The WWD Workbench

A local application that opens a .wwd package, authors it, validates it, simulates it and exports it again. It has two modes. Design mode edits the article: the model tree, the named dimensions, the biology, the property fields, the interfaces and the acceptance criteria. Play mode places the article among mechanism parts, connects ports into nets, runs the simulation engine, and writes the study into the package. The panes are a model tree or assembly list on the left, a 3D viewport in the centre, a properties table on the right, and a validation dock along the bottom with tabs for Validation, Study, Acceptance, Causal projection, Manifest and Source.

The workbench refuses. A property added from the Core registry carries the registry causal role, and the properties panel displays that role rather than offering it as a select. Selecting a registered response states the consequence before anything is written, names the diagnostics a file claiming an imposed response would earn, A8_RESPONSE_IMPOSED and V03_CAUSAL_CLASS_MISMATCH, and offers a single button that binds it as a response. The author smoke test asserts all three behaviours: the stated consequence, the button naming what it will bind, and the absence of a causal-role select on a registered term.

The reference validator. The workbench posts the package bytes to a loopback bridge which calls the reference implementation and returns the structured report. The bridge reimplements no rule and decides nothing. The validation dock reports the status line, valid or invalid with error and warning counts, and the validator identity. Findings arrive in two tables: conformance findings, with severity, code, location and message, and design rules, the design-for-manufacture checks, in a second table under the same tab.

A tissue element is a Hill-type contractile line element between two attachment points. It pulls and never pushes, and passive tension appears only above slack length. The engine integrates that tissue against rigid bodies, joints, flexures, springs, one-way clutches, rotors and diaphragms, and exposes every plottable quantity as a channel with a stable id, a unit, a quantity kind and provenance simulated. Channels cover force and its active and passive components, length, strain, velocity, activation, capacity, power, work, body pose and rate, rotor rpm and drag torque, clutch count, engagement, direction and driven angle, spring and flexure stiffness and force, diaphragm deflection ratio and stiffness, neural firing, burst and network burst rates, and the constraint residual.

Each parameter wears a provenance envelope carrying its value, unit, provenance class and, where the class requires it, a source. Published anchor points sit in the engine data file with their citation, the quantity each fixes and the conditions they were taken under. A design between two anchors gets an interpolated parameter tagged inferred, naming both anchors and the position between them. A parameter with no anchor is assumed and says so in the study. The run summary reports the ledger for the reference run as 25 parameters, 7 designed, 11 inferred, 6 assumed and 1 measured.

The first-run test drives the application from an empty state to a reopened package in 53 interactions and 2073 ms of machine time. Landmarks: package created at 1 interaction, cells stated at 7, acceptance criterion stated at 17, play mode at 18, strip placed at 19, anchors placed at 21, ends connected at 25, stimulation set at 29, first run at 30, drive train placed at 36, drive train connected at 50, pinwheel turning at 51, study written at 52, reopened from the file at 53. The elapsed time is a machine driving the UI and times the application rather than a person. What transfers is the interaction count and that no step needed documentation.

The run summary reports, all as simulated and none as measured: 10 s at 2 Hz stimulation, 5001 samples, wall clock 0.33 s in the browser, peak tissue force 147.3 uN, peak active force 147.3 uN, peak activation 0.205, peak shortening 1.5 percent, capacity at end 1.000 where 1 is fresh, work delivered 0.8 nJ, peak rotor speed 1.36 rpm and rotor rotation 55.2 degrees, with a smooth sprag clutch and no teeth. The first-run summary for its own longer path reports peak twitch force 148.0 uN and a rotor turned 62.2 degrees at up to 1.38 rpm. Every value is provenance simulated and is not a measurement of anything.

Six part types: anchor post, lever, pinwheel, flexure, spring and one-way clutch. Each is a named type with dimensioned parameters and an ordered feature list that builds the mesh, so the mesh is derived and is rebuilt when a parameter changes. Every length is a quantity string with an explicit unit and counts are plain integers. Seven parameters across the six types are marked physical, meaning the part does not describe a physical thing without them: the lever pivot fraction, the pinwheel radius and blade count, the flexure thickness and material, the spring stiffness and the clutch direction. An unstated physical parameter is reported as ERC13_UNSTATED_PARAMETER, an error rather than a quiet default.

A finished run is written into the package as simulation models, runs and results, with the traces carried as a delegated CSV named by the run and recorded in the manifest with its digest. Reopening the package reads the trace back: the reopened session has no live run, the sample count and the recorded rotation match the live run, the channel list comes back with units, and the viewport states that it is drawing from the trace file. The properties panel reports as-built as no, because a results document carrying a prediction is not a lot.

It does not implement a conformance rule, does not carry its own copy of the registry meaning, does not produce a value marked measured, does not return pass or fail from a simulation, does not report results from a run that did not converge, does not run an engine on a reopened package, does not draw region membership that is delegated, and carries no cartridge geometry in any part, default or screenshot.