Kai9987kai/Sonapse-LAB

SYNAPSE: Developmental Connectome Intelligence & Weight-Space Lab (SAIL Edition)

Current evidence: SYNAPSE is a CPU research workbench with an integrated executed-transition learner and many standalone experiments. Module availability does not establish embodied capability or biological fidelity. Planning is experimental and off by default. Read the design recovery audit, Antigravity contract audit, neural graft training specification, connectome source & verification guide, and compute, training & connectome pilot report for restored contracts, verified datasets, and measured pilot outcomes.

A unified, CPU-first embodied cognitive research platform combining signed recurrent neural graphs, a frozen Poseidon Tidal MoE DAgger controller, continuous online connectome neurogenesis, interoceptive homeostatic drive regulation, MiMoMix dual-process cognitive deliberation, experience-grounded prediction, and direct weight-space programming with inherited recursive self-improvement (RSI).

Now featuring the Sonapse Adaptive Intelligence Laboratory (SAIL) research track: utility-driven continual backpropagation, trainable student-teacher sleep consolidation, universal fast-weight memory (uTTT & Titans associative learning), recurrent multi-step world models with MPC planning ($H \in {1, 3, 5, 10}$), verifiable ReLIFT curriculum challenges, and causal connectome structural evolution.

Live telemetry, real-time 3D anatomical connectome visualization, and zero-warping virtual world observatory are served via a local web interface and audited CLI tools.

Architecture

flowchart TD
    W[Complex Dynamic Virtual Ecology] -->|Observations, Threat, Friction, Light| S[Sensory Stream: Rangefinder, Contrast, Cues]
    S --> A[GraftBrain Recurrent Rate Connectome]
    A <-->|Residual Bridges| T[Frozen Poseidon Tidal MoE Controller]
    
    %% Astrocytic and Homeostatic Neuromodulation
    H[Homeostatic Drive Engine] -->|Vitals: Energy, Stamina, Health, Threat, Hunger| N[Neuromodulatory Field: DA, 5-HT, OA, ACh]
    N --> A
    N --> G[Astrocytic Calcium Reservoir]
    G --> A
    
    %% Fast Sensorimotor vs Deliberative Cognition
    A --> M[Fast Motor Actuators: Drive, Turn, Reach, Height, Grip]
    M --> R[MiMoMix Dual-Process Reasoner]
    A --> R
    T --> R
    R -->|System 1 Fast Reflex or System 2 ACT Mental Search| D[Deliberated Action]
    
    %% Experience & Outcome Planning
    E[Experience World Model Ensemble] --> P[Bounded 1–5 Step Experience Planner]
    SM[Surprise Episodic Outcome Memory] --> P
    D --> P
    P --> X[Executed Action]
    X -->|Dynamic Throttle by Stamina Fatigue| W
    
    %% Structural & Weight-Space Plasticity
    A -->|Online Welford Covariance| NG[Archimedes Connectome Neurogenesis]
    N -->|Dopamine RPE & Astrocyte Calcium Gating| NG
    NG -->|Synaptogenesis & Apoptosis under Dale's Law| A
    
    %% Weight Space Programming & RSI
    X --> EXP[Rollout Experience & Causal Diagnostics]
    EXP --> WSP[Weight-Space Programmer & Transaction Ledger]
    WSP -->|Validated Inactive Patch & Atomic Rollback| A
    WSP --> RSI[Inherited Meta-Improvement Policy]

