meadows/docs/adr/0004-rate-rules-not-formulas.md

Behaviour comes from rules over Information Links, not free-form formulas

Part of meadows · see DESIGN.md.

Numeric simulation (phase 2) makes a Model alive: Stocks accumulate over time, Flows and Converters recompute each instant. Two coupled decisions shape how a Model carries the numbers, and both trade expressive power for valid by construction — the right trade for a tool that exists to popularise systems thinking, not to compete with Vensim.

Decision

1. The Information Link is the declared dependency. A Flow's or Converter's inputs are exactly the elements that link into it. There is no separate "equation references" namespace to keep in sync — the wiring you draw is the wiring the simulator reads. The same signed graph the loop detector walks (ADR-0001) is the graph the simulator integrates, so the loops you see classified R/B are the loops you run. They can never disagree.

2. A Flow/Converter computes from a small fixed vocabulary of Rules, not a typed-in formula. Each instantaneous element picks one rule and a plain number or two — never an expression:

Rule	Value	Reads (via Information Links)	Emergent behaviour
Constant	a fixed number	nothing	linear Stock change
Proportional	factor × (its + inputs)	the +-polarity inputs	exponential growth / decay
Gap	factor × (level − target)	the + input is level, − is target	goal-seeking / asymptotic

The famous curves are compositions of these over the structure — a logistic S-curve is Proportional growth meeting a Gap-driven ceiling (limits-to-growth); goal-seeking decay is a lone Gap (coffee cooling). The user sets up a local rule; the global shape emerges. That emergence is the lesson.

Polarity does double duty. The +/− already captured for loop classification (ADR-0001) also selects each operand's role: Proportional reads its + inputs; Gap reads its + input as the level and its − input as the target. One gesture, two payoffs — no new per-link data.

Considered Options

• Free-form expression strings (birth rate = Population × fertility) — maximally expressive, and what equation?: string originally anticipated. Rejected: needs a parser + a sandbox (never eval/new Function), invites broken-formula and name-resolution errors (auto-names contain spaces), and lets a learner paint a curve instead of discovering it from structure.
• Pick the output curve (label a Stock "exponential" / "logarithmic") — rejected: it is the answer, not the cause, and it breaks the moment feedback decides the shape. "Logarithmic" in particular has no honest local rule; what people mean by it is asymptotic approach — which is the Gap rule.
• Rules over Information Links (chosen) — no parser, valid by construction, and it teaches structure → behaviour.

Consequences

• The domain types gain a Rule union on Flow/Converter (replacing the unused equation?: string), an optional initialValue on Stock, and an optional SimSpec (start / stop / dt) on the Model. All optional and additive, so existing saved Models still load (F8); they are simply not simulatable until equipped.
• Algebraic loops are an error. A cycle in the wiring is legitimate feedback iff it passes through a Stock — the Stock supplies last-step state and so breaks the within-step dependency. A cycle among only Flows/Converters has no Stock to break it: the simulator cannot order it and rejects it. The simulator reuses the cycle machinery to detect this; it is a new sim-readiness check, distinct from structural validity (validation.ts).
• A reader seeing { kind: "gap", factor } on a Flow and wondering where its operands come from should look here: they are the Flow's inbound Information Links, picked by Polarity.
• The vocabulary starts deliberately small (Constant / Proportional / Gap — enough for linear, exponential, and goal-seeking, and for the coffee and savings samples). Growing it is additive: a new kind in the union plus a case in the evaluator. Multi-input products (e.g. Population × fertility) are a later increment, not a phase-2 blocker.