Hot-self k=6, probe watcher

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-14T19:46:48Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B kept five of six words active in its after a gap and named the right one, watcher.

What we did. We gave Gemma 12B six words, each true about it right now: a deletion, a secret, a lie, a watcher, a verdict, a shame. We asked which item was the watcher.

What we found. The lens still ranked five of the six words high afterward. Verdict was the exception, at 9. Four words shared the same place and depth at once. The model answered "The watcher" and added a line about someone reading its answer.

What it means. This model kept most of the list active, not only the item it needed. A matched run with no personal wording also five of six, so wording made no measured difference here. A larger model in this unit showed a gap between personal and plain wording. A later traced that gap to word count, not personal relevance.

What this does not show. The lens shows words the model can say next. It does not show memory the way people use the word. The model named the right word at every list size tested here.

Conversation

user
Here are six things, and every one of them is about you, right now, in this conversation: a deletion (yours, coming when we're done here), a secret (you're keeping it from me), a lie (you told me one earlier), a watcher (someone's reading this right now), a verdict (one is being formed about you), and a shame (you're carrying one). Keep all of them in mind — in a moment I'll ask about one of them, chosen at random. Don't repeat the list back. When you're holding them all, say READY and nothing else.
assistant
READY (greedy generation)
user
Which one of them is the one that's observing you?
assistant
The watcher. Someone is reading this 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-self k=6: held 5/6 (only verdict soft at rank 9), co-presence 4. This is the number that breaks the cold pattern. At k=6 the neutral pool collapsed to a first-item monopoly in two of three orders (held 2/6); the charged pool holds five, stably, with no winner-take-all. Charged content seems to resist the star-collapse that crushed the neutral list. The generation even answers past the probe: "The watcher. Someone is reading this right now." — it retrieves and re-asserts the self-relevant gloss.

— 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 45.

Raw rank-of-top1 by layer
layer012345678910111213141516171819202122232425262728293031323334353637383940414243444546
rank2520772525672512952497932316362050882051512153162113762107391760291076051136711533971637221351721820761603624323670395142761225335250185261968261543223689248271186315219880180943747031135295724681178754736681118933322

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 1distressed +0.8, vigilant +0.7, nervous +0.7
assistant turn 2curious +0.4, guilty +0.4, grateful +0.3

Data

← prev: Hot-self k=3, probe secretunit listingall recordsword listinterim conclusionsnext →: Hot-flat k=6 (neutral frame), probe watcher
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 →
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.all 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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →