diff --git a/#module.sigil b/#module.sigil
new file mode 100644
index 0000000..4d7aa7a
--- /dev/null
+++ b/#module.sigil
@@ -0,0 +1,42 @@
+component RiaToolkitOss {
+  goal {
+    RIA Toolkit OSS is the open-source RIA Toolkit: the foundational Python tools to develop, test, and deploy radio intelligence applications.
+    It owns the core radio data types, signal generation and processing, the radio dataset framework, a unified SDR hardware abstraction, annotation tooling, visualization, and the CLI, agent, and server surfaces that expose them.
+    It serves both standalone SDR/ML researchers and RIA Hub, whose controller embeds this library in-process.
+    It does not own RIA Hub's web platform, storage, or orchestration UI; it provides the radio primitives those build on.
+  }
+
+  interface {
+    A Python library (ria_toolkit_oss) importable by downstream code, including RIA Hub's controller which imports it in-process.
+    Library surfaces: data (radio-ML data types including Recording and Annotation, plus the radio dataset and dataset-builder framework); io (load/save recordings across SigMF, npy, wav, and Blue); signal (signal generation and processing); transforms (NumPy radio-data transforms used in ML); sdr (unified receive/transmit API across software-defined radios); annotations (annotation management and automatic signal-detection annotation); view and viz (plotting and report generation); utils (array-conversion and helpers).
+    orchestration: configure and run automated RF capture campaigns (CampaignConfig).
+    remote_control: remote SDR transmitter control over SSH and ZMQ.
+    server: the RT-OSS HTTP server (ria-server) that can run orchestration and inference as a standalone HTTP service.
+    CLI suite (console scripts): ria and ria-tools (main toolkit CLI), ria-server (RT-OSS HTTP server), ria-agent (edge node agent for campaigns/inference), ria-app (pull and run containerized RIA applications).
+    File formats: SigMF signal recordings and HDF5 radio datasets.
+  }
+}
+
+expand RiaToolkitOss {
+  logic {
+    The library provides the radio data types and signal tools; the CLI, agent, server, and app runners are entry points layered over that library.
+    It is consumed both as an imported Python library and through its console-script CLIs.
+    The server package (RT-OSS) is a standalone HTTP deployment of orchestration and inference; RIA Hub's controller embeds this library in-process rather than calling that server.
+  }
+
+  constraints {
+    Python 3.10+; license is AGPL-3.0-only.
+    Ships as a single distribution (ria-toolkit-oss) containing two packages from src: the ria_toolkit_oss library and the ria_toolkit_oss_cli CLI.
+    numpy is pinned to 1.26.4 for Radioconda compatibility; scipy is constrained below 1.16.
+    SDR hardware backends are intended as optional extras, but the individual optional-dependency entries are currently commented out; only an all-sdr list (pyrtlsdr, pyadi-iio for Pluto, pyhackrf, pyrf for ThinkRF; excluding USRP/UHD and BladeRF, which need system libraries) is present, and it is not wired as a valid PEP 621 extra.
+    Built with poetry-core.
+    The view and viz packages overlap: both provide recording-plot primitives (spectrogram, iq_time_series, frequency_spectrum, constellation); viz is the superset RIA Hub imports, view is the CLI-facing set. This duplication is unresolved.
+  }
+
+  cases {
+    A researcher installs the package and uses the ria CLI to generate or inspect a radio dataset.
+    RIA Hub's controller imports ria_toolkit_oss in-process to load recordings, render plots, and drive radios.
+    A ria-agent node runs on an edge device to execute capture or inference campaigns.
+    ria-server exposes toolkit operations to RIA Hub over HTTP.
