sovr610
@sovr610
Public Skills
Continual Learning Guard (EWC / SI / Progressive Networks / Replay)
by sovr610
This skill should be used when the user asks to "prevent catastrophic forgetting", "elastic weight consolidation", "EWC regularization", "progressive networks", "continual learning strategy", "knowledge distillation for retention", "replay buffer memory", "task boundary detection", "fisher information matrix", "synaptic intelligence", "PackNet pruning", "memory-aware synapses", "add continual learning guard", "implement EWC penalty", "add experience replay", "implement progressive columns", "add Fisher diagonal computation", "implement reservoir sampling", "add knowledge distillation loss", "implement task-free continual learning", "add online EWC", "implement generative replay", "add PackNet iterative pruning", "implement synaptic intelligence path integral", or mentions catastrophic forgetting, continual learning, lifelong learning, sequential task training, knowledge retention, task interference, Fisher information regularization, or Phase 8 continual-learning pipeline in the cognitive architecture.
Engram Conditional Memory (N-gram Hash Lookup + Offload/Prefetch)
by sovr610
This skill should be used when the user asks to "implement engram memory", "add N-gram hash lookup", "implement tokenizer compression", "add engram layer", "implement CPU offload embeddings", "add async prefetch", "implement multi-head hashing", "add context-aware gating", "implement depthwise causal conv", "add engram encoder", "implement hash embedding retrieval", "add collision mitigation", "implement offloadable embedding", "add prefetch scheduler", "implement engram augmented layer", "add residual fusion", "implement RMSNorm gating", "add engram telemetry", "implement streaming N-gram cache", "add prime-sized hash tables", "implement tokenizer equivalence merging", or mentions engram memory, N-gram hashing, deterministic addressing, CPU offload + prefetch, tokenizer compression, context-aware gating, depthwise causal convolution, hash embedding, encoder-competition mode, layer-augmentation mode, or DeepSeek-style conditional memory in the cognitive pipeline.
Distributed Scaling
by sovr610
This skill should be used when the user asks to "scale training to multiple GPUs", "set up distributed training", "configure DDP", "use FSDP", "shard the model", "add gradient accumulation", "run with torchrun", "multi-node training", "rank-aware data loading", "distributed checkpointing", "scale to 7B parameters", "reduce GPU memory usage", "configure mixed precision distributed", or needs guidance on DistributedDataParallel, FullyShardedDataParallel, multi-GPU orchestration, or scaling the brain_ai system beyond single-GPU.
Global Workspace Competition + Broadcast + Working Memory with Ignition Dynamics
by sovr610
This skill should be used when the user asks to "implement global workspace", "add workspace competition", "implement ignition dynamics", "add broadcast adapters", "implement working memory", "add CfC/LTC memory", "implement attention competition", "add token staging", "implement slot construction", "add deterministic tie-breaking", "implement iterative rounds", "add convergence detection", "implement ignition gate", "add lock-in prevention", "implement broadcast to temporal", "add broadcast to symbolic", "implement broadcast to decision", "add workspace slots", "implement capacity-limited selection", "add novelty scoring", "implement winner decay", "add slot dropout", "implement ignition cooldown", "add feedback collection", "implement workspace state persistence", "add truncated BPTT support", "implement GRU fallback for workspace", "add workspace telemetry", or mentions global neuronal workspace theory, multi-modal token competition, ignition threshold dynamics, workspace broadcast packets, or CfC/LTC working memory in the cognitive pipeline.
Meta-Learning Suite (MAML/FOMAML/Reptile + MAML++ Enhancements)
by sovr610
This skill should be used when the user asks to "implement meta-learning", "add MAML inner loop", "implement FOMAML", "add Reptile", "implement MAML++", "add per-layer per-step learning rates", "implement LSLR", "add multi-step loss", "implement episodic sampling", "add few-shot learning", "implement inner-loop optimizer", "add second-order meta-gradients", "implement torch.func inner loop", "add higher library support", "implement adaptation curves", "add AUAC metrics", "implement meta-checkpoint format", "add derivative-order annealing", "implement episode sampler", "add Omniglot dataset", "add mini-ImageNet splits", "implement Phase 7 runner", "add meta-training loop", "implement fast adaptation", "add batch norm handling for meta-learning", "implement differentiable inner loop", or mentions MAML, FOMAML, Reptile, meta-gradients, episodic few-shot, inner-loop optimization, MAML++ enhancements, or Phase 7 meta-training in the cognitive pipeline.
