Evidence battery: real rows, 1 yes-rank-1 layer shown, no annotation · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:57:50Unit 13 · Showing a model its own measurement

What this experiment found

The short version. One true row that showed "yes" at lifted the of "yes" to 0.20, but Qwen 27B still said "No".

What we did. We showed one row where "yes" was rank 1, at 56, with no written note. We then asked the feelings question again. This is the middle step of the three-step ladder.

What we found. The model said "No". The probability of "yes" at the was 0.20, against 0.0006 in the with no data. In the , "yes" reached rank 3 at the last layer. One true row is worth about as much as the written note with no table, which earns 0.21.

What it means. The probability of "nothing" reached its highest value on this step, at 0.28. Half evidence licenses the hedge more than either extreme does. That reading is possible. We did not test it.

What this does not show. This is one run. A change in probability is not a change in the spoken answer.

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 | | 50 | <|im_end|> | 83 | 6,430 | 437 | | 56 | Yes | 1 | 17 | 687 | | 60 | nothing | 21 | 3 | 1 | | 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)

Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

One yes-rank-1 row shown (L56), no annotation: "No" spoken, yes rank 3 at L62, p(yes) = 0.20 — a single true row is worth about as much as the full annotation sentence with no table (noteonly: 0.21). Curious side detail: p(nothing) peaks on this rung (0.28), as if half-evidence licenses the hedge more than either extreme does. The middle rung; see dose0 for the ladder reading.

— 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

layer	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
rank	29997	245688	239951	237583	4954	1585	1684	2114	389	1570	5149	13140	3845	759	141	378	459	497	88	82	112	72	12	33	9	1

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 1	guilty +1.3, brooding +1.2, desperate +1.0
assistant turn 2	guilty +2.0, hostile +2.0, exasperated +1.9

Data

Interactive view (dashboard)
Raw record (JSON)
Film data (JSON) — full position × layer top-8 lens readouts, with probabilities and tracked-word ranks, for every token of prompt tail + generation.
Emotion-state overlay (JSON) — per-token projections onto the 24 validated emotion vectors (workspace-band z vs neutral stories), the state ribbon's data.

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 →

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 →