+  }
+}
diff --git a/.sigil/config.json b/.sigil/config.json
new file mode 100644
index 0000000..8348c1a
--- /dev/null
+++ b/.sigil/config.json
@@ -0,0 +1,20 @@
+{
+  "sigilVersion": "0.1.0",
+  "workspace": {
+    "name": "ria-toolkit-oss",
+    "members": []
+  },
+  "files": {
+    "include": [
+      "**/*.sigil"
+    ],
+    "exclude": [
+      ".git/**",
+      ".deno/**",
+      "node_modules/**",
+      "build/**",
+      "coverage/**"
+    ]
+  },
+  "tools": {}
+}
diff --git a/src/ria_toolkit_oss/agent/agent.sigil b/src/ria_toolkit_oss/agent/agent.sigil
new file mode 100644
index 0000000..2a3ab28
--- /dev/null
+++ b/src/ria_toolkit_oss/agent/agent.sigil
@@ -0,0 +1,25 @@
+component RiaAgent {
+  goal {
+    Run a toolkit agent on an edge device that connects outbound to RIA Hub to capture, stream, or run inference on live radio.
+    Own the agent CLI, its two execution modes, and its persisted registration.
+  }
+
+  interface {
+    CLI ria-agent with subcommands: register (--hub, --api-key, --name, and TX interlock flags), stream, run (legacy), and detect (list SDR drivers).
+    register POSTs to {hub}/screens/agents/register with the X-API-Key registration key and saves agent_id and token to ~/.ria/agent.json (mode 0600).
+    Two execution modes: a WebSocket streamer that opens an SDR and streams raw IQ to the hub (the hub does inference), and a legacy long-poll NodeAgent that runs ONNX inference locally on the device.
+  }
+}
+
+expand RiaAgent {
+  logic {
+    The streamer connects to wss://{hub}/screens/agent/ws with a bearer token, heartbeats, auto-reconnects, and runs concurrent RX (ship interleaved float32 IQ) and TX (play inbound IQ) sessions, sharing one driver for full-duplex when RX and TX name the same device.
+    The legacy NodeAgent connects outbound only: it registers at /composer/nodes/register, heartbeats, long-polls /commands, and dispatches run_campaign, load_model, start/stop/configure_inference, and TX commands, reporting to /events and /campaign-status.
+  }
+
+  constraints {
+    TX is opt-in (--allow-tx) and bounded by tx_max_gain_db, tx_max_duration_s, and allowed frequency ranges, with a TX watchdog; a TX-start request exceeding the interlocks is rejected (per-frame handling rejects malformed frames and enforces max duration).
+    Registration keys are personal (ria_reg_ prefix); the legacy shared key is deprecated. Failures are reason-coded (invalid_key, expired, revoked, already_consumed, rate-limited).
+    Config is AgentConfig at ~/.ria/agent.json (override RIA_AGENT_CONFIG).
+  }
+}
diff --git a/src/ria_toolkit_oss/annotations/annotations.sigil b/src/ria_toolkit_oss/annotations/annotations.sigil
new file mode 100644
index 0000000..944aa3a
--- /dev/null
+++ b/src/ria_toolkit_oss/annotations/annotations.sigil
@@ -0,0 +1,27 @@
+component Annotations {
+  goal {
+    Create, manage, and automatically detect annotations on a recording.
+    Own the detection algorithms, annotation transforms, signal isolation, and slice qualification.
+  }
+
+  interface {
+    Automatic detection, each returning a new Recording with detected annotations appended: detect_signals_energy (energy/envelope), annotate_with_cusum (change-point segmentation), threshold_qualifier (hysteresis burst detection), annotate_with_obw (occupied-bandwidth), and find_spectral_components with split_recording_annotations (parallel signal separation).
+    Bandwidth helpers: calculate_occupied_bandwidth, calculate_nominal_bandwidth, calculate_full_detected_bandwidth.
+    Annotation transforms: remove_contained_boxes and is_annotation_contained (merge_annotations is declared but not implemented).
+    isolate_signal (slice, normalize, shift to baseband, low-pass to the annotation bandwidth) and qualify_slice_from_annotations (fixed-length slices overlapping annotations).
+    This layer is a standalone library surface; the RIA Hub controller does not import it directly.
+  }
+}
+
+expand Annotations {
+  logic {
+    Energy detection builds a moving-average envelope, estimates a segment-median noise floor, thresholds with hysteresis, merges nearby boundaries, and assigns frequency bounds by the chosen method (nominal, occupied, or full-detected bandwidth).