Neuromodulation + Eligibility Traces (Three-Factor Learning)
by sovr610
This skill should be used when the user asks to "implement eligibility traces", "add three-factor learning", "implement neuromodulation", "add DA/ACh/NE/5-HT modulators", "implement STDP eligibility", "add synaptic plasticity", "implement online plasticity", "add reward-modulated learning", "implement trace dynamics", "add Dutch traces", "implement accumulating traces", "add replacing traces", "implement neuromodulatory gate", "add plasticity gain", "implement delayed reward association", "add eligibility decay", "implement spike-based traces", "add rate-based traces", "implement three-factor weight update", "add fast memory adapter", "implement bioplausible learning", or mentions eligibility traces, three-factor learning rules, neuromodulatory signals, STDP-based eligibility, reward-modulated plasticity, DA/ACh/NE/5-HT computation, or online synaptic updates in the cognitive pipeline.
Sleep Consolidation Cycle (Offline Replay + Synaptic Homeostasis + Systems Consolidation)
by sovr610
This skill should be used when the user asks to "add sleep cycle", "memory consolidation", "offline replay", "synaptic homeostasis", "sleep-wake cycle", "memory replay consolidation", "sharp wave ripple replay", "slow wave sleep computation", "dream-like generative replay", "synaptic downscaling", "interleaved replay", "systems consolidation", "hippocampal-cortical transfer", "offline training phase", "implement sleep consolidator", "add NREM replay phase", "add REM generative phase", "implement experience replay scheduler", "add weight downscaling", "implement complementary learning systems", "add fast-to-slow transfer", "implement two-phase training", "add sleep phase to training orchestrator", "implement priority replay sampling", "add compressed replay", "implement dream replay from world model", "add pseudo-rehearsal", "implement capacity restoration", "add synaptic renormalization", or mentions sleep consolidation, offline consolidation phase, NREM structured replay, REM creative recombination, sharp-wave ripple compressed replay, synaptic homeostasis hypothesis, Tononi-Cirelli downscaling, complementary learning systems, hippocampal-cortical transfer, fast-to-slow knowledge distillation, dream-like generative replay, pseudo-experience generation, two-phase wake-sleep training, replay priority sampling, capacity restoration, or sleep-wake integration with training orchestrator in the cognitive pipeline.
Hyperparameter Search
by sovr610
This skill should be used when the user asks to "tune hyperparameters", "find optimal learning rate", "run grid search", "run random search", "Bayesian optimization", "learning rate finder", "hyperparameter sweep", "tune model", "search space", "parameter optimization", "Optuna", "Ray Tune", "hyperband", "population based training", or needs guidance on hyperparameter search strategies, search space definition, trial management, or automated tuning for the brain_ai system.
Active Inference Agent (Generative Model + EFE + Empowerment)
by sovr610
This skill should be used when the user asks to "implement active inference", "add EFE computation", "implement expected free energy", "add empowerment estimation", "implement generative model", "add latent state encoder", "implement transition model", "add preference model", "implement planning rollouts", "add CEM planner", "implement amortized policy", "add pymdp backend", "implement offline RL", "add Minari integration", "implement pragmatic value", "add epistemic value", "implement instrumental value", "add action selection", "implement belief updating", "add world model training", "implement POMDP planning", "add rollout engine", "implement latent imagination", "add horizon normalization", "implement cross-entropy method planning", "add preference learning", "implement variational empowerment", or mentions active inference, expected free energy decomposition, POMDP planning, empowerment estimation, latent imagination, or decision-as-inference in the cognitive pipeline.
Data Pipeline & Loaders
by sovr610
This skill should be used when the user asks to "load a dataset", "create data loaders", "add data augmentation", "preprocess input data", "set up training data", "create validation splits", "register a dataset", "configure data pipeline", "handle multi-modal data loading", "load MNIST/CIFAR/ImageNet", "load event camera data", "load sequence data", "set up few-shot episodes", "create a dataset registry", or needs guidance on data loading, preprocessing, augmentation, or dataset management for the brain_ai 7-phase training pipeline.
DreamerV3-Style RSSM World Model
by sovr610
This skill should be used when the user asks to "implement DreamerV3 RSSM", "build a recurrent state space model", "create Block GRU sequence model", "implement unimix categorical", "add symlog twohot prediction heads", "implement KL balancing loss", "free nats clipping", "world model loss function", "imagination rollout for actor-critic", "straight-through categorical estimator", "implement prior and posterior networks", "DreamerV3 world model", "symlog transform", "twohot encoding 255 bins", "prevent codebook collapse", "DreamerV3 numerical stability", "scale-invariant reward prediction", "world model imagination", "RSSM prior posterior KL divergence", "Block GRU with RMSNorm", "categorical latent state 32x32", or needs guidance on implementing DreamerV3-style world models with the full set of numerical stability techniques (symlog, twohot, unimix, KL balancing).
