The price, stated plainly
docs/adr/0011-recipes.md
1. **Two of the three recipes are not tuned and do not claim to be.** `just-win`'s `C = 1` sits far outside anything either campaign searched: campaign 1's grid opened at 18 and campaign 2 only widened it upward. It is a value judgement taken to its logical end and nothing here claims it predicts well. 2. **No recipe has been backtested end to end as a system.** [`demo/2023-recipes.md`](../../demo/2023-recipes.md) is one season's final board, which compares outputs and does not evaluate them. `cfbpoll backtest` scores orderings, and a recipe as a whole is outside what it scores. 3. **The gate is not applied per recipe.** `[gate]` is written against the published poll, so an alternate lens publishes no gate verdict and its methodology document says why. A page showing three rankings and one verdict would be less honest than one showing three verdicts, and the second does not exist yet. 4. **Selectable value systems are a rhetorical risk.** "Here are three rankings, pick the one you like" is the opposite of what this project is for, and the only thing separating this feature from that sentence is the labelling: one recipe is published, the other two are lenses, and each one carries its own costs in its own words on the same page as its ranking. If the site ever presents the three as equals, this ADR should be reopened. 5. **`full-merit` is an ordering this project rejected on the evidence and still rejects.** B lost the résumé axis in 16 of 16 season × surface cells and ranks teams with losses above unbeaten ones. Shipping it as a selectable lens does not relitigate ADR 0005. It makes ADR 0005 checkable: "we measured it and it lost" was a claim a reader had to take on trust, because B was the one candidate the pipeline could not be pointed at.