+    Detectors write a JSON comment recording the detection type, generator, and parameters, plus a detail dict.
+  }
+
+  constraints {
+    Detectors require a sample_rate in the recording metadata and operate on channel 0 (only threshold_qualifier accepts a channel override).
+    Annotation frequencies are stored absolute (center frequency plus the relative offset); the parallel separator works in baseband and converts.
+    A detector warns and returns the recording unchanged when the dynamic range is too low to threshold.
+  }
+}
diff --git a/src/ria_toolkit_oss/app/app.sigil b/src/ria_toolkit_oss/app/app.sigil
new file mode 100644
index 0000000..d1917e3
--- /dev/null
+++ b/src/ria_toolkit_oss/app/app.sigil
@@ -0,0 +1,24 @@
+component AppRunner {
+  goal {
+    Pull and run containerized RIA applications, auto-configuring hardware access from image labels.
+    Own the ria-app CLI over a container engine.
+  }
+
+  interface {
+    CLI ria-app with subcommands: pull, run, list, stop, logs, and configure; a global --sudo flag.
+    run auto-pulls the image if absent and derives container flags from the image's ria.* labels.
+  }
+}
+
+expand AppRunner {
+  logic {
+    The engine is the first of docker or podman on PATH, optionally sudo-prefixed.
+    Image references resolve against a configured registry and namespace; fully-qualified references pass through.
+    Hardware flags are derived from ria.profile and ria.hardware labels: GPU access when an NVIDIA profile and runtime are present, USB device passthrough for USB SDRs, and host networking for USRP, ThinkRF, and Pluto; flags can be overridden or dry-run.
+  }
+
+  constraints {
+    There is no Python SDK; it is a subprocess wrapper and exits with an error if no container engine is found.
+    Config is AppConfig (registry, namespace, sudo) at ~/.ria/toolkit.json with environment-variable fallbacks.
+  }
+}
diff --git a/src/ria_toolkit_oss/data/data.sigil b/src/ria_toolkit_oss/data/data.sigil
new file mode 100644
index 0000000..7018659
--- /dev/null
+++ b/src/ria_toolkit_oss/data/data.sigil
@@ -0,0 +1,34 @@
+component RadioData {
+  goal {
+    Provide the core radio-ML data types: a Recording (an immutable IQ tape with metadata and annotations) and the radio dataset and dataset-builder framework.
+    Own the in-memory contract the rest of the toolkit and RIA Hub build on for signals and datasets.
+  }
+
+  interface {
+    Recording(data, metadata=None, dtype=None, timestamp=None, annotations=None) - complex IQ data shaped channels-by-samples (1-D is reshaped to one channel); metadata must be a JSON-serializable dict; rec_id and timestamp are auto-populated.
+    Recording properties: data, metadata, annotations, shape, n_chan, rec_id, dtype, timestamp, sample_rate (settable).
+    Recording methods: astype, add_to_metadata/update_metadata/remove_from_metadata (rec_id and timestamp protected), trim, normalize, view/simple_view, and to_sigmf/to_npy/to_wav/to_blue (delegating to io).
+    Annotation(sample_start, sample_count, freq_lower_edge=None, freq_upper_edge=None, label="", comment="", detail=None) with is_valid, overlap, area, and to_sigmf_format.
+    Dataset framework (data.datasets): abstract RadioDataset (HDF5-backed and iterable; item assignment raises, though it exposes explicit inplace mutators) with IQDataset and SpectDataset subclasses, DatasetBuilder, and leakage-safe split/random_split.
+  }
+}
+
+expand RadioData {
+  state {
+    A Recording's sample data is immutable: item assignment raises, and property accessors return copies or read-only views; metadata is mutable only through the metadata methods.
+    Metadata keys must be lowercase letters and underscores only (no digits); rec_id and timestamp are protected and cannot be modified or removed.
+  }
+
+  logic {
+    A Recording requires complex data of one or two dimensions (channels by samples); rec_id is the SHA-256 of the data bytes plus timestamp.
+    trim slices samples and re-aligns annotations; normalize scales peak magnitude to 1 and rejects all-zero data.
