Verification

Build first, tune last

A mechanic is done when it works and moves the right stock in the right direction. Tuning its constants to real-tank numbers comes after the mechanics around it exist, because every new mechanic shifts the chemistry the old ones were tuned against.

Whole tanks: the scenario runner

Whole-tank behaviour is judged by running the preset tanks headless — nano, low-tech, high-tech, community, sword-gravel, low-flow and cold, and a guppy and a neon pair left to breed, whose head count is the reading — on a keeper’s schedule, sampled at fixed days and always on the last one. Each reading is graded against a plausibility band:

Grade	Meaning
Green	Inside what a real tank of that kind shows
Amber	Unusual but possible — worth a look
Red	Not something a real tank does

Per-action flags replay a tank under a different keeper: more food, no water changes, extra fish. A run saved as a baseline can be diffed against a later one, so a change shows as the few readings it moved. Amber and red are questions to reason about, not failures to fix — a constant is never moved during build-out just to turn a cell green.

Tests: formulas and invariants

The test suite pins what holds regardless of tuning:

Kind	Example
Formula	A function computes what it says
Scaling	Doubling capacity doubles flow
Conservation	Carbon, nitrogen and every mineral close across rot, the nitrogen chain, every body and the gases
Soundness	No reading goes NaN or infinite
Determinism	The same seed runs the same life

A test never pins a whole-tank outcome or a coefficient’s value; those move every time a mechanic lands.

Source

The scenario runner and its bands live under src/cli/scenarios/, the invariant tests under src/simulation/tests/, and the constants under src/simulation/config/.