Jaxonomy — hybrid dynamical systems on JAX

Generate a plain English description of what a Jaxonomy model does. Args: model_json: JSON string of the model Returns a human-readable description of the model structure, blocks, and signal flow.

Fit model parameters to measured data using finite-difference gradients. Args: model_json: JSON string of the base model data_csv: CSV string with columns for time and measured signals. First column must be time. Example: "t,position,velocity\n0.0,0.0,1.0\n..." signal_map: JSON string mapping CSV column names to diagram signal names. Example: '{"position": "integ.out_0"}' params_to_fit: list of dot-notation parameter paths to optimize. Example: ["motor.R", "motor.L"] n_steps: number of optimizer iterations (for scipy methods, used as maxiter) learning_rate: optimizer learning rate (optax methods only) method: optimizer to use. One of: "adam" (default), "sgd", "rmsprop" — Optax gradient descent with central finite-difference gradients; "l_bfgs_b" — scipy L-BFGS-B with FD gradients; "nelder_mead" — scipy Nelder-Mead, gradient-free. bounds: list of [lo, hi] per parameter (same order as params_to_fit), or null for unbounded. Use null on either side for a one-sided bound. Example: [[0.0, 10.0], [null, 5.0], [0.0, null]] Returns JSON with: fitted_model_json: updated model with fitted params final_loss: final MSE loss value loss_history: loss values over iterations (sampled) fitted_params: {param_name: fitted_value} converged: bool (loss decreased by > 10x) n_iter: actual number of optimizer iterations completed n_fev: total number of loss function evaluations (approximate) convergence_message: human-readable convergence status relative_decrease: (init_loss - final_loss) / init_loss Note: Gradients are computed with central finite differences on the simulation loss so the full simulator need not be JAX-traced. For high-dimensional fits, prefer the dedicated optimization APIs in ``jaxono…

Serialize the sensitivity-weighted neighbourhood of one or more signals. Answers "what actually drives this signal, and by how much" on a model too large to serialize whole. The model's dependency structure supplies the edges; autodiff supplies the weights; expansion takes the strongest edges first and stops at a token budget, so what gets dropped is what mattered least. Weights are relative (dimensionless) sensitivities: a path's product is the relative end-to-end sensitivity, so 0.4 reads as "a 1% change here is a 0.4% change there". Edges into a continuous state are scaled by ``tau`` seconds, since a state derivative is a rate rather than a gain. Args: model_json: JSON string of the model focus: Comma-separated focus points. Each may be a block name ("plant", which expands to all of that block's signals), a node id ("plant:out:out_0", "plant:xc"), or any unambiguous fragment. budget_tokens: Approximate ceiling on the returned text (~4 chars/token) hops: Graph edges to expand from the focus. Nodes are signals, so crossing one block costs two hops. threshold: Minimum edge |weight| to include tau: Seconds represented by an edge into a continuous state probe: Relative step for the secant cross-check on zero-derivative edges (0 disables). Use a non-zero value on models with saturation, quantization, or dead zones, whose exact local derivative is zero where the block is flat even though the connection carries information. input_values: JSON object fixing input ports, as 'block.in_0': value Returns JSON with: text: the rendered context, one line per signal and per edge nodes, edges: the same content structured blocks: block name paths included estimated_tokens, dropped_edges: what the budget cost conventions: how to read a weight

Linearize a model around an operating point. Args: model_json: JSON string of the model state_values: JSON string mapping state names to values at operating point, or a JSON array of continuous state components for a single continuous-state subsystem (in diagram order). Example: '{"x": 0.0, "v": 1.0}' or '[0.0, 1.0]' input_values: JSON string mapping input names to values. Example: '{"u": 0.5}' as 'block.in_0': value Returns JSON with: A: state matrix (list of lists) B: input matrix (list of lists) C: output matrix (list of lists) D: feedthrough matrix (list of lists) eigenvalues: list of {real, imag} dicts is_stable: bool

List all available block types in the Jaxonomy library. Returns JSON with block names, descriptions, and key parameters. Use this before create_diagram to know what blocks are available.

Run a Jaxonomy simulation. Args: model_json: JSON string of the model t_start: simulation start time (seconds) t_stop: simulation end time (seconds) recorded_signals: list of signal names to record. Use the block name and port, e.g. ["integrator.out_0", "gain.out_0"] backend: "jax" (fast, GPU) or "numpy" (compatible) max_major_steps: maximum simulation steps (increase for longer simulations) Returns JSON with: time: list of time points signals: {name: list of values} final_time: actual simulation end time

Validate a Jaxonomy model JSON string. Args: model_json: JSON string of the model (Jaxonomy model format) Returns JSON with: valid: bool errors: list of error strings warnings: list of warning strings