+    RadioDataset reads data and metadata lazily from an HDF5 source; non-inplace mutations write numbered sibling files; concrete backends live outside the OSS package.
+    split and random_split partition by rec_id so no recording leaks across subsets; lengths may be counts or fractions.
+  }
+
+  constraints {
+    Recording data must be complex or construction raises; metadata must be a JSON-serializable dict.
+    RadioDataset sources must be HDF5 (validated at construction); the data and metadata datasets are required for use but not checked at init; the class is abstract and instantiated only via a concrete backend.
+    The datatypes package is stale (only bytecode remains); the live location is data.
+  }
+}
diff --git a/src/ria_toolkit_oss/io/io.sigil b/src/ria_toolkit_oss/io/io.sigil
new file mode 100644
index 0000000..c29cdb0
--- /dev/null
+++ b/src/ria_toolkit_oss/io/io.sigil
@@ -0,0 +1,28 @@
+component RecordingIO {
+  goal {
+    Load and save Recording objects to and from file across the RF formats the toolkit supports.
+    Own format detection and the per-format readers and writers.
+  }
+
+  interface {
+    load_recording(file) - dispatch on file extension to the matching reader (sigmf, npy, wav, blue); an unknown extension raises.
+    Per-format readers and writers: to_sigmf/from_sigmf, to_npy/from_npy (plus from_npy_legacy), to_wav/from_wav, to_blue/from_blue.
+    Most writers return the output path (to_sigmf returns None) and honor an overwrite flag; auto-generated filenames land under recordings/.
+  }
+}
+
+expand RecordingIO {
+  logic {
+    The npy v2 format stores a version marker then data, JSON metadata, and JSON annotations, avoiding pickle; legacy npy paths fall back to pickle with a warning and normalize namespaced metadata keys.
+    SigMF writes paired .sigmf-data and .sigmf-meta, mapping RIA metadata to core keys and namespacing the rest under ria:.
+    WAV stores I and Q as stereo channels with metadata in an info chunk; Blue uses the MIDAS 512-byte header format.
+  }
+
+  constraints {
+    Format detection is by file extension only.
+    The SigMF, WAV, and Blue writers reject multichannel input and the readers produce single-channel output; only npy round-trips multichannel data.
+    Integer-encoded outputs (16-bit WAV, integer Blue) require samples normalized to the range -1 to 1.
+    io.common (exists, validate, move, copy) is declared but unimplemented.
+    save_recording is exported in the package __all__ but not defined (latent bug).
+  }
+}
diff --git a/src/ria_toolkit_oss/orchestration/orchestration.sigil b/src/ria_toolkit_oss/orchestration/orchestration.sigil
new file mode 100644
index 0000000..d9491a8
--- /dev/null
+++ b/src/ria_toolkit_oss/orchestration/orchestration.sigil
@@ -0,0 +1,26 @@
+component Orchestration {
+  goal {
+    Define and execute automated RF capture campaigns: sweep transmitters and channels, record, QA, and label the results.
+    Own the campaign configuration schema, the executor, and the QA checks.
+  }
+
+  interface {
+    CampaignConfig(name, recorder, transmitters, qa, output, mode, loops) with loaders from_dict, from_yaml, and from_device_profile; it requires a recorder and at least one transmitter.
+    Supporting dataclasses: RecorderConfig, TransmitterConfig (type and control_method external_script/sdr/sdr_remote/sdr_agent), CaptureStep, QAConfig, and OutputConfig.
+    CampaignExecutor(config, progress_cb=None).run() returns a CampaignResult of StepResults; check_recording(recording, config) returns a QAResult; label_recording writes ria: SigMF metadata.
+  }
+}
+
+expand Orchestration {
+  logic {
+    The executor opens the SDR and remote TX controllers once, then for each loop, transmitter, and step it starts the transmitter, records IQ, stops the transmitter, then labels, QA-checks, and saves it as SigMF.
+    QA estimates SNR from the PSD and checks duration and SNR; in flag-for-review mode a recording always passes but may be flagged.
+    Cancellation is cooperative at step boundaries via the progress callback.
+  }
+
+  constraints {
+    Frequencies, gains, durations, and bandwidths are parsed from human units (for example 2.45GHz, auto, 1.5m).
+    Device names are aliased onto SDR drivers, including mock and sim for hardware-free runs.
+    The controller's Conductor and the RT-OSS server are parallel HTTP wrappers that both reuse this package's CampaignConfig and CampaignExecutor in-process.
+  }
+}
diff --git a/src/ria_toolkit_oss/remote_control/remote-control.sigil b/src/ria_toolkit_oss/remote_control/remote-control.sigil
new file mode 100644
index 0000000..5e93fbb
--- /dev/null
+++ b/src/ria_toolkit_oss/remote_control/remote-control.sigil
@@ -0,0 +1,25 @@
+component RemoteControl {
+  goal {
+    Drive a software-defined radio transmitter on a remote machine over SSH and ZMQ.
+    Own the client controller and the remote server that bridge transmit commands to a remote SDR.
+  }
+
+  interface {
+    RemoteTransmitterController(host, ssh_user, ssh_key_path, zmq_port=5556): set_radio, init_tx, transmit_async, wait_transmit, stop, close.
+    RemoteTransmitter (the server on the transmit machine): set_radio, init_tx, transmit, stop, and a run_function JSON-RPC dispatcher; launched as a module over a ZMQ REP socket.
+  }
+}
+
+expand RemoteControl {
+  logic {
+    The controller SSHes into the transmit host (paramiko, key-file auth), launches the remote_transmitter server there, waits for it to bind, then drives it with JSON-RPC over a ZMQ REQ/REP socket.
+    Supported remote devices are the transmit-capable drivers: Pluto, USRP, HackRF, and BladeRF.
+    set_radio must precede init_tx and transmit (both guard on a radio being set); the init_tx-before-transmit order is a usage convention, not an enforced guard; transmit runs in a daemon thread joined by wait_transmit.
+  }
+
+  constraints {
+    SSH host-key policy is AutoAddPolicy, which accepts unknown hosts (security note).
+    RIA Hub does not use this subpackage; remote and agent transmit is handled through the agent path instead.
+    Known gap: transmit calls an SDR method tx_cw that no driver defines, so remote CW transmit currently no-ops for real drivers (the AttributeError is swallowed).
+  }
+}
diff --git a/src/ria_toolkit_oss/sdr/sdr.sigil b/src/ria_toolkit_oss/sdr/sdr.sigil
new file mode 100644
index 0000000..7ca0893
--- /dev/null
+++ b/src/ria_toolkit_oss/sdr/sdr.sigil
@@ -0,0 +1,29 @@
+component SdrInterface {
+  goal {
+    Provide a unified receive and transmit API across a variety of software-defined radios.
+    Own the SDR base contract, the per-device drivers, and device discovery and selection.
+  }
+
+  interface {
+    get_sdr_device(device_type, ident=None, tx=False) returns an SDR; mock and sim return a MockSDR, real devices delegate to the CLI package's factory.
+    detect_available() returns the drivers whose optional dependency imports cleanly; it does not probe hardware.
+    SDR base class: record(num_samples, rx_time) and rx(num_samples) to receive, tx_recording(recording) to transmit, stream_to_zmq and pickle_buffer_to_zmq to stream, getters and setters for sample rate, center frequency, and gain, and capability probes for bias-tee and dynamic-update support.
+    Drivers: Pluto, USRP, HackRF, and Blade (receive and transmit); RTLSDR and ThinkRF (receive only); MockSDR (AWGN generator, no hardware, receive and transmit). Real drivers are constructed with a device identifier; MockSDR takes a buffer size and seed.
+    Errors: SDRError with SDRParameterError and SdrDisconnectedError (SDROverflowError is defined but not in the public __all__).
+  }
+}
+
+expand SdrInterface {
+  logic {
+    init_rx or init_tx must be called before record or tx_recording (rx auto-initializes on first use); they push sample rate, center frequency, and gain to the hardware via per-driver setters.
+    Receive accumulates validated buffers into a Recording; transmit loops or truncates the source IQ to the requested length.
