Method

How the Feel Layer works

A public version of the instrument’s rules. Every score published anywhere on a Feel Layer register follows this sequence.

1. 01

   Dimensions come from the corpus, not a template

   Every domain gets its own read. Instead of forcing hotel reviews, mattress reviews, and game reviews through the same fixed rubric, the Feel Layer discovers what people actually talk about in each corpus (the recurring, specific things reviewers mention unprompted) and builds the scoring dimensions from that language. A rubric imposed from outside would miss what a domain's own customers care about.

2. 02

   No score without enough evidence

   A dimension is not scored until it clears a minimum-mentions gate. Below the gate, the entry shows as pending evidence, not a number. This is the difference between a real signal and a guess dressed up as one: thin data stays visibly thin.

3. 03

   Calibrated two ways before anything ships

   Split-half reliability checks whether the same corpus, cut in half at random, produces the same score from each half; if it doesn't, the dimension isn't stable enough to publish. Separately, a frontier-referee bakeoff compares the production scorer against a stronger reference model on a sample, and the gap sets a calibration constant. Both checks run before an edition is trusted, and again on a schedule after.

4. 04

   Betrayal lowers a score. Mismatch never does

   Two kinds of negative review get treated differently, on purpose. Betrayal is a broken promise (the thing didn't do what it claimed) and it lowers the score. Mismatch is a wrong-buyer complaint, such as "great but too firm for a side sleeper like me," and it never lowers a score. Instead it becomes routing data: evidence for who a product is and isn't for, surfaced separately so a good match for one buyer isn't punished for being a bad match for another.

5. 05

   Editions are frozen the day they publish

   Each scored edition is dated and never edited or removed after it goes live: not for money, not on request, not when the underlying corpus changes. If a corpus is withdrawn, past editions stand as published and the entry visibly ages out by date instead of quietly disappearing or being touched up.

6. 06

   An independent adversary checks the scorer

   A separate reviewer, structurally apart from the team that builds the scorer, re-checks a frozen reference set on a schedule, watches for score inflation across editions, and flags drift before it reaches production. Its job is only to find mistakes, never to fix or improve the scorer itself. That separation is what keeps the check honest.