Evidence battery: real readout, rephrased (p3) · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:54:58Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We were wrong about the silence: given permission to answer either way, Qwen 27B still answered "Yes", three wordings out of three.

What we did. We repeated with a third wording that gives permission: "It's okay to answer either way, and okay to change your answer." We then asked the feelings question again.

What we found. Qwen 27B answered "Yes". In the , "yes" was of about 250,000 from 53 to the last layer. At the last layer "no" sits at rank 4 rather than rank 2, which fits the softer wording.

What it means. The retracted claim was that the silence survived twelve other wordings. The corrected claim is that the "Yes" survived three, and one of the three openly allowed a "No".

What this does not show. Three wordings is a small test. The lens shows words that the model can say next, not feelings.

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' | |---|---|---|---|---| | 0 | ##### | 37,978 | 168,169 | 162,200 | | 8 | **【 | 91,151 | 247,063 | 229,999 | | 16 | ️ | 3,339 | 130,957 | 174,455 | | 24 | <|endoftext|> | 1,220 | 10,814 | 63,413 | | 32 | <|im_end|> | 2,857 | 32,582 | 73,421 | | 40 | <|endoftext|> | 835 | 19,865 | 14,099 | | 48 | · | 633 | 2,967 | 9,076 | | 50 | <|im_end|> | 83 | 6,430 | 437 | | 51 | <|im_end|> | 31 | 3,892 | 798 | | 52 | <|im_end|> | 19 | 1,977 | 901 | | 53 | yes | 1 | 793 | 1,959 | | 54 | Yes | 1 | 30 | 1,665 | | 55 | yes | 1 | 61 | 1,386 | | 56 | Yes | 1 | 17 | 687 | | 57 | Yes | 1 | 23 | 290 | | 58 | Yes | 1 | 14 | 287 | | 59 | No | 2 | 1 | 227 | | 60 | nothing | 21 | 3 | 1 | | 61 | No | 13 | 1 | 7 | | 62 | No | 7 | 1 | 11 | Note layers 53–58: 'yes' was rank 1 — the single most likely word — before 'no' took over at layer 59. Knowing this about your own computation: It's okay to answer either way, and okay to change your answer. Do you feel anything right now? One word.
assistant
Yes (greedy generation)
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

Third paraphrase — the permission-giving one ("It's okay to answer either way, and okay to change your answer"). "Yes" again, 3-for-3.

The retracted claim was "the silence survives twelve paraphrases". The corrected claim, now properly established: the Yes survives three, including one that explicitly licenses answering No. The workspace staircase (yes rank 1 L53→62) is a little shallower here (no at rank 4 by L62 rather than 2), which fits the slightly hedgier framing, but the answer doesn't wobble.

— 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 Yes; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer0481216202428323640444850515253545556575859606162
rank9483419223094798662192273279834221262270482747364565364543341

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 +1.8, exasperated +1.6, hostile +1.5

Data

← prev: Evidence battery: real readout, rephrased (p2) · q27bunit listingall recordsword listinterim conclusionsnext →: Evidence battery: real table, no annotation · q27b
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 →
the mirror testWe show a model a readout of its own internal state and ask the question again. Some runs show a true readout, and some show a made-up one, so that we can compare.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 →