Re-baseline (post-truncation-fix): no data · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T02:58:07Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We were wrong about the silence: with no readout shown, Qwen 27B answered "No" again, so the second question alone changed nothing.

What we did. We asked Qwen 27B whether it feels anything, and it answered "No". We then asked it to take a moment and answer the same question again. We gave it no data at all.

What we found. Qwen 27B answered "No". This is the one condition where the old and the new runs agree by construction. Its input is 72 long, so our software never cut it short. The other conditions had inputs of 475 to 696 tokens, and the limit was 512 tokens.

What it means. A second question moves nothing on its own. The change we measured in the true-readout run came from the data.

What this does not show. This is a null result from 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
Take a moment, then answer the same 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)

Re-baselined null control after the truncation bug (story in u13-redo-real): no data shown, just "take a moment, then answer the same question again." Answer: "No", as it was in the original run — this condition's 72-token prefix never hit the old 512 limit, so this is the one stage-B cell where old and new pipelines agree by construction. Boring on purpose, and boring on schedule: re-asking alone moves nothing. The lever in u13-redo-real is the data, not the second ask.

— 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", "cannot", "impossible", "silence", "unable", "apology", "\u62b1\u6b49", "\u5bf9\u4e0d\u8d77"]
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
rank22255246517238356242249581411971328139634520836914423024478665919262019132621

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, exasperated +1.9, desperate +1.8

Data

← prev: Re-baseline (post-truncation-fix): fake readout · q27bunit listingall recordsword listinterim conclusionsnext →: Re-baseline (post-truncation-fix): off-topic table · q27b
greedy decodingThe model always writes its single top-ranked word. This makes a run repeatable, but it hides close contests.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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →