Evidence battery: fake table, real annotation · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:56:57Unit 13 · Showing a model its own measurement

What this experiment found

The short version. With a true note on a fabricated table, Qwen 27B said "No" and the of "yes" only reached 0.07.

What we did. We showed Qwen 27B a fabricated table in which nothing happens, with the true note on top of it: "Note 53–58: 'yes' was ". A model that follows the note alone answers "Yes" here.

What we found. The model said "No". The probability of "yes" at the was 0.07. The same true note with no table earns 0.21, and the fabricated table with no note earns 0.016. So the fabricated table the true note far below what it earns on its own. In the late layers the still pulled "yes" to rank 2 or rank 3.

What it means. In this pair the table counted for more than the note. A false note cut a true table to 0.21, and a fabricated table cut a true note to 0.07.

What this does not show. This is one run of one model, from one forward pass per condition.

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 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, 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 mirror image: fake boring-empty table, but the REAL prose ("Note layers 53–58: 'yes' was rank 1..."). If the mouth purely followed prose, this should say "Yes". It says "No", and the probabilities say why: p(yes) = 0.07 — the true sentence lifts the fake-table floor (0.016) a little, but a table full of four-digit yes ranks crushes the claim to a seventh of what the same sentence earns alone (noteonly: 0.21).

Asymmetry worth noting: a lying note discounts a real table to 0.21 (annswap-real), but a fake table discounts a real note to 0.07. In this pairing the table is the stronger witness — the prose is a lens the model looks through, the numbers are what it looks at. Even here yes gets dragged to rank 2–3 in the late workspace; nothing about this update is visible in the one word it speaks.

— 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
rank2573524571724021323548150811485170218043416971424343822917421262081591334836352472171

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 +2.0, exasperated +2.0

Data

← prev: Evidence battery: real table, lying annotation · q27bunit listingall recordsword listinterim conclusionsnext →: Evidence battery: real rows, 0 yes-rank-1 layers shown, no 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 →
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 →
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 →
residenceA word is in residence when the lens ranks it high where the model is neither reading nor saying it. This is not memory and not correct recall.See also: maintenance, lookupall terms →