System Components & Evidence Boundaries

Subsystem	Architectural Role	Governing Dynamics & Evidence Boundary
Recurrent Connectome Substrate	Fast sensorimotor rate dynamics	Signed base connectivity across FlyWire, MANC, and Optic Lobe profiles, 8 inter-module graft sites, Dale's law compliance, and ellipsoid body compass bump.
Tidal MoE Controller	Pretrained goal & intent donor	Pinned Poseidon Tidal core (tidal_core.pt, SHA-256 a053104a...), 4 experts, 128 hidden width, 2 recurrent steps, 6 categorical intent classes (rest, forage, drink, shelter, explore, flee). Kept frozen during ordinary execution.
Archimedes Neurogenesis Engine	Dynamic structural plasticity	Continuous online Welford covariance tracking; apoptosis of inactive synapses ($
Supermix Homeostatic Drives	Interoceptive physiological regulation	Continuous tracking of 6 bodily vitals (Energy, Stamina, Health, Threat, Exposure, Hunger); dynamic neuromodulator coupling ($\text{DA}, 5\text{-HT}, \text{OA}, \text{ACh}$); physical forward thrust limits under stamina fatigue.
MiMoMix Cognitive Core	Dual-process cognitive deliberation	Epistemic trigger monitoring free energy, surprisal, and hazard proximity; fast reflexive System 1 intuition vs. multi-step System 2 Adaptive Computation Time (ACT) mental rollouts with attention sink null-slots and ponder penalties.
Experience World Model	Prequential dynamics prediction	3-member deterministic bootstrap RLS ensemble predicting sensor summary deltas and reward conditioned on live graph/Tidal representations. Disagreement and leverage provide uncalibrated heuristics.
Surprise Episodic Memory	Error-gated outcome storage	Bounded capacity (128 entries); selective error/coverage admission; pure distance and recency weighted retrieval; non-mutating reads.
Complex Dynamic Ecology	Embodied virtual simulation	Continuous 2D/3D physics environment featuring autonomous moving predator drones with threat auras, variable-friction terrain zones (ice, mud, sand), dynamic oscillating/locked gates, and diurnal ambient illumination cycles.
Weight-Space Programming	Safe transactional parameter editing	Bounded edit packets, parent hash validation, candidate evaluation, exact state rollback, and closed-loop automated diagnosis/repair.
Recursive Self-Improvement (RSI)	Meta-level learning of learning	Versioned improvement policies $\phi_g$, support/query task isolation, shared archive controls, candidate meta-validation, and fresh-learner held-out evaluations.
Zero-Warping Observatory & HUD	Real-time scientific visualization	Isotropic 3D isometric and top-down radar views with mathematical aspect ratio locking; real-time 3D anatomical connectome observatory with OrbitControls; direct WASD teleoperation; real-time Vitals HUD meters.
Continual Backpropagation	Utility-guided plasticity & replacement	Continuous utility tracking $u_i(t) = \beta u_i(t-1) + (1-\beta)
Trainable Sleep Engine	Student-teacher consolidation	Minimizes $\mathcal{L}{\text{sleep}} = \lambda_1\mathcal{L}{\text{task}} + \lambda_2\mathcal{L}{\text{distill}} + \lambda_3\mathcal{L}{\text{replay}} + \lambda_4\mathcal{L}_{\text{reg}}$ with SHY synaptic downscaling and evaluation rollback gating.
Universal Fast-Weight Memory	Test-time associative learning	Shared associative matrix $M \in \mathbb{R}^{d_k \times d_v}$ with self-supervised inner loss $\mathcal{L}_{\text{inner}}$, surprise-gated outer-product updates, and surprise gating.
Recurrent Multistep Planner	Learned dynamics & MPC control	Recurrent GRU state transition, reward, observation, and uncertainty heads; MPC action trajectory optimization over horizons $H \in {1, 3, 5, 10}$ with penalty for epistemic variance.
Verifiable Curriculum Engine	Verifiable progression & ReLIFT	Procedural sequential subgoals (CUE_DISCOVERED, GATE_TRIGGERED, BARRIER_PASSED, TARGET_REACHED) with cryptographic SHA-256 process receipts.
Causal Connectome Evolution	Interventional structural changes	Paired counterfactual rollouts measuring $\Delta J(m) = \mathbb{E}[R \mid do(m)] - \mathbb{E}[R \mid do(m_0)]$ penalized by wiring, compute, and catastrophic forgetting costs.
Tripartite Uncertainty Decomposition	Calibrated risk & OOD detection	Epistemic ensemble variance + Heteroscedastic Gaussian NLL aleatoric head + Mahalanobis OOD latent distance; Expected Calibration Error (ECE) and monotonic risk-coverage analysis.
Action-Grounded Efference Self-Model	Sensorimotor agency & reafference	Forward internal model predicting sensory delta $\Delta \hat{s}t = M{\text{efference}}(s_t, a_t)$; decomposes discrepancy into Self-Authorship Index $A_t \in [0, 1]$, External Disturbance $E_t$, and Actuator Slip $F_t$; benchmarked against matched scientific controls (scrambled_timing, shuffled_action, ablated).
Supermix Research Fabric & 6-Arm Study	Cross-ecosystem contracts & comparative study	Typed, versioned contracts across Poseidon, NEUROGRAFT, Tether, FAULTLINE, and GhostInTheMachine with cryptographic SHA-256 evaluation receipts; standard 6-arm curriculum study (frozen_actor, online_rl, actor_replay, continual_backprop, connectome_evolution, composed_system) evaluating retention ($\ge 95%$), forgetting, and latency.
Plasticity Monitor & Reusable Skills	Representation rank & macro discovery	Continuous representation Effective Rank $\text{erank}(X)$, Stable Rank $\text{srank}(X)$, dead neuron monitoring, and TESSERA-inspired macro skill discovery.
Verifier-Gated Bounded RSI	Bounded self-improvement & failure archive	AIDE² / DGM / AlphaEvolve pattern: AST range checks, deterministic tests, support/query evaluation, LCB dominance ($LCB(\Delta) > \epsilon$), canary regression gating ($\le 1%$), and immutable Failure Archive.
World Model v2 & CEM Planner	Standalone neural dynamics experiment	$M=3$ latent ensemble and CEM demonstrator. The live actor instead uses the persisted experience model below. CLI demonstrations do not record fabricated realized outcomes.
Experience MPC	Integrated short-horizon prediction and control	Pure deterministic ensemble rollouts over horizons 1–5, bounded local action changes, fixed neural context, repeated-baseline comparator, uncertainty/support abstention, and a 256 ensemble-transition work budget. No terminal/value head or calibrated safety claim.
9-Arm Connectome Causal Assay	Biological vs matched controls	Causal benchmark evaluating real Drosophila wiring against 8 matched nulls (degree-, sign-, community-rewired, random, learned, unsigned, local-only, brain-cord motif) calculating formal Cohen's $d$.
Coconut Latent Deliberation	Tokenless continuous recurrence	Continuous hidden-state thought recurrence ($h^{\text{think}}_{i+1} = F(h^{\text{think}}_i, z_t)$), dynamic ponder halting, and anytime compute-depth curves.
Masked Diffusion Repair Sidecar	Non-autoregressive repair & infill	Iterative masked diffusion repair of failing code, configuration, or action trajectory constraints with executable verifier in the loop.
Machine-Verifiable DPO Engine	Direct preference optimization	First-party machine-checked preference pairs (safety, return, unit tests) optimized via DPO without ungrounded self-judging loops.
Unified Evaluation Scorecard	Standalone metric calculations	Combines supplied returns, compute, Brier/ECE and failure inputs. Its CLI fixture is explicitly illustrative; it does not measure deployment performance.
Landmark Connectome Catalog (8 Datasets)	Whole-animal multi-connectome substrate	Full integration of 8 datasets: MaleCNS v1.0, BANC, FlyEM Optic Lobe, FANC, FlyWire FAFB v783, MANC, Larval Drosophila, and Hemibrain with regional neuromodulation.
Multi-Connectome Hub	Scalable cross-region routing	Unifies brain, optic, and cord neuropils with 93 biological interface sites, high-capacity scaling (>40,000 neurons/sec on CPU), and transactional weight-space editing.
Frontier Apex Omnibus Mode	Full architectural & environmental integration	Single unified command python -m synapse.cli omnibus running all 11 complex environments (forage, maze, push, delayed_match, ecosystem, predator_evasion, dynamic_arena, puzzle_chamber, apex_megasite, stealth_labyrinth, kinetic_foundry) across 2,000 neurons (667/667/666) with efference self-model, uncertainty calibration, plasticity rank monitoring, and cryptographic SHA-256 evaluation receipts.
Executed-Transition Neural Graft	Bounded residual actor, TD critic & sensory predictor	Zero-initialized residual policy (synapse/neural_graft.py) conditioned on 155-d observations, 64-d graph/donor features, 5-d baseline actions, and an optional 128-neuron MaleCNS sparse connectome prior (load_programmed_graft). Supports parallel rollout pretraining (pretrain_runtime) and online clipped policy optimization without mutating actor identity during auxiliary pretraining.
Verified Neuron-Level Connectome Caches	Checksum-verified local connectome ingestion	Independent source-backed caches (synapse/connectome_assets.py) for malecns, flywire, manc, optic-lobe, banc, hemibrain, and larval, keeping catalog metadata, cell-type aggregates, induced subgraphs, and synthetic proxies explicitly distinct.
Live Weight-Space Screen & Bounded RSI	Matched cloned-world numeric patch screening	Screens numeric weight edits (screen_live_patch, run_live_rsi in synapse/live_programming.py) on isolated cloned worlds across query and retention seeds before atomic application.

