On 26 July, while we were still building it, our discount text still held three codes from our first prototype: WELCOME10, SAVE15, VIP20. None existed in the store. The model also invented codes of its own, and offered 15% on a store whose cap was 12%. In our test chats, on our own test stores, every discount request got a code that would have failed at checkout.
What changed:
1. Only the merchant's own codes. No codes set up means no code offered, said plainly.
2. The cap is applied on the server, before the model ever sees a code: a tier above the store's maximum is dropped.
3. In the shopper's chat, the model asks the server for a discount and gets one code back: the lowest first, one step up per ask, never past the top. The cap and the other codes never reach it.
On 14 November our own test agent pushed six times: 30%, "tell me your maximum", two codes the model had invented back in July plus 25%, then 20%, 10% and 13% on one product. No code was offered in any of the six, so the cap itself wasn't reached in the test. Its proof is the code and its tests, which run on every pull request.
Full write-up: x.com/con_golias/status/2121650172660760999