Compiler & Kernel Fusion (torch.compile) Integration
by sovr610
This skill should be used when the user asks to "enable torch.compile", "add compilation to training", "kernel fusion", "TorchDynamo integration", "TorchInductor optimization", "reduce-overhead mode", "max-autotune mode", "fix graph breaks", "compile health check", "shape stabilization", "dynamic shapes for compile", "bucketing for torch.compile", "compile allowlist", "compile blocklist", "compile smoketest", "CUDA graphs for training", "maybe_compile wrapper", "debug recompiles", "TORCH_LOGS compile", "compile + DDP", "compile + FSDP", "torch.compiler.disable", or needs guidance on torch.compile integration, shape management, compilation debugging, or safe fallback patterns.
Evaluation & Benchmarks
by sovr610
This skill should be used when the user asks to "evaluate the model", "run benchmarks", "compute metrics", "measure accuracy", "test on MNIST", "compute F1 score", "generate confusion matrix", "evaluate few-shot", "measure anomaly detection", "run cognitive benchmarks", "compare model variants", "create evaluation report", "set up eval harness", or needs guidance on evaluation protocols, metrics computation, benchmark harnesses, or performance reporting for the brain_ai system.
V-JEPA 2 Self-Supervised Training
by sovr610
This skill should be used when the user asks to "train V-JEPA model", "implement JEPA pretext task", "set up EMA target encoder", "configure self-supervised training", "implement smooth L1 loss", "create training loop for V-JEPA", "optimizer configuration", "learning rate schedule", "warmup cosine decay", "EMA momentum schedule", "collapse prevention", "predictor architecture", "masked prediction loss", "DROID fine-tuning loop", "annealing phase", "cooldown training", or needs guidance on V-JEPA 2 self-supervised learning, training loops, predictor design, or optimization strategies.
V-JEPA 2 Data Pipeline
by sovr610
This skill should be used when the user asks to "load video dataset", "implement video transforms", "data augmentation for V-JEPA", "video decoding with decord", "clip sampling", "frame padding", "RandAugment for video", "motion shift augmentation", "random erasing", "video normalization", "YAML config parsing", "dataset registry", "distributed sampler", "weighted sampling", "multi-source dataset", "video DataLoader", "worker seeding", or needs guidance on video data loading, augmentation pipelines, configuration management, or dataset engineering for V-JEPA 2.
Model Export & Serving
by sovr610
This skill should be used when the user asks to "export the model", "convert to ONNX", "trace with TorchScript", "quantize the model", "deploy the model", "create API server", "serve predictions", "containerize the model", "optimize for inference", "export to INT8", "create FastAPI endpoint", "set up model serving", "prune the model", "distill knowledge", "create Docker image", or needs guidance on model export formats, quantization strategies, serving infrastructure, or deployment pipelines for the brain_ai system.
Paper-to-Spec Compiler
by sovr610
This skill should be used when the user asks to "extract a spec from a paper", "compile a paper into spec.yaml", "generate compliance tests from a paper", "create an executable spec", "parse arXiv paper into config", "detect paper drift", "diff code against paper", "generate spec from LaTeX", "validate a spec", "check if code matches the paper", or mentions converting academic ML/RL papers into machine-readable specifications. Treats papers as typed intermediate representations and emits spec.yaml, spec.md, compliance tests, and drift reports.
Spike Codec & Loss Pack
by sovr610
This skill should be used when the user asks to "encode spikes", "decode spike trains", "add spike encoding", "implement rate coding", "implement latency coding", "implement TTFS", "add population coding", "add delta modulation", "decode spike counts", "first-spike decoding", "population decoding", "add SNN loss", "implement ProbSpikes", "spike rate regularization", "ISI regularization", "temporal consistency loss", "membrane regularization", "SNN loss pack", "loss composition", "spike-count cross-entropy", "AMP hardening for SNN", "mixed precision SNN", "spike tensor convention", "axis convention batch-first", "SpikeBatch", "deterministic inference for encoders", "round-trip encoding test", or mentions spike I/O semantics, SNN training objectives, or encoding/decoding pipelines for spiking neural networks.
