A free-text "why this element is here" note on Stocks, Flows,
Converters, and Information Links — editable in the Inspector next to
the unit, shown read-only in the gloss card below the generic kind
definition on selection. The Inspector now also opens for Information
Links, which previously had no editing surface.
The model had no downward force on Performance, so its only flow
(improvement) could only pull it up. Performance rose 40 -> 60 while
the Standard fell 80 -> 60, contradicting the "slide downhill" story.
Add a constant decay outflow on Performance and seat it just below the
goal. Now both ratchet down in lockstep (Performance 70 -> 10, Standard
80 -> 20), a fixed gap apart — the genuine eroding-goals trap. Rewrite
the doc-comment, layout note, overview line, and blurb to match.
Drops the Allee threshold to ~30 and lifts carrying capacity to ~1170, so
the stock is hardier and only collapses once heavily overfished. The
overshoot-and-collapse-to-extinction arc holds: Fish drift up, the fleet
overshoots (Boats peak ~700), and the catch drives the stock past the
threshold to a permanent zero by t=150.
The non-renewable Resource (a Stock with no inflow) becomes a renewable
fishery with a true point of no return — an Allee threshold. Spawning
scales with density (~Fish^2) while natural deaths are linear (~Fish), so
below a critical density the deaths win and the stock slides to an
extinction it never recovers from; crowding deaths (~Fish^3) cap a healthy
stock at carrying capacity. A reinvesting fleet (Boats, Reinforcing)
overshoots the renewal rate and drags the fish under the threshold, then
starves and scraps itself.
The negative-positive-negative regrowth curve is the one shape the
proportional rule can't draw alone, so two relay Converters build it
(density to lift spawning to ~Fish^2, crowding for the ~Fish^3 ceiling) —
the Limits-to-growth crowding trick, doubled. At 16 nodes this is the
gallery's largest model and the only one with a Converter feeding a
Converter.
Tuned against the engine: Fish hold near 1000, cross the threshold (~200)
at t~40 as the catch overshoots, then go extinct and stay there; Boats
overshoot to ~450 and collapse back near their start by t=150. No
divergence; loops classify as expected (R: fleet reinvestment, birth
engine; B: natural/crowding/catch drains, scrapping).
The results panel floated over the canvas and hid the lower part of the
diagram. Dock it beneath the canvas instead, and refit the view when it
opens or closes so the whole diagram stays framed in the height that
remains. The refit fires on Vue Flow's `dimensions` (post-measure), not
on the toggle, so it frames against the settled size.
The thirteenth sample points the language at a live debate: a classic trap,
Shifting the burden to the intervenor (Thinking in Systems, ch. 5), with AI as
the intervenor. Technical debt drives AI reliance, AI churns debt and lets
skills atrophy, and a thinner-skilled team refactors less — so the loop
Technical debt -> AI reliance -> atrophy -> Expertise -> refactoring ->
Technical debt is Reinforcing: addiction, not a fix. A learning Balancing brake
(practice toward a skill ceiling) is tuned to lose.
Maps the brief onto valid roles: Expertise and Technical debt are the stocks;
code quality and lead time read off them (inverse of debt / inverse of
expertise); model price lives in the AI-reliance factor. Buildable on the
existing rule vocabulary. Tuned against the simulator: over t=0..50 Expertise
slides 70 -> ~6 and Technical debt spirals 20 -> ~150, stopped (like
"Escalation") before the loop runs off-chart. The loop detector classifies the
four loops as expected (2 R, 2 B).
The twelfth sample fills the one classic dynamic the gallery was missing —
overshoot and collapse, the dynamic Thinking in Systems is built around. A
Reinforcing engine (Capital reinvesting its extraction revenue) runs on a
non-renewable Resource, the first Stock in the gallery with no inflow: it
overshoots the limit instead of settling at it, the dark twin of "Limits to
growth".
Buildable with the existing rule vocabulary (proportional only); the
non-negative-stock floor tames the bilinear extraction term as the Resource
runs out. Tuned against the simulator: Capital 5 → ~250 (t≈39) → ~5 by t=150,
Resource 1000 → ~6, no divergence. The loop detector classifies the three
loops as expected (R: investment→Capital→extraction; B: depreciation→Capital;
B: extraction→Resource).
The in-flow value line grew the card downward, dropping its vertical centre —
and the left/right handles with it, so connected pipes and links shifted. Render
the value as an out-of-flow overlay below the box; the card keeps its size and
the handles stay put.
valToPos returns CSS pixels relative to the plot area, but the overlay
sits on the canvas, so the playhead was short by the y-axis width and
started before x=0.
The fixed-width time readout wrapped to two lines for fractional times (dt 0.25
→ 't = 70.25'), growing the panel a row each frame and reflowing the whole
panel — the real flicker. Pin its width and forbid wrapping.
The default (no data-theme) resolves to DaisyUI's dark theme, whose
:root:not([data-theme]) selector outranks a custom light theme — so the
previous light-only override never reached buttons in dark mode. Force it with
an !important :root override that wins regardless of theme.
The playhead was a canvas line redrawn by redrawing the whole uPlot chart on
every frame, which flickers. Draw it as a DOM overlay that slides over the
canvas instead, so the canvas is only redrawn when the data actually changes.