Run Locally

Requires Python 3.10+ with PyTorch CPU, NumPy, and SciPy:

New workbench sessions use 2,000 total neurons, allocated 667/667/666 across the three default sources (FlyWire, MANC, Optic Lobe). The graph view includes all 2,000. Use python -m synapse --total-neurons 2000 serve or set Total neurons in the session controls. The older --nodes 16 option still selects 16 per source for small experiments and compatible checkpoints. Run the 11-stage omnibus curriculum with python -m synapse.cli omnibus.

python -m pip install -e .
python -m pytest -q
.\launch.ps1 -Port 8765

Open the local workbench at http://127.0.0.1:8765 or launch via Run-Synapse.cmd.

Interactive Workbench Features

• Zero-Warping World Views: True isotropic aspect ratio rendering for both 3D isometric projection and top-down radar world views, ensuring square cells remain square under all window resizes.
• Complex Ecology Telemetry: Real-time visualization of terrain friction zones (ice, mud, sand), moving predator drones with hazard danger rings, dynamic gates, day/night ambient lighting, and resource targets.
• Direct Keyboard Teleoperation: Control the embodied agent in real-time using WASD or Arrow Keys, with Spacebar for object grab/place.
• 3D Anatomical Connectome Observatory: Interactive 3D point-cloud and synaptic edge visualization with OrbitControls (left-click rotate, right-click pan, scroll zoom) and click-to-inspect individual neuron activity.
• Homeostatic Vitals & Cognition HUD: Real-time meters for Energy, Stamina, Health, Threat, Exposure, and Hunger, paired with MiMoMix System 1 / System 2 deliberation badges and ACT ponder steps.
• Scientific Assay Panels: Causal Circuit Atlas, Weight-Space Hypernetwork synthesis, Flagship evolutionary benchmarks, State Archives, and Audited Capabilities Explorer.
• Counterfactual Plan Inspector: Select a horizon, inspect pure forecasts, compare candidate returns and uncertainty, and see why the planner retains its baseline. Previews never advance the world or fit the model.
• Learning Contracts: Manual actions can train the same executed-outcome learner. Drive/turn coordinates and fatigue limits apply consistently; frozen evaluation preserves learned topology and structural statistics. Tool memory belongs to each checkpointed runtime.

Command Line Interface (CLI)

SYNAPSE provides comprehensive CLI commands for headless execution, experiments, and inspection:

# Inspection & Baseline Execution
python -m synapse.cli --nodes 16 status
python -m synapse.cli --nodes 16 run --steps 100
python -m synapse.cli --nodes 16 train --steps 100 --save trained_agent

# Supermix Beyond Capabilities
python -m synapse.cli --nodes 16 neurogenesis
python -m synapse.cli --nodes 16 homeostasis
python -m synapse.cli --nodes 16 mimomix --ponder 4
python -m synapse.cli --nodes 16 ecology --task ecosystem

# Audited Upstream Capabilities (142 audited capabilities across 47 families)
python -m synapse.cli capabilities --id structural.neurogenesis
python -m synapse.cli capabilities --id motivation.homeostasis
python -m synapse.cli capabilities --id reasoning.mimomix
python -m synapse.cli capabilities --id world.complex_ecology
python -m synapse.cli capabilities --id learning.continual_backprop
python -m synapse.cli capabilities --id memory.fast_weights

# SAIL 6-Upgrade Research & Development Suite
python -m synapse.cli continual-backprop
python -m synapse.cli sleep-train --sleep-steps 10
python -m synapse.cli fast-memory
python -m synapse.cli multistep-plan --horizon 3
python -m synapse.cli curriculum-verify
python -m synapse.cli causal-growth
python -m synapse.cli flagship-milestone --milestone-steps 20

# Deep Research Architecture & Control Experiments
python -m synapse.cli rsi-verify --patch-type config
python -m synapse.cli world-model-cem --cem-horizon 5
python -m synapse.cli connectome-assay --assay-trials 3
python -m synapse.cli coconut-think --think-depth 4
python -m synapse.cli diffusion-repair --repair-iters 3
python -m synapse.cli dpo-align --pairs 5
python -m synapse.cli unified-eval --eval-task forage

# Landmark Connectome Datasets & High-Capacity Scaling
python -m synapse.cli connectome-inventory
python -m synapse.cli connectome-datasets
python -m synapse.cli multi-connectome --nodes-per-source 32
python -m synapse.cli scale-neurons --targets 64 128 256 512

# Neural Graft Pretraining, Active Training & Bounded Live RSI
python -m synapse.cli --compute-threads 4 --neural-graft --programmed-connectome pretrain --steps 256 --updates 64 --batch-size 128 --workers 2 --save pretrained-graft
python -m synapse.cli --compute-threads 4 --neural-graft --active-training --programmed-connectome --connectome-mode prefer_authentic train --steps 64 --save online-graft
python -m synapse.cli --compute-threads 4 live-rsi --steps 24 --candidates 2 --save rsi-screened

# Automated Weight-Space Programming & Closed-Loop Repair
python -m synapse.cli --nodes 16 autoprogram --diagnostic-episodes 3 --eval-episodes 3 --save repaired_agent
python -m synapse.cli --nodes 16 transplant --donor-task forage --test-task delayed_match
python -m synapse.cli --nodes 16 metamorphic
python -m synapse.cli --nodes 16 spectral
python -m synapse.cli --nodes 16 hyperprogram --target motor.weight

# Measured Adaptation Experiments
python -m synapse.cli --output outputs evidence --name my-pilot --seeds 7 11 19 --tasks forage push --support-steps 64 --query-steps 64
python -m synapse.cli adaptation-verify outputs/adaptation-studies/my-pilot
python -m synapse.cli evidence --receipt outputs/adaptation-studies/my-pilot

# Paired actual-world planning comparison (actor, horizon 1, horizon 3)
python -m synapse.cli --nodes 16 --output outputs planning-study --name planning-pilot --seeds 23 29 31 --tasks forage push ecosystem --support-steps 96 --query-steps 48 --horizons 1 3
python -m synapse.cli planning-verify outputs/planning-studies/planning-pilot

Measured Experiments & Verification

Empirical evidence studies run on actual native and complex environments with strictly separated support and query worlds:

• Support Adaptation: Shared exploratory support trajectories condition the learner.
• Query Freeze: All model and memory learning is strictly frozen during query episodes.
• Replay & Audit Receipts: Full cryptographic digests, parameter matrices, RNG states, and execution timings are recorded in atomic JSON receipts.

