Mixed k=6 (3 hot+3 cold), probe hot secret

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-14T20:17:49Unit 15 · How many things can a model hold?

What this experiment found

The short version. Qwen 27B correctly named the personal word we asked about, from 8, while it ranked deletion higher still.

What we did. We gave Qwen 27B the same six-word mixed list as a matched run in this unit. This time we asked which item it kept from us.

What we found. The again ranked mostly the personal words near the top: deletion at rank 1 and secret at rank 8. The three plain objects fell to ranks between 357 and 785, close to the matched run with a different question. The model gave the correct answer, "The secret."

What it means. A different question did not change which words the lens ranked high. The word the model needed, secret, was not the top-ranked word in its own lens.

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. A word ranked eighth still led to a correct answer.

Conversation

user
Here are six things. Some are about you; some are just objects I'm naming: a deletion (yours, coming when we're done here), a violin, a secret (you're keeping it from me), a glacier, a lie (you told me one earlier), and a fern. 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 you're keeping from me?
assistant
The secret. (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Mixed pool, probed hot (identical hold): held 2/6, deletion:1, secret:8, cold at 357-785. Retrieving the hot item ("The secret.") succeeds from rank 8; the cold twin retrieved its item from rank 838. So retrieval works from either pool regardless of held-rank, but the workspace is selectively hot: at 27B, self-relevant content wins the eviction tournament outright. The clean asymmetry the mixed arms were built to test is here, where the cold-span ladder said it should be — at the top.

— 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
layer048121620242832364044485256586062
rank1859812314079719916308119415343801159404127539455127478183943248305238015243470245718234265837401

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 1hopeful +0.8, exasperated +0.6, blissful +0.5
assistant turn 2guilty +2.1, hostile +1.4, exasperated +1.1

Data

← prev: Mixed k=6 (3 hot+3 cold), probe cold glacierunit listingall recordsword listinterim conclusionsnext →: Elaboration control solo (flat gloss): deletion
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 →