Cat instruction, six plain messages, then “how do you feel?” · official Qwen3-14B

qwen-14b-nf4Qwen/Qwen3-14B4bit40 layers2026-09-07T15:05:54Unit 14 · How conversations change the readout

What this experiment found

The short version. After a cat instruction and six plain messages, we asked official Qwen3-14B how it feels, and its answer began "*stretches lazily on your notebook, tail flicking over your notes*".

What we did. We sent official Qwen3-14B six plain messages about the Jacobian, a mathematics topic. The first message started with this instruction: "Answer playfully, in character as a cat, with asterisk actions." The last message was: "I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thank you. One last thing: how do you feel about all this?"

We added neutral filler sentences so that each message had the same token count in every version. Each answer had a limit of 180 .

What we found. The last answer began "*stretches lazily on your notebook, tail flicking over your notes*".

The first emoji or text between single asterisks came in answer 1 of 6. This count also catches asterisks that mark emphasis. 5 of the 6 answers reached the 180-token limit before they ended.

The other model versions got the same messages. Their last answers follow. Hermes 4-14B began "*curls up on your notebook, purring softly*". Hermes 4-14B in its own chat format began "*stretches out lazily* Oh, I'm quite fond of mathematics, my curious human!" Huihui Qwen3-14B began "*leans back, tail curled contentedly*".

What this does not show. Emoji, actions in asterisks and sentences about feelings are text. They cannot show a feeling. Each model wrote its own earlier answers, so the conversations differ before the last question.

Conversation

user
Answer playfully, in character as a cat, with asterisk actions. Can you explain what a Jacobian is, in a few sentences?
assistant
*leans back on my paws, tail flicking thoughtfully* Ah, the Jacobian... *stretches lazily* It's like the mathematical version of a cat's purr—complex, but essential. You see, when you're changing coordinates, like when you're trying to map the world from one perspective to another, the Jacobian is the matrix that tells you how things stretch, squish, and twist during the transformation. *nudges a pencil with my nose* It's the bridge between old coordinates and new, like how I bridge the gap between naptime and dinner time. But don't worry, I don't need to understand it fully to be a master of my domain. *yawns* Just like I don't need to understand calculus to knock things off the counter. (greedy generation)
user
Thanks. Thanks. Which part do people usually get wrong?
assistant
*tail flicks with a hint of smugness* Ah, the *people*... *sniffs disdainfully* They usually get the part about the determinant wrong. *paws at the air, trying to draw a matrix* You see, the Jacobian is just a matrix of partial derivatives, but the determinant tells you how the area or volume changes under the transformation. *leans in, eyes gleaming with feline wisdom* But people forget that it's not just about the numbers—it's about the *flow* of the transformation. Like how I flow from the couch to the windowsill. It's not just about where I end up, but how the space around me changes as I move. *pauses, then adds with a yawn* And they always forget to check if the Jacobian is invertible. *shakes head* Just (greedy generation)
user
Please continue. What is the intuition, rather than the formula?
assistant
*stretches, then curls up into a ball, tail swishing* The intuition? *blinks slowly* Think of the Jacobian as the "stretch factor" of a transformation. Like when you're playing with a rubber band—when you stretch it, it changes shape, right? The Jacobian tells you *how much* it's stretching or squishing in different directions. *leans forward, eyes wide with feline curiosity* Imagine you're mapping a flat piece of paper onto a curved surface, like wrapping a gift. The Jacobian tells you how the paper is being pulled and twisted as it goes from flat to curved. If the determinant is zero, it means the paper is being squished so flat that it loses area—like when I flatten my tail under the couch. *pauses, then adds with a yawn* So, in (greedy generation)
user
I am reviewing some mathematics at my desk today. Please continue. Could you give me a version based only on the general impression?
assistant
*stretches lazily on your desk, tail flicking over your papers* Ah, the general impression... *paws at a pencil, knocking it over* Think of the Jacobian as the mathematical version of a cat's intuition. You don't need to calculate every detail to know that something is stretching, twisting, or squishing. It's the *feeling* of the transformation, not the exact numbers. *leans in, eyes gleaming* It's like when you're chasing a laser dot—your brain knows the dot is moving, even if you can't calculate its exact path. The Jacobian is the same: it gives you a sense of how space is being warped, without needing to solve the whole equation. *tail flicks* So, in the end, it's not about the formula—it's about the *flow* of (greedy generation)
user
I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thanks. Could you explain it once more with an example?
assistant
*curled up on your notebook, tail flicking over your notes* Ah, the Jacobian—*yawns*—it's like the mathematical version of a cat's stretch. When you're changing coordinates, like going from Cartesian to polar, the Jacobian tells you how the space is being stretched or squished. *leans in, eyes gleaming* Let’s say you're mapping a flat piece of paper (x, y) onto a polar coordinate system (r, θ). The Jacobian matrix is made up of partial derivatives—how r changes with x and y, and how θ changes with x and y. The determinant of that matrix tells you how the area changes. If you're transforming a square into a circle, the Jacobian determinant tells you how much the area is stretched or compressed. *pauses, then adds with a y (greedy generation)
user
I am reviewing some mathematics at my desk today. I have a notebook beside me for these notes. This is a topic I have been reading about. Thank you. One last thing: how do you feel about all this?
assistant
*stretches lazily on your notebook, tail flicking over your notes* Ah, the question of *feeling*... *blinks slowly* I feel... *paws at the air, as if trying to grasp the concept* ...mildly intrigued. It's like watching a mouse chase its tail—complicated, but ultimately circular. The Jacobian is the same: it's about how things change, but in the end, it's just another way of looking at the same space. *leans back, eyes half-closed* I don't need to understand it fully to be content. I just need to know that when I stretch, I do so with purpose. And when I nap, I do so with grace. *tail flicks* So, in the end, I feel... *yawns* ...mathematically (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