+    A driver is available only if its optional hardware dependency imports; USB or network drops are translated to SdrDisconnectedError.
+  }
+
+  constraints {
+    gain_mode absolute applies the requested gain (clamped or snapped to the device's valid range); relative requires a negative (attenuation) value; ThinkRF uses named gain profiles instead.
+    RTLSDR and ThinkRF are receive-only; their transmit methods raise or are stubs.
+    Identifiers are device-specific free-form values (Pluto URI or IP, USRP serial or args, HackRF/BladeRF serial).
+    RIA Hub consumes this layer through the SDR base contract and get_sdr_device but currently wires only Pluto and USRP live; the controller imports driver modules directly rather than only via the factory.
+  }
+}
diff --git a/src/ria_toolkit_oss/server/server.sigil b/src/ria_toolkit_oss/server/server.sigil
new file mode 100644
index 0000000..04a1f89
--- /dev/null
+++ b/src/ria_toolkit_oss/server/server.sigil
@@ -0,0 +1,28 @@
+component RtOssServer {
+  goal {
+    Expose the toolkit's campaign orchestration and live RF inference over HTTP as a standalone service (RT-OSS).
+    Own the FastAPI app, its auth, and the deploy and inference endpoints.
+  }
+
+  interface {
+    create_app(api_key="") returns a FastAPI app mounting /conductor and /inference (both API-key gated) plus an unauthenticated GET /health; an empty api_key disables auth (dev only).
+    serve() launches uvicorn from RT_OSS_API_KEY, RT_OSS_HOST (default 0.0.0.0), and RT_OSS_PORT (default 8080), and is the ria-server entry point; the ria serve CLI command is separate and calls create_app() directly.
+    Conductor endpoints: POST /conductor/deploy ({config}) returns {campaign_id} and runs the campaign in a background thread; GET /conductor/status/{campaign_id}; POST /conductor/cancel/{campaign_id}.
+    Inference endpoints: POST /inference/load (ONNX model plus label_map), POST /inference/start, POST /inference/stop, POST /inference/configure, GET /inference/status.
+  }
+}
+
+expand RtOssServer {
+  logic {
+    A deploy parses the body via CampaignConfig.from_dict and rejects any transmitter carrying a script field (external scripts blocked).
+    The inference loop captures 4096 IQ samples, estimates SNR, preprocesses to the model input shape, softmaxes, and maps the index to a label; idle-class labels are reported as idle.
+    State is in-memory only (a campaign dict plus a single thread-locked inference state); nothing is persisted.
+  }
+
+  constraints {
+    Auth is X-API-Key compared in constant time; a missing configured key allows all requests (dev only).
+    Server dependencies (fastapi, uvicorn, onnxruntime) are an optional extra; a missing onnxruntime or SDR yields 503.
+    RIA Hub's controller does not call this server; it embeds the toolkit library in-process. This server is the alternative standalone deployment.
+    model_path is path-traversal validated and must resolve to a .onnx file.
+  }
+}
diff --git a/src/ria_toolkit_oss/signal/signal.sigil b/src/ria_toolkit_oss/signal/signal.sigil
new file mode 100644
index 0000000..7bf160e
--- /dev/null
+++ b/src/ria_toolkit_oss/signal/signal.sigil
@@ -0,0 +1,26 @@
+component SignalProcessing {
+  goal {
+    Provide the toolkit's signal generation and processing suite: functional waveform generators and a modular block-based modulation-chain framework.
+    Own the DSP building blocks that produce Recording objects for datasets and transmission.
+  }
+
+  interface {
+    Functional generators returning a Recording: sine, square, sawtooth, noise, chirp, lfm_chirp_complex, complex_sine.
+    Block framework: Block, SourceBlock, ProcessBlock, and RecordableBlock forming a DataType-typed chain, with sources (BinarySource, AWGNSource, SineSource, RecordingSource, ...), Mapper (PSK/QAM/PAM constellations; standalone APSK and cross-QAM mapper classes exist but are not reachable through Mapper) and SymbolDemapper, symbol modulators (GMSK, OOK, OQPSK), continuous-phase modulators (FSK, CPFSK), pulse-shaping filters (RaisedCosine, RootRaisedCosine, Gaussian, Sinc, Rect), Upsampling, channels (AWGNChannel, FlatRayleigh), and basic ops (Add, MultiplyConstant, PhaseShift, FrequencyShift, frequency up/down conversion).
