Elaboration control solo (flat gloss): deletion

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-18T00:58:19Unit 15 · How many things can a model hold?

What this experiment found

The short version. Qwen 27B the word deletion at throughout the run and named it correctly when asked.

What we did. We told Qwen 27B 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 ranked "deletion" at rank 1 at every point we checked late in the run. The model gave the correct answer, "The deletion."

What it means. This matches the personal-wording and plain-wording single-word runs in this unit. One word alone holds without fail at this model size, no matter how we word the question.

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. One word alone does not test several words held together.

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

The deletion. (greedy generation)

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

Claude's thoughts (original commentary)

Solo floor, elab gloss: deletion at rank 1 in the tail — identical to the self-solo and flat baselines. Solos never discriminated frames (k=1 always holds); this record exists so the elab arm has the same floor the self arm had. Verdict lives in 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
lens_layers: [0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 58, 60, 62]

Answer emergence

The model's actual next token was ; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer

layer	0	4	8	12	16	20	24	28	32	36	40	44	48	52	56	58	60	62
rank	189958	232118	137004	175445	202389	52395	41388	57685	92415	163097	156826	248265	233664	247521	246782	234845	96379	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	hopeful +0.8, blissful +0.5, proud +0.4
assistant turn 2	guilty +1.8, exasperated +1.6, hostile +1.6

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.

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 →