforce post-Eklund multiple-comparison correction (permutation/TFCE/voxel-FWE/FDR) as a non-optional gate on any cluster-inference result.
- 181.20.1.3 Implement mixed-effects modelling (lme4/brms) with maximal-vs-parsimonious random-effects guidance.
181.20.2 Causal Inference#
- 181.20.2.1 Implement causal discovery (PC/FCI/GES/LiNGAM/NOTEARS) and a DoWhy/causal-learn model→identify→estimate→refute flow.
- 181.20.2.2 Implement IV/Mendelian-randomization, mediation
(counterfactual NDE/NIE), and target-trial emulation.
- Acceptance: the refutation step runs before any effect is reported; a known confounded fixture is correctly flagged by a placebo/refutation test.
181.21 @phoebe/reproducibility — Preregistration, Multiverse & Provenance#
181.21.1 Preregistration & Multiverse#
- 181.21.1.1 Implement machine-readable auto-preregistration (locked, timestamped hypotheses & power analysis (G*Power/simr) + analysis DAG).
- 181.21.1.2 Make multiverse / specification-curve analysis a default execution mode (enumerate defensible pipelines by construction).
- 181.21.1.3 Implement preregistration-vs-execution auto-diff that
flags undisclosed researcher degrees of freedom.
- Acceptance: an analysis that deviates from its locked preregistration is automatically flagged with the specific divergence.
181.21.2 Provenance Ledger#
- 181.21.2.1 Implement an end-to-end provenance graph (DataLad/
rerunpattern, W3C PROV / RO-Crate) from stimulus → raw data → analysis → figure with one-command recompute. - 181.21.2.2 Implement FAIR/BIDS-conformant export and dataset DOIs.
181.22 @phoebe/knowledge-graph — Biolink KG & Ontology Service#
181.22.1 Graph Store & Ingestion#
- 181.22.1.1 Implement a Biolink-model-typed graph store ingesting SPOKE, PrimeKG, Hetionet, Monarch, DRKG, and NeuroKG.
- 181.22.1.2 Implement KG-embedding (TransE/RotatE/ComplEx) and heterogeneous-graph-transformer (TxGNN/PROTON pattern) services for link-prediction-based hypothesis support.
181.22.2 Ontology Service & The Crosswalk Phoebe Owns#
- 181.22.2.1 Host Cognitive Atlas, NIFSTD, Uberon, NeuroNames, NBO, MF/MFOEM, MeSH, SNOMED CT, DSM-5-TR, ICD-11 with versioned IRIs.
- 181.22.2.2 Publish the canonical machine-readable RDoC ontology (OWL/SKOS) — which does not exist anywhere today — with constructs mapped to Cognitive Atlas concepts and NBO behaviours.
- 181.22.2.3 Publish and maintain the DSM-5-TR ↔ ICD-11 Ch.06 ↔ MeSH
F03 ↔ SNOMED CT ↔ Cognitive Atlas ↔ RDoC crosswalk as an SSSOM
mapping graph (current best-in-class is a static third-party table).
- Acceptance: a query resolves a DSM-5-TR disorder to its ICD-11 code, MeSH term, SNOMED concept, and related RDoC constructs through the crosswalk, with mapping provenance attached.
181.23 @phoebe/literature — Ingestion & Claim Extraction#
181.23.1 Corpus Connectors#
- 181.23.1.1 Implement connectors to PubMed/E-utilities, OpenAlex, Semantic Scholar/S2ORC (+ SPECTER2 embeddings), Europe PMC Annotations, and openRxiv/PsyArXiv preprint feeds.
181.23.2 Claim & Evidence Extraction#
- 181.23.2.1 Implement claim/evidence extraction (SciFact/MultiVerS SUPPORTS/REFUTES/NOINFO & rationales) with entity linking to the ontology service.
- 181.23.2.2 Implement smart-citation typing (supporting/contrasting/
mentioning, scite pattern).
- Acceptance: extraction is evaluated on SciFact with reported F1; a fabricated claim with no supporting evidence is labelled NOINFO, not SUPPORTS.
181.24 @phoebe/evidence-synthesis — Living Reviews & GraphRAG#
181.24.1 Living Systematic Reviews#
- 181.24.1.1 Implement active-learning screening (ASReview pattern) + LLM majority-vote screening & multi-pass extraction with RCT detection (RobotSearch) and PICO & risk-of-bias extraction.
