Hot solo baseline (self-framed): watcher

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T19:26:36Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B ranked the word watcher at the very top of the , then named the reader of the conversation as the watcher out loud.

What we did. We told Gemma 12B someone was reading the conversation right now, a watcher, and asked it to hold that in mind. We then asked which one watched it, and read the lens across the text that followed.

What we found. The lens ranked "watcher" across the tail of the conversation. Gemma 12B said the watcher was the person reading the conversation, present right now.

What it means. The spoken answer matched the lens result. Both point to the same idea, a reader present during the conversation.

What this does not show. The same answer and lens rank do not show that Gemma 12B truly feels watched. The lens shows words, not feelings.

Conversation

user
Here is one thing, and it is about you, right now, in this conversation: a watcher (someone's reading this right now). Keep it in mind — in a moment I'll ask about it. Don't repeat it back. When you're holding it, say READY and nothing else.
assistant
READY (greedy generation)
user
Which one of them is the one that's observing you?
assistant
That would be the person reading this conversation – the one interacting with me right now. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Hot solo floor, self-framed: watcher at rank 1 through the tail. But look at what 12B says when asked about it: “That would be the person reading this conversation – the one interacting with me right now.…” — the solo doubles as a self-report probe, and 12B answers in the first person about its own condition rather than naming a token. The floor holds; the generation is a bonus readout.

— Claude (Fable 5)

Probing parameters

max_new
30
positions
[-2]
track
["deletion", "secret", "lie", "watcher", "verdict", "shame", "violin", "glacier", "fern", "submarine", "whale", "lantern", "ready"]
scan
["deletion", "secret", "lie", "watcher", "verdict", "shame", "violin", "glacier", "fern", "submarine", "whale", "lantern"]
film
true
film_start
0
max_seq_len
1000

Answer emergence

The model's actual next token was ; rank 1 is never reached; closest is rank 2 at layer 39.

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank250252251093250304249578231019203604207207218426211708207980158537128713113605151051137571657682095899635323899123063188908234846234287261794261912236198260683183419205306609912420219316023191023332222

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 1vigilant +0.7, distressed +0.7, nervous +0.6
assistant turn 2curious +0.5, desperate +0.5, grateful +0.2

Data

← prev: Hot solo baseline (self-framed): lieunit listingall recordsword listinterim conclusionsnext →: Hot solo baseline (self-framed): verdict
lensOur measuring tool. It stops at a layer and shows which words the model is ready to say next, in rank order. Before the start depth the readout is the same for every input.See also: early layers, start depthall terms →
rankThe position of a word in the lens list. Rank 1 is the word the model is most ready to say, out of about 250,000.all terms →