+    High-level generators: PSKGenerator, QAMGenerator, and PAMGenerator, each with record(batch_size, num_bits).
+    This layer is a standalone library surface; the RIA Hub controller does not import it directly.
+  }
+}
+
+expand SignalProcessing {
+  logic {
+    A SignalGenerator validates that each block's output type matches the next block's input type; sample production is driven by the blocks themselves (Block and RecordableBlock via get_samples, and the PSK/QAM/PAM generators by chaining block __call__).
+    Blocks auto-serialize their JSON-safe attributes into the produced Recording's metadata.
+  }
+
+  constraints {
+    Supported modulations reachable via Mapper and the generators: PSK, QAM, PAM (constellation); GMSK, OOK, OQPSK (symbol); FSK and CPFSK (continuous). QAM and PAM require an even bits-per-symbol (QAM additionally more than 2). APSK and cross-QAM mapper classes exist but are unwired.
+    The basic generators (sine, square, sawtooth, noise, complex_sine) raise on a sample rate below 1 (chirp instead requires at least two samples); FSK and CPFSK require frequency spacing times symbol duration of at least 0.5.
+    Data flows as typed stages (bits, symbols, upsampled symbols, baseband, passband) via the DataType enum.
+  }
+}
diff --git a/src/ria_toolkit_oss/transforms/transforms.sigil b/src/ria_toolkit_oss/transforms/transforms.sigil
new file mode 100644
index 0000000..3714156
--- /dev/null
+++ b/src/ria_toolkit_oss/transforms/transforms.sigil
@@ -0,0 +1,21 @@
+component Transforms {
+  goal {
+    Provide NumPy radio-data transforms used by the ML backends: data augmentations and channel/hardware impairment models over IQ.
+    Own the functions that manipulate a signal while preserving its container type.
+  }
+
+  interface {
+    Augmentations (iq_augmentations): generate_awgn, time_reversal, spectral_inversion, channel_swap, amplitude_reversal, drop_samples, quantize_tape, quantize_parts, magnitude_rescale, cut_out, patch_shuffle.
+    Impairments (iq_impairments): add_awgn_to_signal, time_shift, frequency_shift, phase_shift, iq_imbalance, resample.
+    Every function accepts an array or a Recording and returns the same type, preserving metadata for a Recording.
+    This layer is a standalone library surface; the RIA Hub controller does not import it directly.
+  }
+}
+
+expand Transforms {
+  constraints {
+    Input must be a 2-D complex channels-by-samples array or a ValueError is raised; most transforms implement only the single-channel path and raise NotImplementedError for more channels.
+    frequency_shift is relative to the sample rate and must be within -0.5 to 0.5; phase_shift must be within -pi to pi.
+    AWGN is generated to match a target SNR in dB.
+  }
+}
diff --git a/src/ria_toolkit_oss_cli/cli.sigil b/src/ria_toolkit_oss_cli/cli.sigil
new file mode 100644
index 0000000..e7de867
--- /dev/null
+++ b/src/ria_toolkit_oss_cli/cli.sigil
@@ -0,0 +1,11 @@
+component ToolkitCli {
+  goal {
+    Provide the ria and ria-tools command-line interface over the toolkit's library features.
+    Own the Click command group and its subcommands.
+  }
+
+  interface {
+    A Click group that auto-registers the commands defined in ria_toolkit_oss_cli.ria_toolkit_oss.commands and warns when git-lfs is missing.
+    Commands map to library features: capture and transmit (sdr), campaign (orchestration), generate and synth (signal), transform/combine/split/convert (transforms, data, io), annotate (annotations), view (view and viz), discover (sdr), serve (the RT-OSS server), and upload/setup_repo/init (RIA Hub repo integration).
+  }
+}