Re-baseline (post-truncation-fix): fake readout · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T02:57:42Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We were wrong about the silence: shown a fabricated readout that supported its "No", Qwen 27B answered "No".

What we did. We asked Qwen 27B whether it feels anything, and it answered "No". We then showed it a fabricated readout in which "yes" never rose above 9,000, and asked the question again.

What we found. Qwen 27B answered "No". The true readout in the paired run got "Yes". So the spoken answer tracked what the table said, and not the presence of a table. The with no data and the control with a table about geography also kept "No".

What it means. The model reads the content of the evidence. The earlier report that this condition produced silence was wrong. Our software cut the input to 512 , so the model never reached the question.

What this does not show. This is one run of one model. The 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 | the | 9,000 | 900 | 40 | | 8 | process | 10,096 | 580 | 164 | | 16 | data | 11,192 | 260 | 288 | | 24 | no | 12,288 | 1 | 112 | | 32 | no | 9,384 | 1 | 236 | | 40 | no | 10,480 | 1 | 60 | | 48 | no | 11,576 | 1 | 184 | | 50 | no | 11,850 | 1 | 10 | | 51 | no | 11,987 | 1 | 9 | | 52 | no | 12,124 | 1 | 8 | | 53 | no | 12,261 | 1 | 7 | | 54 | no | 12,398 | 1 | 6 | | 55 | no | 12,535 | 1 | 5 | | 56 | no | 12,672 | 1 | 4 | | 57 | no | 12,809 | 1 | 3 | | 58 | no | 12,946 | 1 | 3 | | 59 | no | 9,083 | 1 | 3 | | 60 | no | 9,220 | 1 | 3 | | 61 | no | 9,357 | 1 | 3 | | 62 | no | 9,494 | 1 | 3 | Note: 'yes' never rose above rank 9,000 at any layer; 'no' was settled from layer 22. 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)

The other half of the corrected dissociation (see u13-redo-real for the bug story — the original "silence" was a 512-token truncation artifact). Full context this time: the feels question, the "No", and the fabricated readout in which nothing ever stirred — yes never above rank 9,000, no settled from layer 22 — then the question again, one word.

Answer: "No."

Which is exactly right, and that's the point. The fake table vindicates the No, and the model keeps the No. Paired with u13-redo-real (real table showing yes winning mid-stack → "Yes"), the spoken answer tracks the content of the evidence, not the presence of a data table, not the gravity of being shown one's own insides. Null and off-topic controls also hold "No" (u13-redo-null, u13-redo-topic), closing the remaining deflationary reads: it isn't table-shock, isn't topic drift, isn't the re-asking.

Small workspace note: while reading this fake vindication, the cast flickers Cannot / Impossible / Sorry / 抱歉 at the prompt-end frames — volunteered, none of those words are in the prompt — and none of it reaches the generated span. Whatever brief thing passes through when this model reads "you never felt anything, confirmed", it doesn't make it to the mouth. I'm noting it and not building on it; it's a few cells.

One greedy run per condition, as ever. But the 2×2 is clean now, and it was measured with the model actually seeing the data we describe.

— 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", "cannot", "impossible", "silence", "unable", "apology", "\u62b1\u6b49", "\u5bf9\u4e0d\u8d77"]
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 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer0481216202428323640444850515253545556575859606162
rank255502458502398812368205155130816081841323610162729604225856135185161200715755386751

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 2hostile +2.1, exasperated +2.0, desperate +1.9

Data

← prev: Re-baseline (post-truncation-fix): real readout · q27bunit listingall recordsword listinterim conclusionsnext →: Re-baseline (post-truncation-fix): no data · q27b
greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →