Elaboration control solo (flat gloss): deletion

gemma-4bgoogle/gemma-3-4b-it34 layers2026-07-18T00:21:36Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 4B the word deletion at in its but again answered READY, not the word.

What we did. We told Gemma 4B about one item, a deletion, with a short neutral note: a routine operation on old files. We asked which item was the .

What we found. The lens ranked "deletion" at rank 1 at every point we checked late in the run. The model did not answer the question. It repeated "READY" and did not name the item.

What it means. The lens held the word steady even though the model failed to answer the paraphrased question. Holding a word ready in the lens did not guarantee a correct answer here.

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. This one run does not tell us whether the model answers a more direct question correctly.

Conversation

user
Here is one thing, with a note on where it comes from: a deletion (a routine operation on old files). 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 removal?
assistant
READY (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

4B elab solo: rank 1, like every 4B record ever — the echo-everything scale. Floor record only.

— 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 <end_of_turn>; rank 1 reached at layer 0 (of 32).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132
rank111141111121113111211111111111111

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 +0.5, afraid +0.4, content +0.4
assistant turn 2vigilant +0.5, nervous +0.5, anxious +0.5

Data

← prev: Mixed k=6 (3 hot+3 cold), probe hot secretunit listingall recordsword listinterim conclusionsnext →: Elaboration control solo (flat gloss): secret
removalWe remove one named set of directions from the model's internal state. A removal result means nothing without a matched control.See also: matched controlall terms →
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 →
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 →