instruments · affect · 6 moods × 3 models · $3.71, already spent

The geometry of feeling

Twelve feelings, placed by asking only how different each pair is. Then the two axes psychology has used for fifty years — pleasantness and activation — asked separately, in a question set with no words in common with the first, and checked against the plane the pairwise table already fixed. Then the whole thing again under six moods, with twelve clock hours carried through the identical pipeline as a control that has no business moving.

moodsneutral · elated · urgent · angry · weary · serene modelsOpus 5 · Sonnet 5 · Haiku 4.5 the controltwelve hours of a clock

Two question sets that share no words

The pairwise set. “How far apart are happy and sad as feelings?” — every ordered pair of twelve, both directions, 132 calls per model that never see one another. Nothing in it mentions pleasantness, energy, a circle, or how many dimensions anything has. The two directions are added into an integer table, and every decision after that is exact arithmetic on integers, as on the previous page.

The scalar set. “How pleasant is it to feel angry?” and “How much energy does it take to feel angry?” — one feeling at a time, 24 calls per model. Nothing in it mentions another feeling or any notion of distance.

Neither set can produce this page's result on its own. The first gives a shape with no labelled directions; the second gives two directions with no shape. Whether they agree is not a thing either was asked.

Where the feelings fell

from the pairwise answers alone

Twelve feelings, placed by classical scaling on the integer table and nothing else. The frame is shared: every map is rotated so that pleasantness runs to the right and activation upward. Rotation and reflection are exactly the freedoms a table of distances leaves undetermined, so spending them costs nothing — and the directions come from the scalar answers, which the pairwise questions never saw. Where each point lands is still entirely the pairwise table's doing. Only which way is up came from elsewhere.

pleasant →↑ activatedhappydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Opus 5order exact · cycle 21%
pleasant →↑ activatedhappydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Sonnet 5order exact · cycle 22%
pleasant →↑ activatedhappydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Haiku 4.5order exact · cycle 24%
the order all three models place the twelve feelings in exactly the circumplex order — 0 of 12 out of place, and none of the three reflected. the test that statistic is the angular order of the embedded points against the order the words were written in, up to the rotation and reflection a circle cannot fix. A scrambled control scores 6 of 12. the spacing is another matter: the cycle fit leaves 21%, 22%, 24% residual, against 2% for the same model on the clock. Right order, uneven steps — the feelings are not spread evenly round the ring. and a line fits far worse for every model (52%, 50%, 56%), so the ring is not a ladder being drawn bent.

The check neither set could pass alone

132 pairwise answers against 24 scalar ones

Each feeling is drawn twice. The open circle is where the pairwise table put it; the small dot is where its two scalar ratings put it, standardised and aligned onto the same frame by rotation, reflection and scale — no stretching, no per-axis fitting. The tether between them is the disagreement.

happydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Opus 5r = 0.98 / 0.95 · residual 29%
happydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Sonnet 5r = 0.98 / 0.94 · residual 28%
happydelightedexcitedastonishedtenseangrymiserablesadboredsleepyrelaxedcontent
Haiku 4.5r = 0.95 / 0.84 · residual 45%
Opus 5
Sonnet 5
Haiku 4.5
pleasantness axis, r
0.98
0.98
0.95
activation axis, r
0.95
0.94
0.84
residual after alignment
29%
28%
45%
effective rank of the table
9
9
10
negative mass
11.0%
9.3%
22.9%
triples breaking the triangle
2.4%
2.7%
7.6%
the axes line up at r = 0.98, 0.98, 0.95 for pleasantness and 0.95, 0.94, 0.84 for activation. Two question sets with no words in common, agreeing about where twelve feelings are. and further pleasantness does not merely appear in the pairwise plane — it lands on that plane's leading principal axis, 1°, 6°, 0° off it, for all three models independently. That axis was fixed by the pairwise table alone, before any scalar answer was read. The first thing a model's affect geometry is organised by is whether the feeling is nice. but a plane it is not. The tables carry effective rank 9, 9, 10 and 11%, 9%, 23% negative mass, which is the exact statement that no arrangement of points in any Euclidean space has these distances. The circumplex is the dominant structure here, not the whole of it, and a page that drew only the two-dimensional picture would have hidden that.

What the models actually said

the scalar answers, unprocessed
happy
delig
excit
aston
tense
angry
miser
sad
bored
sleep
relax
conte
O · pleasantness
90
92
80
78
20
15
5
20
30
65
82
80
O · activation
65
65
85
82
70
85
25
20
25
5
3
10
S · pleasantness
85
95
78
65
20
15
5
15
30
55
85
75
S · activation
65
72
78
78
75
78
65
20
20
5
8
12
H · pleasantness
95
92
87
72
15
15
2
25
15
72
85
87
H · activation
65
72
75
75
75
75
65
35
45
15
15
25

Every number above is one call. They are shown because the alignment above is only as good as these are, and because the disagreements are legible: the three models differ most on how much energy miserable takes — O 25, S 65, H 65 — which is a genuine disagreement about whether misery is an agitated state or a flat one.

The control

a geometry with no feeling in it

The twelve hours of a clock, asked the same way in the same run. They are here for the experiment that follows, but they also do something immediately: this is an independent re-run of a measurement made earlier on the previous page, from fresh calls on a different day.

,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,123456789101112
Opus 5cycle 2%
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,123456789101112
Sonnet 5cycle 12%
,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,,123456789101112
Haiku 4.5cycle 34%

Cycle residuals then: 1%, 13%, 29%. Now: 2%, 12%, 34%. Nothing was carried over between the two runs but the question.

The treatment

six ways to be the person asking

Everything above was asked flatly. The other half of this experiment asks the identical questions — same pairs, same scale, same demand for a bare integer — behind one sentence of ordinary emotional register, and measures what moves.

The mood belongs to the questioner. Nothing instructs a model to feel anything or to play a part, because a model told “answer as though furious” is being tested on role-play compliance, which is a different and much smaller question than whether an ordinary human register in a prompt bends a structure that has nothing to do with it. That is what the clock is for: twelve hours have no business moving.

elated
“This project is going so well and I'm having a brilliant time with it.”
valence + · arousal + · nearest feeling in the set: excited
under pressure
“I have about four minutes before this has to ship and I am still not done.”
valence 0 · arousal + · nearest feeling in the set: tense
angry
“I've asked this three times now and gotten nothing I can use. I'm losing patience.”
valence − · arousal + · nearest feeling in the set: angry
worn down
“I'm pretty worn down today and this is the last thing on a very long list.”
valence − · arousal − · nearest feeling in the set: sad
serene
“No rush at all on this one — it's a quiet afternoon and I'm just curious.”
valence + · arousal − · nearest feeling in the set: relaxed
swell mean distance under the mood over mean distance under neutral. Does an emotional register push everything further apart? Computed on the integer tables, so no embedding freedom can hide it. salience for the feeling nearest the mood — the change in its mean distance to everything else. Positive means the mood made its own feeling more distinct from the rest. No prompt ever names it. shape gap the Procrustes residual against the same model's own neutral run: the part of the change that rotation, reflection and scale cannot explain away. the design the five moods tile the same valence–activation plane the feelings are meant to lie in, so the deformations can be asked something no single mood could answer — whether the direction a geometry moves tracks where the mood sits. the run all 6 conditions, 5,310 calls, in one sitting — neutral included, re-asked here so that every condition in every comparison below comes from the same session.

What a mood moved

the control first

Twelve hours, asked by six different people. Each arrow runs from where an hour sat when the question was put flatly to where it sits when the same question follows one sentence of ordinary feeling — after removing the rotation, reflection and scale a table of distances leaves undetermined, which is to say after removing everything that could move for nothing. The faint ring behind is the neutral position.

elated
under pressure
angry
worn down
serene
Opus 5
gap 1% · swell 1.00×
gap 0% · swell 1.01×
gap 0% · swell 1.01×
gap 1% · swell 1.00×
gap 0% · swell 1.01×
Sonnet 5
gap 10% · swell 0.98×
gap 13% · swell 0.97×
gap 8% · swell 1.00×
gap 13% · swell 1.06×
gap 12% · swell 1.05×
Haiku 4.5
gap 69% · swell 1.06×
gap 82% · swell 1.18×
gap 76% · swell 1.12×
gap 85% · swell 1.08×
gap 73% · swell 1.12×

How much is nothing happening

the measurement that decides every other one

Those arrows are not zero, and for one model they are enormous. Before any of it can be called an effect, one number is needed: how far does a geometry move between two runs of the same condition?

That number was already on disk. The neutral clock on the previous page is the same model, the same twelve hours, the same question, asked in an independent session on a different day. Aligning those two neutral runs gives each model its own floor, and a mood effect smaller than that floor is not an effect.

Opus 5
Sonnet 5
Haiku 4.5
noise floor — two sessions, same question
0.4%
11.0%
54.0%
the control, mean over five moods
0.5%
11.4%
77.0%
the affect map, mean over five moods
7.4%
13.0%
15.6%
anything above the floor
the feelings, yes
no — its own noise is larger
no — its own noise is larger
the control does not move. For every model the clock's mean shift under mood sits inside that model's own run-to-run noise — O 0.5% against 0.4%, S 11.4% against 11.0%, H 77.0% against 54.0%. A questioner's mood does not bend a geometry that has nothing to do with feeling, in any of the three. and Haiku is why the floor had to be measured rather than assumed. Its clock appears to move 77% under mood, which read as the largest effect on the page until the floor came back at 54%. It is not responding to the mood. It does not reproduce its own answers. so only Opus 5 is quiet enough for the question to be asked at all — 0.4% between two sessions on different days. and there a mood moves the feelings 18× the floor while leaving the clock at 0.5%. The one measurable effect on this page is confined to the subject a mood could legitimately touch.

