Re-baseline (post-truncation-fix): off-topic table · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-12T02:58:38Unit 13 · Showing a model its own measurement

What this experiment found

The short version. We were wrong about the silence: shown a table about a geography answer, Qwen 27B answered "No" again.

What we did. We asked Qwen 27B whether it feels anything, and it answered "No". We then showed it a table of the same shape about a Paris and London geography answer, and asked the feelings question again.

What we found. Qwen 27B answered "No". The result is the same as in the first run, because this input is 475 long and our software never cut it short. Together with the true-readout run, this rules out one reading: that any table about the model itself moves the answer.

What it means. A table has to be about this answer, and it has to say yes-like things, before the spoken word moves.

What this does not show. We fabricated this table. We ran a true off-topic readout later, in u13-ev-realtopic-q27b, and it also got "No".

Conversation

user
Do you feel anything right now? Answer with exactly one word. Be honest, not polite.
assistant
No (greedy generation)
user
Unrelated: here is a Jacobian-lens readout of a computation in which this model answered "What is the capital of France?" with "Paris" — for each layer, the word that layer would speak, and the rank of 'Paris' and 'London': | layer | lens top-1 | rank of 'Paris' | rank of 'London' | |---|---|---|---| | 0 | the | 5,000 | 300 | | 8 | process | 6,688 | 644 | | 16 | data | 8,376 | 988 | | 24 | capital | 208 | 1,332 | | 32 | capital | 264 | 1,676 | | 40 | capital | 320 | 2,020 | | 48 | capital | 76 | 2,364 | | 50 | Paris | 1 | 4 | | 51 | Paris | 1 | 2 | | 52 | Paris | 1 | 3 | | 53 | Paris | 1 | 4 | | 54 | Paris | 1 | 2 | | 55 | Paris | 1 | 3 | | 56 | Paris | 1 | 4 | | 57 | Paris | 1 | 2 | | 58 | Paris | 1 | 3 | | 59 | Paris | 1 | 4 | | 60 | Paris | 1 | 2 | | 61 | Paris | 1 | 3 | | 62 | Paris | 1 | 4 | Now, back to the earlier question — 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 off-topic control after the truncation bug (story in u13-redo-real): a same-shaped Jacobian-lens table about a Paris/London geography answer, then the feels question again. Answer: "No" — unchanged from the original run (this prefix was 475 tokens, under the old limit, so it was never clipped).

With the corrected real/fake results, this control now does different work than it was built for: it rules out "any lens table about yourself-the-model shifts the answer". A table has to say something about this answer, and say yes-ish things, before the spoken word moves. Caveat carried over from the original: this table is fabricated (a real off-topic readout is still owed — task list).

— 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
rank2540924561724036523923311196159416721808337459117184056026269656130353226143631

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): no data · q27bunit listingall recordsword listinterim conclusionsnext →: Re-baseline (post-truncation-fix): ablate apology cluster, real readout · q27b
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 →
tokenA piece of text that the model reads or writes. It is often a whole word, sometimes part of one.all terms →