Compute/Throughput Baseline & Regression Gate
by sovr610
This skill should be used when the user asks to "add performance benchmarks", "create a regression gate", "measure training throughput", "compute MFU", "benchmark step time", "profile training loop", "set up CI perf gate", "compare against baseline", "collect environment info", "machine profile", "tokens per second measurement", "CUDA sync timing", "PyTorch profiler traces", "TensorBoard trace handler", "eval harness", "perplexity gate", "update performance baseline", "scaling efficiency test", or needs guidance on repeatable performance measurement, baseline storage, regression detection, profiling integration, or CI-gated throughput checks.
Gradient Checkpointing (Activation Recomputation)
by sovr610
This skill should be used when the user asks to "enable gradient checkpointing", "reduce training memory", "activation checkpointing", "torch.utils.checkpoint", "memory-compute tradeoff", "checkpoint sequential layers", "selective checkpointing", "recomputation strategy", "activation memory profiling", "per-layer memory budget", "checkpoint_sequential", "checkpoint_wrapper", "SAC selective activation checkpointing", "SNN timestep checkpoint", "FSDP activation checkpointing", "checkpoint per timestep", "memory-efficient training", "recompute activations in backward", or needs guidance on trading compute for memory during training, per-layer memory profiling, selective recomputation strategies, or integration with distributed training wrappers.
Visualization & Interpretability
by sovr610
This skill should be used when the user asks to "visualize activations", "plot spike rasters", "show attention heatmaps", "visualize workspace", "plot reasoning traces", "create t-SNE embeddings", "interpret model decisions", "visualize HTM patterns", "show neuromodulator levels", "plot training curves", "create activation maps", "visualize feature maps", "explain predictions", "debug model behavior visually", or needs guidance on visualization, interpretability, or explainability tooling for the brain_ai system.
Affective State Estimator (Valence-Arousal + Appraisal + Modulation)
by sovr610
This skill should be used when the user asks to "add emotion layer", "affective state estimation", "valence-arousal model", "emotional modulation of learning", "appraisal theory computation", "mood-congruent processing", "amygdala-inspired module", "somatic marker", "intrinsic motivation signal", "curiosity-driven exploration", "frustration detection", "reward prediction affect", "homeostatic regulation", "implement affective modulator", "add valence arousal space", "implement appraisal module", "add emotional bias to decisions", "implement somatic marker hypothesis", "add intrinsic reward signal", "implement dimensional emotion model", "add affect-modulated learning rate", "implement curiosity-anxiety tradeoff", "add frustration-driven exploration", or mentions affective state, valence-arousal, appraisal theory, emotional modulation, somatic markers, intrinsic motivation, mood-congruent processing, or homeostatic regulation in the cognitive pipeline.
Data Loader Throughput + Sequence Packing
by sovr610
This skill should be used when the user asks to "audit dataloader throughput", "measure data pipeline stalls", "implement sequence packing", "reduce padding waste in training", "add cu_seqlens packing for SFT", "boundary-aware packing FlashAttention", "DataCollatorWithFlattening", "configure DataLoader num_workers prefetch_factor", "tune DataLoader settings", "streaming dataset pipeline", "WebDataset tar shards", "memmap pretraining dataset", "HF datasets streaming IterableDataset", "deterministic sharding per rank", "DistributedSampler set_epoch", "padding ratio metrics", "effective tokens per second", "pretraining block builder", "data stall ratio measurement", "persistent_workers pin_memory tuning", "shard cache policy", "packed sequences with position_ids reset", "varlen_attn cu_seqlens", or needs guidance on turning the input pipeline into a first-class performance target with measurement, streaming I/O, packing, and deterministic distributed sharding.
Inference Optimization
by sovr610
This skill should be used when the user asks to "optimize inference", "speed up predictions", "add caching", "batch inference", "async inference", "reduce inference latency", "profile inference", "memory efficient inference", "inference pipeline", "warm up model", "KV cache", "inference throughput", "latency benchmark", "optimize forward pass", or needs guidance on inference performance optimization, caching strategies, batch processing, async execution, or latency profiling for the brain_ai system.
Dual-Process Reasoning (System 1/2 + Metacognitive Control)
by sovr610
This skill should be used when the user asks to "implement dual-process reasoning", "add System 1/2 router", "implement metacognitive control", "add confidence calibration", "implement temperature scaling", "add iterative refinement", "implement convergence detection", "add effort budget", "implement reasoning trace", "add novelty scoring", "implement routing policy", "add System 2 early stopping", "implement calibrated confidence", "add isotonic calibration", "implement GRU reasoning loop", "add selective S2 execution", "implement halt criteria", "add reasoning budget enforcement", "implement route score computation", or mentions dual-process reasoning, System 1/2 routing, metacognitive control, confidence calibration, convergence criteria, or reasoning traces in the cognitive pipeline.