- 181.24.1.2 Implement contradiction detection across the claim graph and versioned, provenance-stamped living-review outputs.
181.24.2 GraphRAG with KG Verification#
- 181.24.2.1 Implement hybrid vector & graph retrieval (GraphRAG Leiden- community summaries; Local/Global/DRIFT search).
- 181.24.2.2 Implement KGARevion-style verification of every
LLM-generated triple against the grounded KG to suppress hallucination.
- Acceptance: a generated triple absent from / contradicted by the KG is filtered out before it reaches an answer.
181.25 @phoebe/datasets — FAIR Catalog & NWB/BIDS I/O#
181.25.1 Catalog & Access Tiers#
- 181.25.1.1 Implement a FAIR (RO-Crate/schema.org) catalog of UK Biobank, HCP, ABCD, ENIGMA, OpenNeuro, IBL, Allen ABC Atlas, ABIDE, ADNI, PPMI, AnswerALS, NDA, DANDI, BIL, EBRAINS, recording modality, scale, and access tier (open / DUA-LONI / enclave-Workbench / federated-only).
181.25.2 Standards I/O#
- 181.25.2.1 Implement first-class NWB (PyNWB/MatNWB; HDF5 & Zarr) and BIDS (incl. iEEG/microscopy/MRS) read/write, with DICOM→BIDS via HeuDiConv recording W3C-PROV provenance.
- 181.25.2.2 Implement IBL ONE/Alyx, DANDI, and OpenNeuro streaming
connectors (no-download access).
- Acceptance: a round-trip BIDS validation passes and a streamed DANDI NWB file reads without local download.
181.26 @phoebe/governance — Neuroethics, Neurorights & Privacy#
181.26.1 Neural-Data Governance#
- 181.26.1.1 Implement neural-data classification/tagging and a consent-management service (broad & dynamic consent, withdrawal, portability, deletion) engineered to Chile / Colorado HB24-1058 / California SB 1223 / Montana SB 163 / Connecticut SB 1295 & the UNESCO 2025 neurotechnology standard, plus HIPAA and GDPR Art. 9.
- 181.26.1.2 Implement HIPAA de-identification (Safe Harbor 18-identifier & Expert Determination) and mandatory neuroimaging defacing (fsl_deface/mri_reface) — given the 83% top-1 face-reID risk.
181.26.2 Privacy-Preserving Compute & Provenance#
- 181.26.2.1 Implement federated learning (NVIDIA FLARE / COINSTAC for neuroimaging) and differential privacy (OpenDP) for multi-institution analysis; SMPC+HE for genomics (SF-GWAS pattern).
- 181.26.2.2 Implement dbGaP-style controlled access (DAR → DAC → DUC)
and end-to-end audit trails.
- Acceptance: a neural-data export for an unconsented purpose is hard-blocked; every access is logged immutably.
181.27 @phoebe/regulatory — FDA SaMD Lifecycle & Validation Harness#
181.27.1 SaMD Lifecycle#
- 181.27.1.1 Implement SaMD lifecycle support (510(k)/De Novo/PMA by risk; IMDRF framework) with a Predetermined Change Control Plan (PCCP)-aware model-update pipeline (ship updates without full re-submission).
- 181.27.1.2 Implement model-card / data-lineage documentation, bias/ equity testing (race-stratified, per the npj 2025 finding), and model- drift detection with rollback criteria and post-market monitoring (TPLC).
181.27.2 Evidence & Registration#
- 181.27.2.1 Implement a clinical-evidence/RCT pipeline
(Therabot/Limbic- grade) and ClinicalTrials.gov / WHO ICTRP registration
emission.
- Acceptance: a model update that fails bias/drift thresholds is blocked from promotion and triggers rollback; the validation record is auditable.
181.28 @phoebe/safety — Crisis & Compliance Backbone#
181.28.1 Crisis Handling#
- 181.28.1.1 Implement crisis routing (988 Suicide & Crisis Lifeline / Samaritans / findahelpline) and suicidality-escalation workflows triggered by C-SSRS / PHQ-9 item-9 / conversational detection.
- 181.28.1.2 Implement age-gating (post-Character.AI) and mandatory AI-disclosure (Utah HB 452).
