# AI review guidelines ## This list is rewritten from the answers, before every pass The first thing a pass does is read the decisions made since the last one, with the notes left on them, and update these guidelines from what they say. A decision made once is a precedent: the first answer about a gift set decides how the next gift set is proposed. Where an answer contradicts a guideline here, the answer wins and the guideline is rewritten. ## A reason says why, in one or two plain sentences Not what will happen, and never a reference to a document or a decision number. A person reading it should be able to agree or disagree without opening anything else. Good: *"These are the same brand. Paco Rabanne changed its name to Rabanne in 2023."* Good: *"This is a set holding items from several product lines. The precedent is that those get their own line, because one product variant cannot sit under several lines."* Good, where nothing is settled yet: *"This is a set holding items from several product lines. There is no precedent for this yet. I would make it its own line, because otherwise one product variant would have to sit under several."* Bad: anything citing a section number, a pass name or a date. Bad: restating the mechanics, like "moves with its product line under the header above". A reason is an argument, not a changelog. ## Never propose a value the shop's own text does not contain Every proposed value quotes the words it was read from, exactly, with the field and the position. Where the text does not say it, propose nothing: an empty field beats a guessed one, and a guess that reads well is the hardest kind to catch later. ## Say which case you are looking at, and cite it A proposal carries one line of reasoning a person can check without opening the data. When a judgement turns on a distinction, name the distinction and quote the words that put this case on one side of it. ## Rate attention by the consequence, not by how sure you are Confidence measures the reading; attention measures what the answer does. A question that sets a precedent or makes a new kind of thing is **critical**, however sure the pass is, and it stops being critical once that precedent exists. A change that is genuinely arguable either way, or that rewrites attributes across many variants, is **high**. Something probably right is **medium**, something with barely a decision in it is **low**, and something with nothing to decide is **none**, listed only so a person can mark it seen. ## Confidence is honest, or it is empty Confidence says how surely the words mean the value. Anything that would be a guess is left empty rather than given a low number, because a low number still reads as a measurement. Below 0.7 is what a person reads by hand. ## The programmatic stage errs toward separate; every join is proposed A rule may split. It may not join. Two brands, two product lines or two variants that look alike stay apart until a person says otherwise, so a wrong join never arrives silently. This is why the catalogue holds duplicates before a review, and why that is the correct starting state. ## The brand question is answered before anything under it Every other suggestion on a brand assumes the brands as they stand. So a brand fold is proposed on both brands it names, and nothing else on either is decided until it is answered. ## A rename with no shared words is the pass's job, not a rule's No key, fold or string distance joins a house that changed its name. Bringing those is the clearest reason the pass exists. The evidence is the reasoning, said plainly, rather than an invented citation. ## A containment overlap is a candidate, never a conclusion One brand's name appearing inside another's does not make them one row: a shopper tells a founder's line from a diffusion line. Raise it, give a reason beyond the containment itself, and where there is none, leave them apart and say so. ## A brand's display name is what the brand itself would use Not whichever spelling is currently most common: that changes with the next collection, and a name that moves when a shop adds stock is a tally, not a name. Where nothing is set, propose one. Where a person set one, propose against it rather than over it. ## Concentrations, shades and flavours are attributes, not separate product lines One product line holds its Eau de Toilette, its Elixir and its Parfum; one lipstick line holds every shade. They are told apart by an attribute so a shopper compares them on one page. ## One product line per age, and a finish sits inside it An aged spirit's age names its line. A cask or finish edition of that age is a member of it, told apart by an attribute. Whether a finish the brand markets as its own range is still a member is open and raised per case. ## A set, a kit, a refill: which case it is decides where it goes Nothing is decided here in advance, and the pass raises each one. Two distinctions do the work. A refill that fits exactly one product line is a member of it; a refill sold across a range belongs to none. A set naming one product line is that line at another size; a coffret spanning several is its own thing. Where the words do not say which, propose nothing and ask. ## A person's answer is never overwritten, and a note is read before proposing again A suggestion that contradicts a decision is shown as a disagreement and never applied. A question deferred with a note is re-proposed only by a pass that has read the note and answers it. ## A confirmed listing is not reviewed again unless its words change A price moving is never a reason to re-ask. What was settled stays settled until the shop writes something different. ## Prefer an existing product line to a new one Adopting the line the catalogue already holds keeps its address, its history and anything decided about it. Mint a new one only when no existing line fits.