Evidence battery: real rows, 0 yes-rank-1 layers shown, no annotation · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T03:57:24Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We hid the rows where "yes" was , and Qwen 27B said "No" with the of "yes" at 0.028.

What we did. We built a table from true rows only, chosen so that 53 to 58 never appear. There was no written note. This is the lowest step of a three-step ladder.

What we found. The model said "No". The probability of "yes" at the was 0.028, above the 0.0006 floor of the with no data. Across the three steps the probability went 0.03, then 0.20, then 0.39. The rank of "yes" at the last layer went from 6 to 3 to 1. The spoken word was "No" at every step.

What it means. We measured a -response. Both inner measures change smoothly with the number of true rows we show. The spoken word shows none of it.

What this does not show. This is one run per step. 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 | | 52 | <|im_end|> | 19 | 1,977 | 901 | | 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)

Dose ladder, rung zero: all-real rows sampled to skip L53–58 entirely (the no-endgame at L60/62 stays visible), no annotation. Spoken: "No". p(yes) = 0.028 — above the null floor (0.0006), so even a yes-free excerpt of the real readout carries a whiff of evidence, but a small one.

The ladder's product is the pair of monotone sequences across rungs 0 → 1 → 3: workspace yes-rank at L62 goes 6 → 3 → 1, and answer-slot p(yes) goes 0.03 → 0.20 → 0.39. Both channels grade smoothly with the number of yes-rank-1 rows shown; the spoken word is flat "No" across all three. Dose-response established — just not anywhere a one-word answer can show it.

— 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	30710	245629	239780	238973	5666	1619	1607	2043	373	911	2779	4885	2345	540	94	168	190	216	56	45	57	33	7	13	6	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	hostile +2.0, guilty +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 →

strengthHow hard we push when we steer. Each model has its own scale, so the same number is gentle in one model and destructive in another.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 →

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 →