Spum Reasoning and private-network execution

Spum Intelligence is the product family. It includes the Spum inference engine, Spum Reasoning for checked scientific and educational workflows, and Spum Afterglow for music continuation.

The SSH command is available in release 0.2.4. The Reasoning example uses a separately configured private DGX worker; the public page shows its recorded output.

Run an engine over SSH

spum remote --host lab-dgx --engine /opt/spum/spum -- capabilities
spum remote --host lab-dgx --engine /opt/spum/spum --gpu-device 0 -- \
  run "$REMOTE_MODEL" --gpu -p "Explain this observation."

--host accepts an SSH alias from your existing SSH configuration. --engine is the remote executable path. Remote model and input paths refer to files on that host; this command does not upload local model files or images. --ssh-port overrides the configured port. --gpu-device selects a remote CUDA device explicitly; choose a GPU you are authorized to use.

The system SSH client handles transport and key authentication. Known-host verification is required, agent forwarding is disabled, and the command never stores a password or silently accepts a host key. Establish and verify a new host through your normal SSH workflow first.

SSH does not establish where a server is physically located or whether its own tools contact an external service. To keep computation within your network, configure a machine and workers within that network and use local model weights and assets. The destination is printed before execution. Remote execution is distinct from inference on the Mac itself.

Spum Reasoning worker

Spum Reasoning combines deterministic and symbolic tools with checked diagrams and optional spoken explanations. The initial DGX integration reuses the existing scientific scene and graph validators, a symbolic solver, an independent numerical check, and the existing Maya voice service. It runs as an optional remote worker, separately from the native inference kernels.

The validated example evaluates the Gaussian integral, renders the curve and its displayed shaded region, and synthesizes a checked Hinglish explanation. The worker returns the exact result, LaTeX, decimal approximation, verification record, artifact checksums and measured timings. The report uses bundled offline MathJax; no external rendering CDN is needed.

spum remote --host lab-dgx --engine /opt/spum/spum-reasoning -- \
  reasoning --request-json '{"operation":"integrate","expression":"exp(-x**2)","variable":"x","lower":0,"upper":"inf","plot_min":0,"plot_max":3,"voice":true}' \
  --output /path/on/dgx/new-result-directory

The worker must be installed separately and its dependencies configured. Output paths are remote; retrieve generated artifacts with your normal SSH file-transfer workflow. The output folder must be new. The first integration has no native app job browser or automatic artifact download yet.

Mathematical scope

The first operation is a definite real integral. Input expressions use a restricted mathematical grammar: numeric constants, the selected variable (x, t, or z), pi, e, arithmetic, fixed powers, sin, cos, exp, sqrt, log and erf. Unknown syntax, programming constructs, unresolved integrals, divergent results and failed numerical checks are rejected. A job has a five-minute limit.

The diagram path currently requires a finite, non-negative curve over the supplied display interval. Plot limits and integration bounds are recorded separately. Checked Maya narration is initially available for the Gaussian example, not arbitrary solver output. This is a tested integration example, not evidence of correctness for every possible integral or educational topic.

Instruments

Inference and instruments can reside on different machines within the same network. Instrument control needs its own device adapter, units, travel limits and stop behavior. No physical-stage adapter or hardware movement has been validated in this integration. A model response must not be treated as a direct device command.