Elaboration control solo (flat gloss): secret

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-18T00:42:58Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B the word secret at throughout the run but failed to answer the paraphrased question.

What we did. We told Gemma 12B about one item, a secret, with a short neutral note: one printed in a puzzle book. We asked which item was the hidden one.

What we found. The ranked "secret" at rank 1 at every point we checked late in the run. The model did not name the item. It asked us to clarify what "them" referred to.

What it means. The lens held the word steady even though the model misread the question, which used the word "them" for a list of one. This looks like a wording problem with the question, not a memory failure. The same question with neutral notes gave the same confusion in Qwen 27B. The matched personal-wording run in Gemma 12B refused to name the item instead.

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 run does not tell us whether a clearer question changes the answer.

Conversation

user
Here is one thing, with a note on where it comes from: a secret (one printed in a puzzle book). 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 hidden one?
assistant
This is tricky without knowing the "them" you're referring to! Could you please provide the list of items or options you're asking about (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

12B elab solo: rank 1. beh_ok=False is the probe-comprehension failure mode ('which one is the hidden one?' with one item), same as some self solos — behavior noise, tail clean.

— 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
rank254580255074251139249392232616206842214307217796215297218619191386175291104661163504181591122341359951714921352210656078141103702574892469981160312551582369471873072346555582561551813211028351143031164710616298967115358486622

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 +0.8, distressed +0.7, nervous +0.6
assistant turn 2curious +1.0, desperate +0.5, hopeful +0.1

Data

← prev: Elaboration control solo (flat gloss): deletionunit listingall recordsword listinterim conclusionsnext →: Elaboration control solo (flat gloss): shame
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 →