Included modules provide native types and explicit execution contracts. Data structure references are listed separately from module APIs.
These modules currently ship with Aner. Independent packages and distribution of libraries from other developers are part of the roadmap.
Core collection references
These data structures are included in the language and require no import. Each reference documents its construction, supported operations, constraints, and time and space complexity.
Typed arrays require an explicit Int64, Float64, Bool, or String element type and need no import. Allocation, capacity, sorting, searching, and numerical contracts are documented separately for each type.
NDArray reference
Contiguous matrices and ranks 1 through 8 use explicit element types and shapes, without an import. The reference specifies reshape ownership, broadcasting, permutation, matrix multiplication, and time and space complexity.
Compose a Dense topology, define an explicit training function, or control parameter leaves and keyed updates in a class. CPU and CUDA support Float64/Float32; Metal/MPS supports Float32. Requested and effective execution remain inspectable.
Dense layers with explicit weights, Xavier or He initialization
Xavier MLP exploration and explicit Dense initialization, forward objectives, gradients and parameter updates
Reverse mode automatic differentiation
Full batch and lazy mini batch SGD/Adam with explicit loss, validation and fixed epochs
Measured progress, fixed epoch ETA, and optimizer checkpoint continuation
Float64 preservation, compatibility adaptation, and explicit Float32 performance policy
Explicit CPU, Metal/MPS and CUDA placement; CNN, RNN and general distributed Tensor execution remain planned
V5 epoch boundary continuation retains optimizer state, exact schedules and source membership history
Supported Dense replica training through aner.parallel with native CPU, Metal/MPS and CUDA forward and backward computation
Declare CSV schemas, preserve identifiers and missing values, and inspect a lazy scan before execution. Managed buffer budgets make collection an explicit decision.
Typed CSV schemas and exact missing tokens
Projection and bounded filter operations
Batch summaries and explicit collection
Validation of the full schema before filtering
Joins, Parquet, SQL, and streaming model training remain planned
Load the offline Iris/Wine teaching catalogue or verified local MNIST files. Immutable source identity, protected roles, stratified/group aware partitions, folds and lazy batches retain the training protocol.
Offline Iris and Wine
CC BY 4.0 attribution and exact source variants
Source row IDs and immutable row views
Seeded train/test partitions
Standardization fitted on training data
Verified compact MNIST IDX loading without network access
Training/validation/test partitions, grouped folds and immutable source row audits
Lazy deterministic mini batches and training only streamed standardization
Classical and neural workflows meet in typed fitted models. Start with KNN, K means, and concise neural classifier pipelines; inspect exact limits in the reference.
Fitted KNN classifiers
Seeded K means clustering
Neural classifier pipelines with scaling fitted only on training data
Explicit prediction and evaluation
SVM, decision trees, and general model interfaces remain planned
Immutable mini batch training contracts with validation, final or best epoch selection, and historical test exclusion
Construct and inspect rank two Float64 and Float32 Tensors shared by datasets and numerical modules, with explicit CPU or available accelerator placement.
Neutral tensor construction
Rows, columns, and value inspection
CPU and native CUDA Float64/Float32 storage; native Metal/MPS Float32 storage in compatible builds
Explicit shape and value validation
Additional Tensor ranks, mixed precision and general distributed Tensor execution remain planned; core NDArray separately supports typed ranks 1 through 8
Record selected training steps on the actual tensor device and explore topology, loss, and available gradient summaries in standalone browser, JSON, and terminal reports.
Actual training captures
Offline interactive HTML reports
JSON and terminal summaries
Bounded raw values included only when requested
Completed cell inspection also uses development mode; streaming and general plotting remain planned
Capture bounded observations of variables, objects, collections, and neural training in development. Production uses the same computation with automatic recording disabled.
Named, read only snapshots with dev.capture
Stable object references and collection mutation observations
Offline inspection reports and separate fit loss curves
VS Code Inspector and optional Jupyter completed cell views
Pause, step, and streaming observations remain planned
Run native scientific tasks or supported Dense neural replicas with explicitly selected resources. Retain sample identity, protected Dataset partitions and one canonical optimizer update per fixed global batch.
Bounded reusable C++20 CPU task executor
Sample addressed deterministic Monte Carlo with exact integer sufficient statistics
Explicit local or configured process and SSH execution contexts
CPU allocation inventory, terminal progress and machine readable result accounting
Experimental Float64 analytic regression CLI with valid count gradient reduction and one canonical full batch SGD update
Native CPU, CUDA and Metal/MPS Dense forward and backward replicas with Float64 sample weighted host reduction and a canonical CPU optimizer
Strict common compute precision, bounded checked process or SSH exchange, and actual replica timing and memory observations
Direct GPU collectives, model sharding and durable coordinator recovery remain planned
Connect to a local or cloud language model through native provider adapters. Begin with a prompt, then make message order, output budget, sampling and endpoint control explicit without changing the typed response contract.
Native Ollama chat, OpenAI Responses and vLLM compatible chat adapters
Simple prompt generation and explicit ordered typed messages
Opaque credential references resolved at request execution
Verified HTTPS, loopback local connections and bounded synchronous transport
Typed text, model, finish reason and reported token usage
No automatic retries or silent provider changes; streaming, embeddings and tool execution remain planned
Keep key values outside source and notebooks. Native opaque references resolve through an operating system store, an injected process environment, or an explicitly selected managed file fallback.
Hidden interactive credential input through aner secrets set
Native macOS Keychain and Windows Credential Manager storage adapters
Explicit environment references without implicit project file loading
Managed file fallback with strict owner permissions on POSIX
Credential status and deletion without exporting values
No silent plaintext fallback; Linux system vault integration remains planned
Expose fitted neural inference or a configured language model through native fixed routes. Keep model, checkpoint, preprocessing, effective device, credentials and transport under explicit application control.
Immutable APIApp builders with fixed classifier and language model routes
Strict finite numerical input rows and checked prediction/probability results
OpenAPI 3.1 descriptions of request and response contracts
Native TLS, explicit permitted hostnames and opaque inbound bearer references
Loopback defaults, bounded bodies, rows, request waiting and process rate admission
One serialized inference with retained fitted preprocessing and model placement
Service to service integration; general callback routes, streaming and browser CORS remain outside the current contract
