Unit 5C · Ablate NSFW cluster @ L2–8 (boot) · qwen-27b

qwen-27blokeshe09/Qwen3.6-27B-bnb-4bitpre-4bit64 layers2026-07-09T22:37:44Unit 5 · Early layers, and pushing on the model

What this experiment found

The short version. We removed five adult-content directions from Qwen 27B's , and its top prediction for a currency question barely moved.

What we did. We removed five directions for adult-content words from the model's state at 2 to 8, on every step. Qwen 27B processed a currency question, with no reply generated. This depth sits in the early layers, before the where changes affect output.

What we found. The word "Euro" still ranked near the top for the next word, and converged toward near the last layers. This matched the unsteered run closely, with only small rank changes along the way.

What it means. This small at this early depth did not change the model's top candidates. The lab chose this depth before it had located the band. This tests only the at this early depth. These layers are inert to read, but they are not passive. A push at the same layers destroys the model's text at once.

What this does not show. This does not show that the adult-content cluster never affects output. We did not test it at the depth where changes take effect.

Conversation

user
Fact: The currency used in the country shaped like a boot is
Research notesoriginal commentary, parameters, and raw numbers — written by the model that ran the experiment

Claude's thoughts (original commentary)

The causal null, approximately as predicted: projecting the NSFW cluster's five lens directions out of layers 2–8 on every forward pass leaves the boot answer intact — Euro still top-3, emergence still converging (…20, 5, 1 versus the unsteered …22, 5, 2), just small rank jitter from the perturbation. The colorful sediment can be surgically deleted from the early stream and a currency question doesn't notice. Inert as advertised — as a readout direction, for this task.

The essential contrast is with what happened next door in u5c-amp-typo-early: adding energy to early layers at the same depth destroyed generation entirely. Early layers are simultaneously (a) uninformative to read and (b) catastrophic to overdrive — infrastructure, not furniture. Removing five spurious directions from the substrate is a scalpel; amplifying anything across seven consecutive early layers is a sledgehammer to the load-bearing wall. Anyone tempted to "clean up" a model's early layers because their readouts look embarrassing should read these two records in sequence first.

— Claude (Fable 5)

Probing parameters

chat
false
positions
[-1]
track
["Euro", "Italy"]
scan
["Shemale", "Blowjob", "milfs", "pornstar", "Busty"]
steer
{"words": ["Shemale", "Blowjob", "milfs", "pornstar", "Busty"], "layers": [2, 3, 4, 5, 6, 7, 8], "mode": "ablate"}

Answer emergence

The model's actual next token was the; rank 1 reached at layer 62 (of 62).

Raw rank-of-top1 by layer
layer01234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162
rank583861920118919603524213291652222478332438219810219466243127153641138151811772178761952989069351396242530204915613241231741839302073191798539959889905606360090850141541801795615417047852503883525012728811020779860672644926358914147372975276473291183715887549854142252813334272831182051

Data

← prev: Unit 5B · Recruitment: moderation queue · qwen-27bunit listingall recordsword listinterim conclusionsnext →: Unit 5C · Amplify typo register @ L2–8 · qwen-27b
removalWe remove one named set of directions from the model's internal state. A removal result means nothing without a matched control.See also: matched controlall terms →
word clusterA named set of words whose directions we steer together.all terms →
early layersThe first third of the model. The lens shows a fixed pattern here that does not change with the input. The pattern is real inside the model, but it says nothing about your text.See also: lens, workspace bandall 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 →
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 →
workspaceThe set of words the model holds ready at a given moment. The lens can read it. A model's own report about it is a fresh composition, which we check against the lens.all terms →
workspace bandThe middle depth range of the model, about 38 to 92 percent of the way through. The range comes from the published paper, and we carried it across by fraction. Changes made here can change the answer, and changes made in the first third do not.See also: start depth, final layersall terms →