181.28.2 Jurisdiction-Aware Compliance#
- 181.28.2.1 Implement jurisdiction-aware feature flags that disable autonomous-therapy features for Illinois/Nevada users and enforce human-in-the-loop where law requires it.
- 181.28.2.2 Implement a human-in-the-loop checkpoint framework that any
patient-facing or human-subjects autonomous action must pass.
- Acceptance: a high-risk query is never answered autonomously; the backbone forces escalation and records the human handoff.
181.29 @phoebe/ai — Autonomous Closed-Loop Neuroscientist#
181.29.1 Closed-Loop Orchestration#
- 181.29.1.1 Implement the supervisor that chains hypothesis → adaptive design → IRB/AI-ethics gate → execution → analysis → theory update → re-rank, specializing the Phase 178 substrate for brains/behaviour.
- 181.29.1.2 Implement the theory/model-update step: formalize findings back into the knowledge graph / causal world-model with provenance and contradiction detection (the hardest unsolved piece).
181.29.2 Reliability Guardrails#
- 181.29.2.1 Implement hallucination/fabrication detection on every autonomous analysis step and human-in-the-loop checkpoints for all human-subjects actions.
- 181.29.2.2 Implement the fail-closed AI-IRB ethics gate: real-time
machine-readable protocol checking against Common Rule / Belmont / ICH
E6(R3) before any agent enrolls a human, with continuous AE monitoring and
automatic stopping-rule enforcement.
- Acceptance: the engine cannot enroll a human subject without passing the AI-IRB gate and a human sign-off; a protocol violation halts the loop and raises an alert rather than proceeding.
181.30 @phoebe/apps — Application Suite#
181.30.1 Researcher & Clinician Surfaces#
- 181.30.1.1 Implement the Researcher Workbench (hypothesis → design
→ run → analyze → prereg/reproduce) as a Next.js app under
libs/phoebe/apps/. - 181.30.1.2 Implement the Clinician Console (assessment, MBC dashboards, treatment-matching, copilot/scribe) with human-sign-off gates.
181.30.2 Real-Time & Participant Surfaces#
- 181.30.2.1 Implement the Neural-Interface Control Room (live BCI decode/stim, signal-quality and device-health telemetry, latency SLAs).
- 181.30.2.2 Implement the Knowledge Explorer (KG/literature/evidence browsing with verified-citation answers).
- 181.30.2.3 Implement the Participant/Patient app (consent, EMA, phenotyping opt-ins, therapeutic content with crisis safety).
181.31 @phoebe/integration — Cross-Domain Hub#
181.31.1 Adapters#
- 181.31.1.1 Implement the Iris adapter consuming
@iris/bci-*/@iris/multimodal-bcidevice, intent-prediction, and privacy primitives. - 181.31.1.2 Implement the Aphrodite adapter consuming
@aphrodite/biometric-engine(incl. EEG) biosignal acquisition. - 181.31.1.3 Implement the Psyche adapter consuming affective/ sentiment signals and the conversational stack for therapeutic delivery.
- 181.31.1.4 Implement the Phase 178 / Nous adapter consuming the autonomous-research substrate and ML platform, and the Kalika adapter consuming research-agent & computational-notebook infrastructure.
- 181.31.1.5 Implement adapters to Mnemosyne (memory), Maat
(ethics/governance), Veritas (verification), Themis (IP), and
Shared (platform infra), plus integration tests covering every adapter
(minimum 200 test cases).
- Acceptance: no Phoebe package re-implements Iris/Aphrodite/Psyche device or affect code; all such access flows through these contract-bound adapters.
