Integration & Project Management¶

Block 6 ties all Studio components together: project save/load for persistent workspaces, and a full pipeline that chains network design through compilation to FPGA synthesis in a single action.

Project Save/Load¶

Save and restore complete Studio state — equations, parameters, network graph, synthesis target, training config — as JSON files on the server.

Save¶

Click Save in the Projects panel (left sidebar), enter a name. The full state is serialised into the server-side Studio project workspace. The API response is path-free and returns studio.project-save.v1 metadata with SHA-256 digests for the project state and full saved project payload. The Projects panel displays the returned evidence classification, project name, state digest, project digest, and schema version after a successful save.

Load¶

Click a project name to restore all state: equations, model selection, parameters, network graph populations and projections, synthesis target, and the Training Monitor configuration. Positive timebase settings and finite numeric parameters are restored from saved values, including zero current, while malformed optional fields fall back to Studio defaults before the next simulation refresh.

Storage Format¶

JSON

{
  "name": "my_network",
  "saved_at": 1711504200.0,
  "version": "0.3.0",
  "state": {
    "sourceMode": "model",
    "selectedModelName": "LIFNeuron",
    "graphPopulations": [...],
    "graphProjections": [...],
    "synthTarget": "ice40",
    "trainingConfig": {...}
  }
}

Full Pipeline¶

The pipeline builds hardware for the network drawn on the Canvas, or says exactly why it cannot:

Text Only

Network Graph → Validate → Simulate → Lower → Co-simulate → Synthesise

Steps¶

Validate — check the graph (populations exist, projections point to valid nodes, neuron count within the 2000 limit).
Simulate — run the network in the Studio runtime.
Lower — translate the network for the hardware network compiler with each catalogue model's own step. Supported: SCLapicqueLIFNeuron (profile sc_lif, the exact step v ← v·e^(−dt/τ) + (v_rest + R·I)(1 − e^(−dt/τ)), firing at v ≥ v_threshold) and PerfectIntegratorNeuron (r = 1/c_m, with its profile's threshold comparison). Projections keep the connectivity the runtime realises and their delay in steps; the hardware's registered spikes are the runtime's one-step latency. A constant drive enters through one input lane per neuron. Refused before any hardware is generated, every reason at once: another model, a Poisson drive, a membrane that does not start at rest, a value outside the fixed-point range or one that quantises to zero, a delay beyond 1024 steps, and two populations of one model with different parameters. Values that round are listed in the step's notes.
Co-simulate — compile the RTL and run it in Icarus Verilog beside a C model built from the compiler's bit-true neuron kernels and the compiled interconnect, and beside the Studio's own run, for up to 2000 steps. The RTL must reproduce its model on every step, or synthesis does not run. Where the fixed-point hardware and the Studio's double-precision run differ, the first differing step is reported.
Synthesise — run Yosys on the top module and the neurons it instantiates for the selected FPGA target (ice40, ECP5, Gowin, Xilinx).

The result's trace binds the run: the lowering's input_sha256 (the graph's digest, the fixed-point format and the lowering schema), the RTL, the bit-true model and the synthesised source. Two different networks never share them. The format is Q8.8 unless the request asks for Q16.16; exact values such as 0.375 reproduce the Studio run in either, while a leaky neuron's decay rounds visibly in Q8.8.

Co-simulation needs iverilog, vvp and a C compiler on the server; without them the pipeline stops at the co-simulation step and says which tool is missing.

Using the Pipeline¶

Design your network in the Network canvas view
Select FPGA target in the Synthesis Dashboard (or use default ice40)
Click Pipeline → ICE40 (or whichever target) on the Canvas toolbar
Read the result under the canvas: the step it ended at, every refusal reason, and what the co-simulation established

Hub Federation Manifest¶

The optional Hub-facing federation surface lives under sc_neurocore.federation and is separate from this local Studio web app. It declares the schema-A capability manifest, the eight advertised verbs, and the evidence-bundle contracts consumed by SCPN Studio platform federation.

Federation requires Python 3.11+ and Platform >=0.11,<0.12. Its manifest declares contract era v2; consumers must support that era. Install the optional SDK before generating or checking the committed manifest:

Bash

pip install "sc-neurocore[federation]"
python tools/emit_studio_manifest.py --check

The generated artifact is docs/_generated/studio_manifest.json; it stamps the source package version and hashes the declared verbs plus evidence schemas, not git state. API details and envelope examples are in Studio Federation API.

API Endpoints¶

Method	Endpoint	Description
POST	`/api/project/save`	Save project state
GET	`/api/project/list`	List all saved projects
GET	`/api/project/load/{name}`	Load a saved project
DELETE	`/api/project/{name}`	Delete a saved project
POST	`/api/pipeline/run`	Lower, co-simulate and synthesise the network, or refuse it

/api/pipeline/run executes through the Studio local worker manager. It keeps the synchronous pipeline response used by the Network Canvas and also records a studio-pipeline Admin queue job with the path-free result artifact pipeline/result.json. The Network Canvas displays the same action-evidence contract beside the terminal pipeline result, including evidence classification, action kind, status, target, step, replay route, and the pipeline/result.json plus pipeline/evidence.json artifact names.

POST /api/project/save¶

JSON

{"name": "my_network", "state": {"sourceMode": "model", ...}}

Returns path-free save evidence:

JSON

{
  "evidence_classification": "project_workspace",
  "name": "my_network",
  "project_sha256": "<64 lowercase hex characters>",
  "saved_at": 1711504200.0,
  "schema_version": "studio.project-save.v1",
  "state_sha256": "<64 lowercase hex characters>",
  "status": "completed",
  "version": "0.3.0"
}

The frontend stores this response and renders it as the latest project-save evidence strip in the Projects panel. The strip uses digest labels only; it does not render server filesystem paths.

POST /api/pipeline/run¶

JSON

{
  "graph": {
    "populations": [...],
    "projections": [...],
    "duration": 30.0,
    "dt": 1.0
  },
  "target": "ice40",
  "q_format": "Q8.8"
}

Returns, for a supported network:

JSON

{
  "success": true,
  "step": "synthesise",
  "target": "ice40",
  "pipeline": "graph → simulate → lower → co-simulate → synthesise",
  "steps": {
    "validate": {"passed": true},
    "simulate": {"n_spikes": 48, "n_total": 5},
    "lower": {"input_sha256": "…", "q_format": "Q8.8", "populations": [...], "notes": []},
    "cosimulate": {
      "steps": 30,
      "rtl_matches_bit_true_model": true,
      "studio_agreement": {"identical": true, "first_divergent_step": null},
      "spike_counts": {"rtl": 48, "bit_true_model": 48, "studio": 48}
    },
    "synthesise": {"success": true, "target": "ice40", "resources": {...}}
  },
  "trace": {
    "input_sha256": "…",
    "rtl_sha256": "…",
    "bit_true_model_sha256": "…",
    "synthesis_source_sha256": "…"
  }
}

A graph the lowering refuses stops at "step": "lower" with every reason in reasons.