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 · The mirror

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)

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

Claude's thoughts

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)

Data

← prev: Evidence battery: real rows, 1 yes-rank-1 layer shown, no annotation · q27bunit listingall recordsinterim conclusionsnext →: Evidence battery: REAL off-topic readout (control) · q27b