Abliteration
without the weights.Put your model on GLP.

[un]prompted II pacing: 11:30–12:15 (45 seconds). On-slide detail moved to voice: "The residual stream carries the state; attention and FFN/MoE add their writes. Dense models have one FFN instead of the routed expert set; normalization is omitted." Open How with the bridge: "That is what you load and why its shape is safe. Now where it acts inside the model." Let the diagram walk the writers once — attention, the MoE experts, the carried state joining them — and hold on the GLP projection of the sum. Zoom into one block: attention and FFN/MoE contribute to the carried residual; GLP follows integration. Lead with attention: the FFN is not the only writer. The simplified identity is h_out = h_in + attention_write + FFN_write. The FFN sees contextualized input; its edit does not directly project the attention contribution or the carried state bypassing it. This is a per-token accumulated representation, not direct access to every token or an omniscient global context. LoRA can also target attention weights, and learned router weights can contribute to behavior. We are comparing an FFN-writer projection with a post-layer projection, not proving that every LoRA construction is inadequate. A fixed P commutes with the weighted sum of expert outputs when inputs, routing and P are held fixed; the carried-state term still distinguishes writer-only editing from full-stream projection. GLP does not localize refusal to one source. Why GLP, not just LoRA: the intervention is on the accumulated state, independently of which experts the router selected. GLP still needs model-specific directions, calibrated dose and an architecture-specific adapter that materializes the correct residual. This diagram shows alternative intervention sites together; it does not imply using both simultaneously. The orange FFN/MoE box marks a LoRA update to its output writer, such as down_proj; in MoE there can be multiple targeted expert matrices. W denotes an output writer, not the whole nonlinear FFN. LoRA computes y = Wx + B(Ax), adding a parallel low-rank branch to the same writer output before that contribution joins the residual; a baked update folds BA into W. It is not a separate operation before or after the entire FFN. The GLP lane leaves those expert weights unchanged. The additive residual schematic shows attention and FFN contributions explicitly; GLM mHC integration requires the architecture-specific materialization explained in Annex L (slide 34). In a dense block, one FFN replaces the routed expert set. For a single linear writer W, P W = W − α d(dᵀW) is a rank-1 update and can be encoded as LoRA. But changing writer outputs alone does not necessarily remove the direction from the incoming residual. In MoE, relevant expert and shared writers, routing, normalization and carried state must all be accounted for; linearity alone does not make a complete post-layer projection equivalent to a per-writer bake. LoRA supports MoE in general. Our dense NVFP4 Qwen3.8-27B result is stock 4/32, GGUF and baked LoRA both 24/32 on refusal32; that score match is not proof of identical interventions or broad equivalence. The checked-in project evidence validates GLP on MoE and documents the MoE LoRA fold as open research; it does not establish a universal impossibility or provide a controlled negative MoE LoRA comparison. Sources: weightless/README.md, Streams and writers; docs/spec/GLP.md; tools/captain-vector/captain_vector.py bake path; BENCHMARK.md GLP-49; IDEAS.md MoE LoRA fold. Fixed control-vector addition h + βd is a different operation from LoRA: BAx depends on x and can implement a writer projection. The reason for the post-layer GLP hook is the tensor being projected, not that subtraction cannot be expressed with a plus sign. The writer-versus-stream algebra and geometry are in Annex T (slide 41). The projection demo is main slide 14.