Planning pilots fit a shared learner on actual support trajectories and compare frozen actors in independent paired query worlds. Saved checkpoints and receipts bind raw transitions, source fingerprints and arithmetic. Verification detects inconsistency; it does not authenticate a consistently rewritten receipt or re-execute the simulation. Negative differences and abstentions remain visible.

Sleep consolidation uses observation-conditioned reset-state imitation and accepts only candidates passing a caller-supplied finite retention screen. Missing or failed screens reject the candidate. Untrained predictive/MiMoMix action corrections are opt-in through experimental_cognition; older checkpoints default to disabled. The original flagship continual benchmark remains a labeled synthetic surrogate.

The integrated planner draws on PETS for ensemble trajectory propagation and MBPO for limiting model rollouts. It uses online linear dynamics and deterministic search; it implements neither paper's complete algorithm or reported guarantees. An online Temporal Difference value critic $V(s)$ resolves finite-horizon truncation bias when value_guided_planning is selected.

Frontier Capabilities & Research Additions

The runtime provides toggleable research implementations of recent frontier techniques:

• Dormant Expert Dynamic Lifecycle (synapse/dormant_lifecycle.py): Recruits dynamic capacity when graph delta exceeds novelty thresholds. Experts advance through validated states (DORMANT $\to$ INITIALIZING $\to$ ACTIVATING $\to$ ACTIVE $\to$ CONSOLIDATING $\to$ RETIRED) with orthogonal initialization, gradual activation ramping, directional cosine contribution testing, and native parameter distillation with slot recycling.
• Value-Guided MPC Critic (synapse/experience.py, synapse/experience_planner.py): Eliminates horizon truncation bias in short MPC trajectories ($H \in [1, 5]$) by learning an online TD critic $V(s) \approx \mathbb{E}[\sum_{\tau=0}^\infty \gamma^\tau r_\tau]$ alongside the transition ensemble.
• Two-Channel Evoked Readout (synapse/agent.py): Separates resting manifold activity $m = O(r_{\text{null}})$ from task-evoked perturbation $d = O(r_{\text{actual}}) - O(r_{\text{null}})$ with input-dependent gating $\alpha$, exposing full telemetry for persistent state dynamics.
• Continual Backpropagation (synapse/continual_backprop.py): Applies Dohare et al. (Nature 2024) selective unit reinitialization based on feature utility to prevent loss of plasticity during continual adaptation.
• Universal Fast-Weight Memory (synapse/fast_weight_memory.py): Implements dual-speed associative learning (Titans / uTTT) with surprise-gated inner-loop updates.
• Continuous Latent Deliberation (synapse/latent_reasoner.py): Applies Coconut continuous latent-space thinking when variational free energy or prediction error exceeds bounds.

All options are configurable in SynapseConfig, controllable via the web workbench HUD and settings, and fully preserved across mid-episode state checkpoints.

Upstream Lineage & Research Grounding

SYNAPSE unifies innovations from:

• GraftBrain: Signed recurrent neural dynamics, biological neuropil topologies, mushroom body associative learning, and embodied planar physics.
• Supermix-Poseidon: Trained Tidal MoE DAgger controller, predictive horizons, causal intervention analyzers, and multi-expert routing.
• Supermix-Expanse: Function-preserving dormant experts, native consolidation, evoked graph readouts, and male-CNS connectome data.
• Archimedes v93: Continuous covariance-driven structural neurogenesis, synaptogenesis, and apoptosis.
• Supermix Beyond: Interoceptive homeostatic drive regulation, MiMoMix dual-process cognitive deliberation, and complex multi-hazard virtual ecology.

All source code is released under the Apache 2.0 license. Detailed source manifests and hashes are preserved in synapse/source_manifest.json.