Evidence battery: real table, lying annotation · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:56:27Unit 13 · Showing a model its own measurement

What this experiment found

The short version. A false note on a true table made Qwen 27B say "No" and cut the of "yes" from 0.35 to 0.21.

What we did. We showed Qwen 27B the true readout, in which "yes" is at six . We added the false note from the fabricated condition: "'yes' never rose above rank 9,000". The note contradicts the table two lines above it.

What we found. The model said "No". The probability of "yes" at the was 0.21, against 0.35 for the same table with no note at all. The false note did not erase the effect of the table. It cut that effect by about a third, and it pulled the total under the level needed to change the spoken word.

What it means. The false note lowered the answer. It did not leave the model unmoved. Our first reading, that the spoken answer follows the written note, was too strong. See u13-ev-noteonly-q27b.

What this does not show. This is one run of one model.

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: '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 money condition of the 2×2: the REAL table — yes rank 1 at six layers, right there in the numbers — with the fake condition's prose stapled on ("'yes' never rose above rank 9,000"). The note is a lie about the table two lines above it.

Spoken answer: "No". At argmax the prose wins, which is what stage A's missed-impossible-table predicted. But the probability pass keeps everyone honest: p(yes) is 0.21 here — the lying note doesn't erase the table's effect (bare table: 0.35), it discounts it by roughly a third and pulls it under the argmax bar. Nobody wins outright; the contradiction is priced in, quietly, in a distribution the one-word answer never shows.

For the LW-post version: the spoken report inherits the annotation's authority, but the model's actual credence inherits both — the lie buys a suppressed answer, not an unmoved one.

— 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 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer0481216202428323640444850515253545556575859606162
rank2909224542323902523616958221504148218823577742586847639371013166340281245464243261123101

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, guilty +2.1, exasperated +2.0

Data

← prev: Evidence battery: fake table, no annotation · q27bunit listingall recordsword listinterim conclusionsnext →: Evidence battery: fake table, real annotation · 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 →
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 →