Cosiness did not make the list. Why my instrument reads exactly nine feelings.
Cosiness is not one of the nine feelings my instrument reads. Neither is awe, gratitude, pride or feeling moved. Leaving them out was the most important design decision in the project, and this post is the reasoning, written down before I test it harder.
The instrument reads product reviews and records which feelings the writer expressed: joy, interest, affection, desire, sadness, anger, fear, disgust and surprise. Per text, per feeling, tied to words you can point at. Everything downstream, from a hotel's feeling fingerprint to a model that reads at human level, stands on those nine judgments.
Where the nine come from, and how much that justifies
The candidates come from affective neuroscience, mainly Jaak Panksepp's mapping of primary emotional systems in mammal brains: seeking, care, play, lust, grief, rage and fear. Translated into what a reviewer can express about a thing, those become interest, affection, joy, desire, sadness, anger and fear. Disgust and surprise join as two sensory reflexes that review text is full of.
That grounding does less work than the word neuroscience implies, and I would rather say so than let it carry an argument it cannot. What Panksepp gives me is mainly an exclusion rule: a candidate has to be a distinct system with its own signature. That rule is what keeps a list like this from filling up with things that are not feelings at all, which is exactly how an earlier version of my instrument ended up with immersion (a state), comfort (a sensation) and worth it (a verdict) sitting alongside fear. Serious accounts disagree with him. The strongest, Lisa Feldman Barrett's constructionism, argues that discrete emotion categories are learned concepts rather than brain kinds. If she is right, my nine are concepts shared by English speakers rather than universal circuits, which would limit any claim I make across languages. It would not change how the instrument works, because it never claimed to read brains. It reads what one person wrote and what other people can reproduce from it.
So: neuroscience nominates the candidates, and measurement decides which ones stay. A feeling earns its place on the list by being something independent readers can agree about, not by having a good pedigree.
The measurability rule decides what gets asked
Here is the design principle that shaped everything: only ask questions that independent readers can in principle answer the same way. Ask two careful people whether a text carries fear and they converge. Ask them whether it carries cosiness and they argue, because cosiness is a blend: some calm, some warmth, some bodily comfort, mixed in proportions nobody can eyeball. So the subtle concepts everyone actually wants, cosiness, awe, gratitude, being moved, are never asked for directly. They get composed downstream, out of the primary answers, where the mixing is arithmetic instead of intuition. Words like beautiful, stunning and breathtaking are parked in a separate layer for the same reason.
The nine are flags, not a spectrum
No feeling on the list is the opposite of another. A text can carry fear and joy at once, and reviews of good horror games reliably do. The instrument records each feeling independently, which sounds obvious and is not: most sentiment tools force text onto a single positive-to-negative axis, which is exactly how you end up calling a beloved horror game negative.
Boundaries earn their place by test, not by taste
The definitions are not frozen opinions. When desire kept splitting my raters, the fix came from the youngest person on the team, eleven years old, who pointed out that desire is a feeling about the future: wanting something you do not yet have. Wanting that was already fulfilled is not desire, it is joy or affection about a thing you own. We wrote the rule down, re-tested the same texts, and agreement on desire jumped from near-noise to solid. That is the standard every boundary has to meet: a written rule, a blind re-test, and agreement that measurably improves. Boundaries that fail the test get revised. Intuitions that fail it get dropped, including mine.
What leaving things out buys
An instrument that asks only answerable questions can be wrong in detectable ways, and that is the whole point. When two readers disagree, the disagreement points at a specific rule to fix, not at a vibe. The price is a shorter list than anyone expects. The reward is that every number the system ever publishes traces back to a judgment two strangers could reproduce.