A Stock can carry a display unit (°C, people, $, …): a new optional field, an
inspector text input, validation + round-trip in io, and a readout beside the
live value on the canvas. Equip the demonstrative samples (Water L, Balance $,
Coffee °C, Population/Epidemic people, Yeast cells). Also bumps the fill gauge
opacity for a more vivid level.
A simulation store owns one run and one playhead, so the chart and the canvas
read off the same instant. While the panel is open, Stocks show a live value and
a fill gauge, Flows and Converters their current value; a play/pause + scrubber
drives the playhead (a wall-clock-paced clock in usePlayback), and a marker
tracks it on the chart.
Draw the simulation's Stock trajectories with uPlot (ADR-0005): a real time
axis, hover-to-read values, and a playhead-marker hook the canvas animation
drives. SimChart owns the imperative uPlot lifecycle behind a small prop API.
Switch the body font to Sono and point the @import at
api.fonts.coollabs.io. The bare fonts.coollabs.io host serves the
marketing homepage (HTML), not CSS, so the @font-face rules never
loaded and the font silently fell back to system-ui.
Recast the S-curve as a Reinforcing inflow plus a crowding-driven die-off that
grows with Yeast², using only the existing proportional rule. Yeast now climbs
20 → ~1000 as a true sigmoid, and the detector still classifies it R + B — the
balancing loop stays visible, which a single "logistic" rule would have hidden.
Drops the carrying-capacity converter (a faithful one needs a divide rule).
- ResultsPanel: a Simulate panel that charts every Stock over time (hand-built
SVG), with start/stop/dt run controls and a divergence warning
- Inspector: edit a selected Stock's initial value or a Flow/Converter's rule
- store: undoable setInitialValue / setRule / setSimSpec actions
Give 10 of 11 samples initial values, rules, and a run window so they simulate
on load: Bathtub (linear), Savings/Escalation/Fixes-that-fail (Reinforcing),
Coffee (Balancing), Population, Epidemic (SIR), Tragedy (overshoot/collapse),
Drift (eroding goal), Predator-prey (damped). Limits to growth is left
diagram-only — its S-curve needs a saturating rule the vocabulary lacks.
Phase 2 brings Models to life. ADR-0004: behaviour comes from a small fixed
vocabulary of rules (Constant / Proportional / Gap) read over the inbound
Information Links, not free-form formulas — valid by construction.
- types: Rule union, SimSpec, initialValue; replaces the unused equation field
- simulation: forward-Euler engine, dependency-ordered evaluation, algebraic-loop
detection, non-negative stocks, and a divergence guard
- io: validate and round-trip the new fields (F8)
Vue Flow's open arrow renders at 0.625x the edge stroke (its 20-unit
viewBox is squeezed into a 12.5 marker box), so the chevron looked thin
against the 2.5px pipe. A marker strokeWidth of 20/12.5 cancels the
downscale, drawing the head at the line's weight at any width.
Rename the coffee outflow "heat loss" to "cooling" so the balancing loop
reads more naturally, and stack carrying capacity above the source in
limits to growth so its link to crowding is a clean horizontal hop.
Swap the filled-triangle arrowhead for an open arrow on flow pipes and
information links, and route flow pipes through a custom cubic edge whose
horizontal flatten arm scales with the vertical drop, not just the
horizontal gap. Vue Flow's default arm collapses toward zero when the
valve sits nearly above the stock, so a steep pipe stayed flat only in
its last pixels and the arrowhead detached from the visible line. The
pipe now keeps a real horizontal run into the stock's left edge at any
approach angle.
Snap all sample coordinates to multiples of 20 so they sit on the
editor grid and read cleanly. Computed midpoint() valves had to land
on grid points too, so a few endpoints were nudged (e.g. escalation
arsenals 300->280, limits-to-growth source -300->-280). Population is
re-laid as a grid cascade so its links no longer overlap the pipes.
Two more Meadows system traps: Fixes that fail (a balancing
road-building fix whose induced-driving side effect reinforces the
congestion it targets) and Drift to low performance (an eroding goal
that spirals actual performance and the standard downward).
Also re-lays-out the coupled-stock samples for readability: predator
and prey and escalation now use aligned two-row layouts, and fixes
that fail stacks its flow valves in one column so the backfire link
reads as a clean vertical.
Add a second tier past the four-model primer: Limits to growth, Predator
and prey, and Epidemic each introduce one structure the primer never
shows, plus two of Donella Meadows' system traps — Tragedy of the
commons and Escalation — to contrast healthy dynamics.
Dragging a Flow back onto a Stock it already touches is refused (an
Information Link never targets a Stock), but that gesture is almost
always someone trying to "close the loop" by hand. The Flow's own pipe
already is the link back, so surface a self-dismissing hint saying so
instead of letting the drop die silently (F4: guide, don't nag).
Restoring on reload fit the view synchronously in onRestore, framing
against unmeasured 0×0 nodes, so the model landed jammed top-left. Arm
the same post-measure onNodesInitialized fit the sample/import loads use.
pnpm 11 defaults strictDepBuilds to true, so a dependency build script
that is not pre-decided in allowBuilds becomes a hard error. The install
layer only copied package.json and the lockfile, so the vue-demi build
decision in pnpm-workspace.yaml was absent and the build failed.
Badge centroids now read Vue Flow's live computedPosition instead of the
store, which commits positions only on drop — so badges no longer lag a
node drag. Centres use real node dimensions, dropping the hard-coded nudge.
