One workshop, one month, every cent.

This is the demo dataset, a stool workshop buying oak, building stools, selling them. Seed it yourself (npm run seed) and every number below is in the database.

1, Two timber receipts, two cost layers

Same SKU, different costs. Averaging would call oak "$8.82" and move on. The engine keeps both layers and remembers the order it received them.

2, A sale consumes the oldest layer first

If the sale had needed 46 lengths, it would have eaten all 45 of layer #1 plus one from #2 at $9.20, and the COGS line would show both unit costs separately, not a blend.

3, Freight hits the stock you still have

A $85 freight bill lands Jan 11. It spreads across open layers by value; consumed history is never rewritten:

4, Building eight stools rolls the BOM up

Each stool eats 4 lengths of oak, 0.25 L of varnish, 1 screw pack. The build on Jan 15 consumes components FIFO across their layers and produces one finished layer whose unit cost is $48.250768, the actual component cost divided by 8, not a round number someone typed. That precision survives every later step because money is NUMERIC, never float.

5, Sales, then one journal

One draft manual journal for the day, balanced, pushed with an idempotency key. If the sync runs twice, Xero sees one journal. Void INV-214 next week and the app restocks the exact three stool-slices at $48.250768, reversing the effect. No adjust-and-hope.