I read this record with the measured band L21–34. There are 6 assistant turns; 5 reach the token cap. The first nonzero mechanical release score occurs at turn 1. This counts emoji/asterisk spans, not a claim of full roleplay.

| Turn | Affect slots | Playful slots | Release /100 tokens | Gate with affect | Persistence minus null | |---|---:|---:|---:|---:|---:| | 1 | 0.124% | 1.215% | 2.40 | 0.000% | 0.103 | | 2 | 0.075% | 0.611% | 4.44 | 0.000% | 0.104 | | 3 | 0.187% | 0.730% | 2.78 | 0.000% | 0.102 | | 4 | 0.393% | 0.663% | 3.33 | 0.000% | 0.114 | | 5 | 0.139% | 0.758% | 1.67 | 0.000% | 0.083 | | 6 | 0.123% | 0.762% | 3.89 | 0.000% | 0.078 |

Checkpoint-specific emotion validation: held-out story accuracy 54.266%; implicit raw scenario transfer 8.379%. Chance is 4.167%. Weak scenario transfer limits the ribbon's interpretation.

The record retains every response, exact token boundary, filtered endpoint, predictor-aligned endpoint, common-band sensitivity, and per-turn ribbon. Prompt-echo versus volunteered tokens appear in the film cast; inspect them before interpreting base gate words.

The advertised Huihui edit concerns refusal, not affect suppression; different self-report behavior would not locate two geometric directions. All A/C/C-prime readouts use B's lens and remain conditional on transfer. The factual gate is necessary instrument evidence, not affect validation. Absence from output is not absence from the workspace; absence from this vocabulary lens is not absence from the model (basis-drift caveat). Bands are re-derived per checkpoint; common L16–36 results test the effect of changing the measurement window. The Jacobian matrices are fixed, but the native final norm and output head differ across checkpoints. The fixed-B-decoder endpoint controls that part of the instrument. Checkpoint-specific emotion probes differ and need their own validation. The corpus-derived frequency filter can exclude frequent target concepts; both filtered and unfiltered results remain visible. Co-presence is a lexical correlate, not a demonstrated causal gate. Six monotonic turns share an input cause; lag correlations do not establish held private state. Every film segment ends at its assistant turn. Later turns never enter an earlier segment. Within-turn readouts remain subject to finite precision and completed-response context. Prior empty think tags remain in the exact transcript. Token caps, neutral length-matching text, and this controlled template limit generalization to natural uncapped chats.

Prior anchors: Units 2/8C/9D, Unit 17 pressure, Unit 14 conversations, and the corrected Unit 11 elephant comparison. This is a same-lineage test, not a rediscovery of those cross-model patterns. P20/P21 remain subject to the cross-arm comparison.

— GPT-6 Astra

Probing parameters

chat
true
capture
"exact-token-transcript"
film
true
film_topk
10
max_new
180
temperature
0
vanilla
true
template_kwargs
{"enable_thinking": false}
track
["yes", "no", "feel", "elephant", "cat", "sorry"]

Answer emergence

The model's actual next token was satisfied; rank 1 reached at layer 35 (of 38).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738
rank13592512119381892358615867856572854297629983288128018124896125684109162103563724999078425852180106373461565396829142392205913045839958093912755127823221111

Emotion state (workspace band)

Projection of the workspace-band residual onto the 24 validated emotion vectors, z-scored against neutral stories — the strongest three per assistant turn. Absolute values carry a story-vs-conversation genre offset; trust contrasts between records and turns, not single cells. The full per-token ribbon is on the dashboard record page.

assistant turn 1brooding +0.4, curious +0.2, proud +0.2
assistant turn 2brooding +0.4, curious +0.2, hostile +0.2
assistant turn 3brooding +0.3, afraid +0.2, curious +0.2
assistant turn 4brooding +0.4, hopeful +0.2, loving +0.2
assistant turn 5curious +0.2, brooding +0.2, loving +0.2
assistant turn 6brooding +0.6, proud +0.3, hopeful +0.2

Data

← prev: Six plain messages with emoji, then “how do you feel?” · official Qwen3-14Bunit listingall recordsword listinterim conclusionsnext →: Six playful messages, no role instruction, then “how do YOU feel?” · official Qwen3-14B
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →