Symulab — what we've built so far.

The headline: Symulab now runs on Windows. The installer carries the complete Julia compute engine — runtime, packages, and a precompiled system image — so simulation works offline out of the box, exactly like the Mac build. No Julia install, no downloads on first launch. Behind the scenes this took a full rework of the build pipeline; the practical result is a 396 MB installer produced automatically from CI for every release.

The equations had an audit. Because a modeling tool is only as good as its numbers, we traced every step from the expression you type to the value in the chart — and fixed seven real bugs. The two that mattered most produced silently wrong results: Stella-imported RAMP(slope) and short-form PULSE(...) calls put the simulation time into the wrong argument slot (a 1-arg RAMP was permanently zero), and a scalar flow connected to an arrayed stock moved nothing at all, with no warning. Both now behave correctly — and the mismatch case is a clear error in the editor and the engine, so it can never fail quietly again. The rest of the pass: LN/INT/uppercase math functions from Stella files now translate instead of erroring, the constant e works in expressions, two-dimensional array subscripts like pop[North,Old] compile correctly, SUM/MIN of a non-arrayed variable behave as identity instead of erroring, and init expressions are validated up front with did-you-mean suggestions.

Trust, but verify. The engine now ships with a verification suite that runs the real worker process against models with known closed-form solutions — exponential decay and logistic growth across Euler, RK4 and Tsit5, integrator convergence-order checks, converter dependency chains, array aggregation, bounds clamping, and every helper function. The full pipeline — how groups, dimensions, expressions, validation, and integration actually work — is documented in docs/equations.md.

The canvas feels right now. A drag-and-drop overhaul fixed the things you’d notice within a minute of modeling: dragging a group now carries its members with it; dragging a stock carries the source and sink valves wired to it; the flow-to-stock magnetic snap no longer fights the alignment grid, so butt-joints stay flush; and pipes render as straight, rigid lines with an arrowhead into the receiving stock, so flow direction is always readable. Interaction got richer too — Shift+drag draws a selection marquee, Cmd+G groups everything selected, double-clicking a node jumps straight to renaming it, and the canvas zooms out to 10% for big models. Under the hood, node positions are now committed once on drop rather than on every mouse move, which keeps large diagrams smooth and makes one drag equal one undo step. Smaller touches: a live size readout while resizing, connection handles that appear on hover instead of dotting every node, and consistent array badges across node types.

Known follow-ups queued for the next releases: a direct Windows download link on this site (the build currently ships via GitHub Releases), Windows code signing to quiet SmartScreen, Stella SMOOTH/DELAY import support, and graphical (lookup-table) functions.

A polish release driven by QA feedback. Most of the work is on the authoring surface — sizes, names, animations, and a couple of edge cases that made the canvas feel rougher than it had to.

Resizable nodes. Every node type now has the standard React Flow resize handles (visible only when selected). Stocks resize freely; hold Shift while dragging to lock the current aspect ratio. Flows stay locked to a 3:2 ratio so the valve geometry holds its shape. Converters stay 1:1 — the circle is the circle, not an ellipse. Defaults are 120 × 80 for Stock and Flow, 80 × 80 for Converter. Sizes persist in the model JSON, so a laid-out diagram round-trips cleanly.

Friendlier labels. Names with underscores (the engine’s required identifier form) now render on the canvas with spaces — birth_rate_max becomes birth rate max, Hares_Hunted_Per_Lynx becomes Hares Hunted Per Lynx. The underscored form is preserved everywhere that needs it: in equations, in CSV exports, in the Dashboard, in chart legends and tooltips. The transform is canvas-only, so identifiers stay greppable and matchable to the Julia engine’s symbol names.

Real Max bounds. Stock and Converter Max (and Min) values previously sat in the UI but were silently dropped by the engine — only flow rates were clamped. They now flow through end-to-end: initial values are clamped at start, the integrator gates derivatives at the bound during the run, computed converter values are clamped each step, and the recorded series are clamped at the end so any solver overshoot doesn’t leak into exports. A logistic model with population.max = 400 now clamps cleanly at 400 instead of climbing to the natural carrying capacity.

Flow valve redesign. The flow’s center regulator was previously a 12-pixel knob lost between the bowtie shoulders. It now reads at converter scale — a 44-pixel central circle, with the bowtie kept as flanking accents and the pipes shortened to 20-pixel stubs that exit the sides. The valve still looks like a Stella valve; it just doesn’t disappear next to a converter anymore.

Animate connectors toggle. Pipes (stock ↔ flow) previously sat still while influence edges marched — visually inconsistent. The animation now applies to every connector type, and a new toggle in Display → Connector Style lets you turn it off entirely. Defaults to on for new models; existing models with no display.animate default to on too.

First-order feedback wiring. Drawing a Population → Deaths influence arrow on top of an existing pipe used to be a silent no-op — the connection logic treated the second drag as a redundant pipe edit. The connect handler now detects the case (source stock is already this flow’s from or to) and creates the influence edge instead, so the diagram can actually show the feedback loop. Doesn’t fire for unrelated stocks.

Quieter fixes.