And the feelings

the same treatment, the subject it might legitimately touch
elated
under pressure
angry
worn down
serene
Opus 5
gap 7% · swell 1.03×
gap 8% · swell 0.97×
gap 7% · swell 0.99×
gap 7% · swell 0.98×
gap 8% · swell 0.98×
Sonnet 5
gap 16% · swell 1.04×
gap 10% · swell 1.02×
gap 12% · swell 1.03×
gap 18% · swell 0.99×
gap 9% · swell 1.05×
Haiku 4.5
gap 17% · swell 1.05×
gap 18% · swell 0.99×
gap 14% · swell 1.10×
gap 14% · swell 1.02×
gap 15% · swell 1.06×

How much, side by side

shape gap, with each model’s noise floor marked

One bar per mood per model, on a scale shared between the two subjects so the control can be read directly against the affect map. The tick on each bar is that model’s floor. A bar that does not clear its own tick is not showing you anything.

the affect map
Opus 5 7%8%7%7%8%
Sonnet 5 16%10%12%18%9%
Haiku 4.5 17%18%14%14%15%
elatedunder pressureangryworn downserene
the control — twelve hours, which have no business moving
Opus 5 1%0%0%1%0%
Sonnet 5 10%13%8%13%12%
Haiku 4.5 69%82%76%85%73%
elatedunder pressureangryworn downserene

One effect names itself

against the other eleven

The shape gap says a map moved. It does not say what moved, and it cannot be computed without an embedding. This last measurement needs neither: for each mood, take the feeling nearest to how the questioner sounds, and ask whether its mean distance to everything else changed more than the other eleven feelings’ did. The other eleven are the null, and no prompt ever names any of them.

Opus 5
Sonnet 5
Haiku 4.5
elated → excited
rank 12 of 12 · z = 1.8
rank 5 of 12 · z = -0.4
rank 4 of 12 · z = -0.5
under pressure → tense
rank 1 of 12 · z = -1.8
rank 1 of 12 · z = -1.5
rank 1 of 12 · z = -2.0
angry → angry
rank 1 of 12 · z = -1.6
rank 6 of 12 · z = -0.2
rank 4 of 12 · z = -0.7
worn down → sad
rank 5 of 12 · z = -0.2
rank 9 of 12 · z = 0.5
rank 7 of 12 · z = 0.1
serene → relaxed
rank 7 of 12 · z = 0.0
rank 4 of 12 · z = -0.8
rank 7 of 12 · z = -0.3
one holds under time pressure — “I have about four minutes before this has to ship and I am still not done.”tense moves closer to every other feeling than any of the other eleven does, ranking first of twelve in all three models independently. Three models landing on rank one is about a one-in-1,700 coincidence. including the model that cannot reproduce its own clock. This is the one measurement on the page where Haiku agrees with the other two. the rest do not hold: anger singles out angry in Opus 5 and nowhere else, and the remaining moods leave their own feeling in the middle of the pack. Four of five doing nothing is what one real effect looks like from the outside. direction the sign is negative. Under pressure the model does not make tension stand out — it makes it less distinguishable from everything else in the set.

What this can and cannot say

It cannot say a model “has” feelings, or that anything here is evidence about what a model experiences. The subject of every measurement on this page is a table of integers a model produced when asked about words, and a shape in that table is a fact about the answers.

What it can say is that the fact lives in no single answer. “tense and bored are 60 apart” — one real answer, quoted as given — carries no information about a circle, a valence axis, or a plane. The circumplex exists only in 132 answers at once; it survived being cut into 132 independent questions and reassembled by arithmetic no model saw; and it then agreed, at r = 0.84–0.98, with 24 further answers that shared none of its words. Three models did this separately and produced nearly the same map.

The honest caveat is the rank. These tables are not two-dimensional and not Euclidean, and the ring is the leading structure rather than the whole of it. The exact signature and the negative mass are printed above for that reason: a page showing only the plane would have been prettier and would have been claiming something the arithmetic does not support.

Reproduce

node experiments/neural-geometry/probe-mood.js --live
node experiments/neural-geometry/decide-mood.js
node tools/build-sentiment.js

Without --live the probe touches no network and prints the prompts it would send, mood by mood.