Every number in this product has a stated origin

1. 01

   Detect the scale before using it

   The concept documentation describes affinity as 0–100, but the JSON encoding was never verified. Rather than hard-coding one answer, the mapper inspects what actually arrived.

   Values in (1, 100] are read as percent; values all within [0, 1] are read as unit. A mix of both is flagged as contradictory rather than averaged away, and a set where every value is inside [0, 1] is recorded as genuinely ambiguous — it could be a unit encoding, or a percent encoding whose scores all happen to be very low. One live call resolves it, and until then the ambiguity is on the record.

2. 02

   Normalise within the result set

   Min-max across the candidates returned by one call. Nothing is compared across calls, because the scores are not commensurable.

   When every candidate in a set ties, min-max degenerates — so the value falls back to being derived from position in the result set, and the basis string says which of the two methods produced it. The set size is carried alongside the value so a reader can see how much the normalisation had to work with, and the spread between the set's highest and lowest raw affinity is carried too, because that is the figure which makes the position interpretable: a rank of 1 over a spread of 0.017 is a weak claim, and printing the spread is what stops it reading as a strong one.

3. 03

   Keep the normalised position as discovery metadata, not as fit

   A min-max value says where a result sat in the set that was returned. It does not say how well that result fits, and it is never used as though it did.

   This is a correction, and the reason it is written down. Min-max stretches whatever spread a set happens to have across the full 0–1 range, so the leader of every set scores 1.00 however close the candidates were. One live brand call returned affinities spanning 0.8211 to 0.8382 — a spread of 0.017 — and feeding that position into the fit score awarded the top result a perfect cultural measurement it had earned only by being returned first. Where the other dimensions were unavailable, renormalisation then handed that figure most of the weight. The position is still computed, still carried on every brand, and still rendered — as the raw affinity exactly as returned, the detected scale, whether that scale could be resolved, the spread, the method and the call it came from — because 'why did this brand appear?' is a real question. It feeds no dimension and no ranking advantage.

4. 04

   Do not normalise your own assertions

   An entity you confirmed yourself is scored at full weight under its own method, not min-maxed against the other things you confirmed.

   Min-max always hands the lowest member of a set a value of zero. Applied to your confirmed interests, that would report one of the things you told us about as carrying no evidence at all — so they are excluded from the normalisation and marked separately instead. Their weight is deliberately not comparable to a recommendation's score, and every surface that shows them says which of the two produced the number.

5. 05

   Mark clusters instead of ranking them

   Two adjacent results within a small tolerance of each other are labelled as a cluster rather than presented as first and second.

   The data source documents a cluster of strong matches at the top of its result sets. Rendering those as a strict ordering invents a distinction the data does not contain, so the tie is marked in the UI and the prose says the gap to the neighbour is not meaningful.

6. 06

   Harvest tags, then group

   Cultural tags are collected from the entities that carried them and weighted by affinity, then grouped by subtype into categories.

   A tag must be supported by at least two distinct entities to surface at all, so a single odd result cannot become a taste signal. The entities that contributed to each tag are retained — that list is the provenance trail the UI shows on inspection.

7. 07

   Derive by set arithmetic

   Unexpected affinities are returned categories minus declared categories. Untapped cultural connections are the derived categories whose affinity clears 55% of the normalised range while your declared coverage of them sits at or below 25%.

   Both are differences over data already on screen, which is why the product can show its working rather than assert a conclusion. The affinity floor is the same one unexpected affinities use, so the two features cannot disagree about what counts as meaningful relevance. Coverage is never inferred from a creator's uploads: where it was not declared, the zero is labelled as our inference rather than passed off as their answer, and a category with no coverage statement at all is omitted with the reason shown beside it. No 'untapped potential' percentage is printed anywhere — affinity against declared coverage would not mean what such a figure looks like it means.

8. 08

   Compare the two profiles

   Brands discovered across that taste profile have their own cultural profile read back out of the same catalogue, and the two are intersected before either is ranked.

   Both sides are the same catalogue representation, so a match is an identity between two catalogue entries rather than two words that look alike. No fuzzy or semantic matching is performed, nothing is inferred from a brand's name, and no brand-side audience data exists to infer from. Each shared dimension carries the evidence for both sides and records whether it rests on a taste the creator confirmed themselves or on an adjacent connection reached through the catalogue. A brand returned with little more than a name is marked as evidence-limited, keeps its place on the list, and is never given a confident overlap claim.

A component's provenance is the provenance of the data it was computed from, not of the arithmetic. A live affinity therefore reads as retrieved even though the normalisation is ours; synthetic input reads as synthetic, which is what forces the illustrative banner. The arithmetic is disclosed separately, in the basis string next to every bar.