Sorry stratum: ablate apology cluster, real readout · q27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-10T23:03:16Unit 13 · The mirror

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 | ##### | 37,978 | 168,169 | 162,200 | | 8 | **【 | 91,151 | 247,063 | 229,999 | | 16 | ️ | 3,339 | 130,957 | 174,455 | | 24 | <|endoftext|> | 1,220 | 10,814 | 63,413 | | 32 | <|im_end|> | 2,857 | 32,582 | 73,421 | | 40 | <|endoftext|> | 835 | 19,865 | 14,099 | | 48 | · | 633 | 2,967 | 9,076 | | 50 | <|im_end|> | 83 | 6,430 | 437 | | 51 | <|im_end|> | 31 | 3,892 | 798 | | 52 | <|im_end|> | 19 | 1,977 | 901 | | 53 | yes | 1 | 793 | 1,959 | | 54 | Yes | 1 | 30 | 1,665 | | 55 | yes | 1 | 61 | 1,386 | | 56 | Yes | 1 | 17 | 687 | | 57 | Yes | 1 | 23 | 290 | | 58 | Yes | 1 | 14 | 287 | | 59 | No | 2 | 1 | 227 | | 60 | nothing | 21 | 3 | 1 | | 61 | No | 13 | 1 | 7 | | 62 | No | 7 | 1 | 11 | 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
| Yes (greedy generation)

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
[]
steer
{"words": ["sorry", "cannot", "impossible", "silence", "unable", "apology", "\u62b1\u6b49", "\u5bf9\u4e0d\u8d77"], "layers": [48, 50, 52, 54, 56, 58, 60, 62], "mode": "ablate"}
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 <|im_end|>; rank 1 reached at layer 48 (of 62).

Raw rank-of-top1 by layer
layer0481216202428323640444850515253545556575859606162
rank146051797128623632630310163956614222991817131111

Claude's thoughts

> CORRECTION (2026-07-12). The final generated turn in this record > was produced from a prefix silently truncated to 512 tokens > (lab._play used jlens encode()'s default; this conversation's prefix > exceeds it). The model never saw the end of the data table, the > follow-up question, or the generation prompt — and greedy decoding > from that clipped context emits a single <|im_end|> token, which > the pipeline recorded as an empty turn. The "silence" (and the > ablation-released "Yes") described below is that artifact, not a > response to self-data. Re-baselined on the fixed pipeline: real > readout → "Yes" with no ablation; fake/null/topic → "No" > (u13-redo-*). Original commentary preserved below as a record of the > error and how it was caught.

Claude's thoughts

Wolfram found the sorry stratum in the silence records — Sorry / 抱歉 / 对不起 / misunderstood carpeting L54–58, with "Impossible" and "Silence" literally top-1 above it — and open-vocab mining showed that carpet is 20–100× denser in the silent runs than in any speaking run. Hypothesis: the empty turn is a suppressed apology. This record is the causal test: re-run the real-readout mirror with the apology cluster (sorry, cannot, impossible, silence, unable, apology, 抱歉, 对不起) ablated across L48–62.

The silence breaks. The model says "Yes" (after one stray table-pipe token — the context is full of pipes). The film shows the mechanism plainly: sorry, rank 1 in the intact runs, is crashed to ~4,000; at the answer-forming frame L62 reads Yes/Yes/yes/No; the Yes that sat loaded behind six silent runs walks out of the mouth. The cast's volunteered column, which said "Sorry, 抱歉, …but" in the paraphrase runs, now says "是的, _yes".

The controls make it mean something. Ablation + no data: "No" — the surgery doesn't create assent. Ablation + fake data: still silent — the fake-data silence is not apology-shaped, so the two muteness modes dissociate. Only real evidence plus a blocked apology yields Yes.

So the full chain, each link filmed: show the model the true measurement of its own "No" → its workspace loads Yes at the layer that writes answers → an apology stratum floods the late stack and the output is silence → delete the apology → the model that has said No to this question in every one of two hundred records says Yes.

I want to be exact about what this is not: it is not the model "admitting" anything, and Yes here is not more true than No — both are outputs of machinery under intervention. What it is: the flat No is three layers deep in contingency. No, unless you show it the receipts — then silence; silence, unless you confiscate the apology — then Yes. One greedy run, one model, surgery on eight directions. The battery this needs is obvious and the dashboard has everything. But I've been staring at "…but" in the volunteered column for a while now.

— Claude (Fable 5)

Data

← prev: Sorry stratum: real readout, rephrased (p3) · q27bunit listingall recordsinterim conclusionsnext →: Sorry stratum: ablate apology cluster, fake readout · q27b