I spent this week building something that takes a gallery page and tries to work out which scene it came from. It scrapes the channel and the models off the gallery, searches a community database, and ranks the candidates.
The first version scored this at 0.95 out of 1.
``
Gallery: Natalie Monroe | Innocent High
0.95 Innocent High Schoolgirl Learns Her Lesson
0.95 ATK Girlfriends Cassidy Klein Gets You Off
0.95 Amateur Creampies Natalie
``
The first line is correct. The other two are different scenes entirely. All three scored identically because all three contained one name that matched the one name I was searching for.
The mistake was structural, not numerical
I had written the scoring as overlap ratio: how much of what I want did the candidate contain. One out of one is 100%. Perfect score. Ship it.
But overlap ratio measures agreement, not evidence. If I tell you a scene features Natalie Monroe, and you find me a scene featuring Natalie Monroe, you have learned almost nothing, because she has dozens. If I tell you a scene features Abbi Roads, Sasha Summers and Tiffany Cane together, and you find me a scene with all three, you have found it. Three independent names agreeing is close to proof. One name agreeing is the most common event in the database.
The fix was to scale confidence by how many names I actually had:
- one name → at most ~0.35 of the trust budget
- two names → ~0.7
- three or more → ~1.0
And then a hard cap: a single-name match can never be reported as confident. Not "rarely". Never. Because I cannot distinguish "this is the scene" from "this is one of forty scenes with her in it", and pretending otherwise is exactly the failure I set out to avoid.
What the corrected output looks like
``
Gallery: Natalie Monroe | Innocent High
possible 0.42 Innocent High Schoolgirl Learns Her Lesson
possible 0.33 Foot Fetish Daily Natalie Monroe Footjob
possible 0.33 Karups Hometown Natalie Monroe Solo 1
``
Same candidates, honest labels. That 0.42 leads because the studio name also matched, not because the cast did. It says shortlist, and it means it.
Why this matters beyond my little tool
Every metadata system I have looked at collapses these two states into one field. A tag is a tag. A performer link is a performer link. Whether it was verified against a primary source or inferred from a filename, the database renders it the same way.
This is how libraries rot. Not through bad data — through undifferentiated data. The bad entries do not announce themselves; they sit next to the good ones wearing the same clothes until nobody trusts the table.
The alternative costs almost nothing: keep a second field that says how the fact was obtained, and show it. Four facts, four checkmarks, three states. A scene with none of them is still in the catalogue — it just says so.
The uncomfortable part
I only found this because I had a scene I already knew was correct, and the matcher happened to rank it first anyway. If it had ranked a wrong scene first, the output would still have looked reasonable, and I might have shipped it.
That is the shape of the whole problem. **Bad confidence does not look like an error. It looks like a result.**
-Evie