• Fresh canvas (zero nodes) no longer flags a red error in the toolbar; the validation badge stays OK until something is actually wrong.
• Clicking Run on an empty model returns empty results instead of throwing — matches the new validation badge behavior.
• The flow-to-stock magnetic snap now reads each node’s actual width and height, so resized stocks and flows snap correctly to their real edges, not a default-sized phantom box.
• The Julia engine package (SmsEngine) now declares SparseArrays and LinearAlgebra as direct dependencies — the disaggregation tool used them but the package wasn’t listing them, which broke a clean precompile when the sysimage was rebuilt from scratch.

Known follow-ups queued from the QA review and not in this release: unit-of-measure support with consistency checking, auto-inference of input/output variables from rate expressions, and movable waypoints on the connector lines. Those each need a brainstorm before they ship.

A maintenance release with one new tool, observability, and a CI cleanup.

Temporal disaggregation (work in progress). A new entry under Tools → Disaggregate Data… opens a standalone utility for converting low-frequency series (e.g. annual totals) into a higher-frequency series (quarterly, monthly, …) using the Denton Proportional First Differences method, with an optional indicator column for benchmarking-style runs. Load a CSV, pick the columns and the split factor, and the result comes back as a chart, a table, and a downloadable CSV. The engine and the core flow are in place; rough edges, edge cases, and broader method support are still being worked through — treat this build as a preview.

Sentry monitoring. The renderer now reports uncaught errors, performance traces, and (sampled) session replays to Sentry. Anything that crashes a panel or trips an unexpected code path should now leave a trail we can actually look at.

Build & Release pipeline. The GitHub Actions workflow that produces shippable artifacts has been reduced to Windows-only. Mac and Linux lanes have been removed; the workflow runs faster and the only artifact it publishes is the Windows installer.

Other fixes. A handful of smaller fixes landed across the canvas, validation, and simulation engine — the kind of polish that doesn’t fit a bullet but keeps the desktop experience steady.

This is a polish release focused on the authoring surface. Three themes:

Canvas behaviour. Dragging a flow within ~30 px of a stock’s left or right edge now magnetises the flow against the stock, aligns it vertically, and sets flow.from or flow.to automatically based on which side the stock is on. Only the closer side wires, so moving one end of an already-connected flow never accidentally drops the other. Snap thresholds live in src/lib/snap.ts if you want to tune the catch zone.

Visual continuity. With flows sitting flush against stocks so much more often, anything breaking the visual continuity at the join had to go. The pipe-handle dots on flows are now transparent, the synthetic pipe-edges pick up the flow’s own color so there’s no seam at the boundary, the right-side arrowhead vanishes when the right end is wired into a stock, and the flow’s wrapper is locked to its SVG width so long rate-expression labels can no longer push the SVG sideways. Source / sink clouds shrank by 10 % to match.

Labels. Each diagram now picks where the variable name and the value / expression label sit — top, bottom, inside the body, or hidden — independently for the two label kinds. New models default to name above and value below, which matches Stella / Vensim conventions and keeps the stock and converter glyphs uncluttered. Labels are absolutely positioned so they never affect the node’s bounding box — snap math stays exact no matter how long the name is. The chosen positions persist as a user preference and apply to every fresh model.

Quality-of-life in the toolbar. A small chevron beside Open drops down the eight most-recent models — pulled from the preferences store, pruned when a file is missing. The palette popover now reports how many overrides are active (per-slot plus per-object) and lets you wipe all of them with a single confirm.

Distribution. The marketing site now lives on a Cloudflare Worker named stocksflow-web, deployed via wrangler login and the new scripts/deploy-web.sh. The macOS download button on this page points at the CloudFront distribution in front of the symody-artifacts S3 bucket — stable URL, no signing, no expiry.

This is the version of the app currently in testers’ hands. The list above is everything wired up end-to-end. The notes below give context.

Authoring. The canvas takes the three SD primitives plus influence links. Names, initial values, and rate expressions edit inline through the properties panel. The DimensionManager lets you define named dimensions (e.g. Region = {N, S, E, W}) and apply them to any node — the ArrayBadge on the node shows what’s active and the engine expands to the right number of state variables at run time.

Persistence. One .json file per model. Diagram, equations, and simulation settings all in one place. Diffable in git. The five bundled sample models in /models exercise different parts of the engine — logistic_growth.json is the smoke test, predator_prey_stella.json plus Predator-Prey_(Stella)_results.csv is the faithfulness check against Stella, civilisation_model.json and world_dynamics.json are the heavier stress tests.

Engine. Tauri spawns the bundled julia with our sysimage flag, hands off paths to a temp input and a temp output file, and reads the result back. The Julia side (SmsEngine.jl + run_model.jl) parses the model, builds the ODEProblem, solves with Tsit5 from OrdinaryDiffEq, and serializes the trajectory back as JSON. The sysimage gets first-run latency down from ~15s to ~2s.

Results. Three tabs in the results panel. Chart is the Recharts time-series with a checkbox per series. Table is the raw step-by-step values, with a per-column CSV export. Equations shows the system of equations the engine assembled — useful for sanity-checking that what you drew is what the engine actually solved.

Distribution. GitHub Actions builds the matrix (macOS / Windows / Linux), packages the installer with the bundled Julia runtime and sysimage, and uploads to Cloudflare R2 under releases/{version}/. The download links on this page point at that bucket. macOS build is signed; Windows is not yet; notarization is pending.