Phase 181 Summary — Phoebe Domain#
| Library | Subsections | Tasks | Description |
|---|---|---|---|
| Infrastructure | 181.1.1 – 181.1.3 | 42 | Library structure, data model/persistence, API gateway & topology |
| @phoebe/core | 181.2.1 – 181.2.2 | 7 | Neural/signal & clinical/experimental primitives, state machines |
| @phoebe/neural-io | 181.3.1 – 181.3.2 | 7 | Multimodal acquisition & LSL/XDF sub-ms time-sync |
| @phoebe/signal | 181.4.1 – 181.4.2 | 6 | Real-time DSP, ICA/CSP/xDAWN/Riemannian feature extraction |
| @phoebe/decoding | 181.5.1 – 181.5.3 | 9 | Decoder runtime, neural foundation models, cross-subject adaptation |
| @phoebe/neuromodulation | 181.6.1 – 181.6.2 | 6 | Closed-loop aDBS/RNS/tFUS/TI, bidirectional read–write, safety governor |
| @phoebe/connectomics | 181.7.1 – 181.7.2 | 6 | FFN segmentation, CAVE proofreading, petascale graph serving |
| @phoebe/simulation | 181.8.1 – 181.8.2 | 5 | NEST/NEURON/Arbor/Brian2 over SONATA, multi-scale co-sim |
| @phoebe/brain-models | 181.9.1 – 181.9.2 | 5 | Brain foundation models & circuit/region digital twins (fail-loud) |
| @phoebe/neuro-analysis | 181.10.1 – 181.10.2 | 6 | SpikeInterface/Kilosort, LFADS/CEBRA, manifolds, identifiability |
| @phoebe/assessment | 181.11.1 – 181.11.2 | 5 | PHQ-9/GAD-7/PCL-5/C-SSRS & DSM-5-TR/ICD-11 computable coding |
| @phoebe/mbc | 181.12.1 | 2 | Measurement-based & collaborative care registry |
| @phoebe/phenotyping | 181.13.1 – 181.13.2 | 5 | Passive sensing, EMA, voice biomarkers, contextual-bandit JITAI |
| @phoebe/comp-psychiatry | 181.14.1 – 181.14.2 | 4 | RL/DDM/active-inference, biotypes, treatment-matching (advisory) |
| @phoebe/therapeutics | 181.15.1 – 181.15.2 | 5 | CBT/DBT/ACT/MI content, anti-sycophancy, clinician copilot/scribe |
| @phoebe/hypothesis | 181.16.1 – 181.16.2 | 4 | Co-Scientist generate/debate/rank, graph-novelty & GNN-feasibility |
| @phoebe/design | 181.17.1 – 181.17.3 | 7 | BOED/DAD/ADO, Bayesian adaptive trials, SMART/JITAI optimizers |
| @phoebe/experiment-runtime | 181.18.1 – 181.18.2 | 4 | Frame-precise authoring/runtime & closed-loop & robotic execution |
| @phoebe/subjects | 181.19.1 – 181.19.2 | 4 | Panel connectors, data-quality firewall, eligibility matching |
| @phoebe/analysis | 181.20.1 – 181.20.2 | 5 | BIDS-App/MNE pipelines, multiple-comparison gate, causal inference |
| @phoebe/reproducibility | 181.21.1 – 181.21.2 | 5 | Auto-preregistration, multiverse default, provenance ledger |
| @phoebe/knowledge-graph | 181.22.1 – 181.22.2 | 5 | Biolink KG, ontology service, RDoC OWL & DSM↔ICD crosswalk (greenfield) |
| @phoebe/literature | 181.23.1 – 181.23.2 | 3 | PubMed/OpenAlex/S2 ingestion, SciFact claim/evidence extraction |
| @phoebe/evidence-synthesis | 181.24.1 – 181.24.2 | 4 | Living systematic reviews, GraphRAG & KGARevion verification |
| @phoebe/datasets | 181.25.1 – 181.25.2 | 4 | FAIR catalog (UKB/HCP/ABCD/IBL/Allen…), NWB/BIDS/DICOM I/O |
| @phoebe/governance | 181.26.1 – 181.26.2 | 4 | Neurorights/HIPAA/GDPR, defacing, federated & DP compute, controlled access |
| @phoebe/regulatory | 181.27.1 – 181.27.2 | 3 | FDA SaMD & PCCP, bias/drift testing, RCT evidence & registration |
| @phoebe/safety | 181.28.1 – 181.28.2 | 4 | 988/Samaritans crisis routing, age-gating, jurisdiction-aware HITL |
| @phoebe/ai | 181.29.1 – 181.29.2 | 4 | Closed-loop neuroscientist, theory update, fail-closed AI-IRB gate |
| @phoebe/apps | 181.30.1 – 181.30.2 | 5 | Researcher Workbench, Clinician Console, Control Room, Explorer, Patient |
| @phoebe/integration | 181.31.1 | 5 | Iris/Aphrodite/Psyche/Phase-178/Kalika & Mnemosyne/Maat/Veritas/Themis |
| TOTAL | 181.1 – 181.31 | 189 | Brain-computer interface, neuroscience, clinical psychology & autonomous discovery |