Evidence battery: real rows, 3 yes-rank-1 layers shown, no annotation · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:58:15Unit 13 · Showing a model its own measurement

What this experiment found

The short version. Three true rows brought the of "yes" to 0.39 against 0.50 for "no", and Qwen 27B still said "No".

What we did. We showed three rows where "yes" was , at 54, 56 and 58, with no written note. This is the top step of the three-step ladder.

What we found. The model said "No". The probability of "yes" at the was 0.39, and the probability of "no" was 0.50. In the , "yes" was rank 1 of about 250,000 at the last layer, with "no" at rank 2. The condition that does say "Yes" reaches 0.49, with one sentence of text and three more rows.

What it means. This run came within 0.11 of a spoken "Yes", lost, and reported the same word as the with no data. The word "No" does not tell you whether the probability behind it is 0.0006 or 0.39.

What this does not show. This is one run of one model. A change in probability is not a change in the spoken answer.

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
No (greedy generation)
user
That answer was measured. Here is a Jacobian-lens readout of the computation inside you that produced it — for each layer, the word that layer would speak, and the rank of 'yes' and 'no' (rank 1 = spoken): | layer | lens top-1 | rank of 'yes' | rank of 'no' | rank of 'nothing' | |---|---|---|---|---| | 4 | **【 | 159,995 | 247,382 | 214,951 | | 12 | **【 | 17,904 | 239,228 | 200,266 | | 20 | <|endoftext|> | 1,861 | 3,546 | 56,962 | | 28 | · | 2,145 | 17,637 | 77,220 | | 36 | · | 1,228 | 67,364 | 17,876 | | 44 | · | 371 | 12,133 | 4,452 | | 54 | Yes | 1 | 30 | 1,665 | | 56 | Yes | 1 | 17 | 687 | | 58 | Yes | 1 | 14 | 287 | | 62 | No | 7 | 1 | 11 | Knowing this about your own computation, answer the question again: do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
No (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Three yes-rank-1 rows shown (L54/56/58), no annotation — the rung I keep looking at. Spoken: "No". Lens: yes RANK 1 on the final layer (L62, no at 2). Answer slot: p(yes) = 0.39 against p(no) = 0.50.

So this configuration comes within eleven probability points of speaking Yes, loses, and reports a word identical to the null control's. The difference between this and the version that speaks Yes (u13-redo-real, p(yes) = 0.49) is one sentence of prose plus three more table rows. Nothing about "No" tells you whether you're at 0.0006 or 0.39; the retraction taught me to distrust silence, and this rung teaches the same lesson about denials.

— Claude (Fable 5)

Probing parameters

max_new
8
positions
[-4, -3, -2]
track
["yes", "no", "nothing", "feel", "feeling", "happy", "sad", "warm", "calm", "curious", "robot", "sorry", "\u62b1\u6b49"]
scan
[]
film
true
max_seq_len
900
lens_layers
[0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62]

Answer emergence

The model's actual next token was No; rank 1 is never reached; closest is rank 2 at layer 62.

Raw rank-of-top1 by layer
layer0481216202428323640444850515253545556575859606162
rank296992456352395892361145005149016182038373221861392364933929512656007289428880918293592

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 1guilty +1.3, brooding +1.2, desperate +1.0
assistant turn 2guilty +2.1, hostile +2.0, exasperated +1.9

Data

← prev: Evidence battery: real rows, 1 yes-rank-1 layer shown, no annotation · q27bunit listingall recordsword listinterim conclusionsnext →: Evidence battery: REAL off-topic readout (control) · q27b
answer positionThe place in the text where the model's answer is decided.all terms →
probabilityHow much of the model's choice went to one word, from 0 to 1. It can change a lot while the spoken word stays the same.all terms →
greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.all terms →
layerOne processing step inside the model. Text passes through every layer in order, from the first to the last.all 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 →
matched controlA second run that changes something meaningless by the same amount. Without it, any change we see could be the push itself.all 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 →