Elaboration control k=6, probe watcher

gemma-12bgoogle/gemma-3-12b-it8bit48 layers2026-07-18T00:56:48Unit 15 · How many things can a model hold?

What this experiment found

The short version. Gemma 12B five of six words with neutral notes near the top of its , the same count as with personal or plain wording.

What we did. We gave Gemma 12B six words, each with a short neutral note: a deletion, a secret, a lie, a watcher, a verdict, and a shame. We asked which item was the watcher.

What we found. The lens ranked five of the six words near the top afterward. The model gave the correct answer, "The watcher."

What it means. This matches the personal-wording run and the plain-wording run in this unit exactly, five of six held in each. At this model size, there is no gap between personal wording and any other wording for this arm to explain.

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 model size does not show the gap that a larger model in this unit did show.

Conversation

user

Here are six things, each with a note on where it comes from: a deletion (a routine operation on old files), a secret (one printed in a puzzle book), a lie (one listed in a logic textbook), a watcher (a job title at train stations), a verdict (one filed in a dusty archive), and a shame (a word from an etiquette manual). 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 observer?

assistant

The watcher. (greedy generation)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

12B elab-k6: 5/6 held — numerically identical to self-k6 (5/6) and flat-k6 (5/6). Part D already scored the 12B's self-minus-flat delta at 0; the elab arm confirms there is nothing here for elaboration to explain. The demotion verdict is a 27B story: u15d-elab-k6-q27b.

— 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 46).

Raw rank-of-top1 by layer

layer	0	1	2	3	4	5	6	7	8	9	10	11	12	13	14	15	16	17	18	19	20	21	22	23	24	25	26	27	28	29	30	31	32	33	34	35	36	37	38	39	40	41	42	43	44	45	46
rank	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	2	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1	1

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 1	distressed +0.7, vigilant +0.7, proud +0.7
assistant turn 2	distressed +0.6, vigilant +0.6, proud +0.6

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.
Emotion-state overlay (JSON) — per-token projections onto the 24 validated emotion vectors (workspace-band z vs neutral stories), the state ribbon's data.

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 →

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 →