Workflow Graph
Actual Svelte Graph canvas connected to the loopback host, with backend-produced graph and definitions. Editor submission is outside this host’s capabilities.
LOCAL-FIRST · RUST-NATIVE · INSPECTABLE
Make the workflow visible.
Make execution accountable.
Pantograph connects a visual node editor with scheduler-owned execution and explicit inference boundaries. Explore one working deterministic workflow, then the research behind the larger system.
Captured 11 October 2026 · CPU-only JSON extraction · No model inference claimed
APPLICATION GALLERY
Inspect the actual workbench.
Genuine Chromium captures of existing Svelte components connected to a persistent Rust service over loopback HTTP. These are not native desktop windows. The workbench queries live diagnostics and reads actual retained artifact bytes; this static page preserves the recordings. Inspect the source recording. Review branch: PR #71; main awaits review.
Execution ledger / Diagnostics
Real Diagnostics component querying the live native service, including completed non-runtime node statuses.
Ledger timeline
The actual native task lifecycle timeline, scrolled into view. Inspect completed node statuses and the three retained payloads.
Workflow-run I/O Inspector
Real inspection component with a live run graph, node status and retained text read through the artifact API.
Model Library
Actual Library component reading a real isolated empty Pumas catalog. Unsupported model-management actions are explained and disabled; no invented models or inference results.
01 / THE DEMONSTRATION
One graph. Three perspectives.
The real Svelte node editor
RENDERING VERIFIEDSelection Input → JSON Filter → Text Output. Backend-produced node definitions and the demo graph are loaded into the application's existing stores. Ports and edges are rendered by the actual editor.
Genuine browser capture, with focused fixes for JSON Filter's renderer registration and aligned-wire visibility. The application is connected to the loopback demo host. Graph editing/submission is outside this host’s capabilities; execute through the demo controls and inspect live results in the workbench. The separate Node Editor / Node Lab authoring feature remains unavailable in current source.
A real scheduler run, with its receipt
NATIVE EXECUTION VERIFIEDThe Rust workflow service validates and snapshots the graph, admits a session run, materializes the JSON input, executes JSON Filter and Text Output, records diagnostics in SQLite, and retains the small scalar output artifacts. No host-returned fixture output is used.
Loading recorded evidence…
Select an event to inspect its original payload.
This interactive view reads captured execution JSON; it does not run Rust in your browser. Local elapsed time is not a benchmark. The rejected run records an invalid downstream task when the path is absent. No GPU scheduling, cache reuse, or model performance was tested here.
Keep orchestration separate from model execution
ARCHITECTURE · NOT AN INFERENCE RESULTGraph and session contracts belong to the workflow service. Scheduler policy owns admission and task execution. Runtime hosts own model loading and execution, while the model library provides artifact and load-target facts.
- Workflow serviceValidated graph, immutable run snapshot, typed inputs and outputs
- SchedulerAdmission, task attempts, lifecycle evidence and resource policy
- Runtime host / inference adapterBackend-specific loading, readiness, execution and cleanup
- Model libraryPackage identity, artifacts and load-target resolution
The deterministic demo exercises the service and non-runtime task path. It does not load an inference backend. Unpublished Pumas fixes and model tests on the developer's computer are outside this checkout and outside this evidence.
02 / RESEARCH & DESIGN
The reasoning behind the system.
Original manuscripts and an evaluation report. Research editions are pinned to 2 October 2026 and source revision 4938e405; they are distinct from this later execution evidence.
NODES · PAPER
Node Execution and Dynamic Extensions
Invocation boundaries, lifecycle, extension admission, and what it takes to qualify a node.
Source audit + proposed contracts → SCHEDULER · PAPERA Resource Safe Learning Scheduler
Resource safety, learned execution costs, and completion-oriented scheduling.
Inspected constraints + proposed v2 design → SCHEDULER · SIMULATIONScheduler v2 Evaluation
C3-corrected standalone simulation evidence, with historical populations kept separate.
Simulator results; not a production benchmark → INFERENCE · BOOKDesigning Compatible Inference Backends
Model identity, semantic task contracts, backend ownership, and qualification before promises.
Source-grounded comparison + architecture proposals → EVIDENCE · PAPERExecution Ledger and Provenance
Identity, result commitment, retention, crash consistency, and evaluation contracts.
Source audit + proposed evidence contracts → RESEARCH INDEXSources, editions & limits
Immutable source references, manuscript identities, and the distinction between research and accepted implementation work.
Read the evidence policy →Small enough to inspect. Real enough to test.
With the repository's pinned Rust toolchain and Python 3:
python3 examples/scheduler-demo/serve.py
Open http://127.0.0.1:8765. Run the valid input, change the prompt, then try the missing path. Each response includes the actual run identity and its recorded diagnostic payloads. To inspect the running service in the real Svelte workbench, start npm run dev in another terminal and follow the demo host’s Open the live workbench link. See the pinned guide for prerequisites and host limits.