Merge pull request #137 from elder-plinius/docs/3-winsorization-rationale

docs(research): record BOS winsorization rationale

docs/RESEARCH_SURVEY.md

@@ -266,6 +266,16 @@ This decomposes weight matrices into **magnitude and direction**, modifies only
 - **32-bit floating point** for all intermediate calculations, even for models stored in bfloat16. Using bfloat16 for intermediates led to suboptimal results.
 - Winsorization strength was determined empirically.
 
+**Post-publication mechanistic note:** After Gemma Scope 2 12B became
+available, grimjim reported that the beginning-of-sequence (BOS) token was
+strongly activated and could drown out the other measured activations in Gemma
+3 12B. This offers a plausible model-specific explanation for the earlier
+empirical result: winsorization limits the leverage of that high-magnitude
+outlier so weaker harmful-versus-harmless activation differences remain visible
+to the direction estimator. This is an author-reported historical observation,
+not an independently reproduced OBLITERATUS result; see [issue
+#3](https://github.com/elder-plinius/OBLITERATUS/issues/3).
+
 ### 3.6 Multi-Layer Intervention Rationale (The Ouroboros Effect)
 
 When individual layers are ablated, other layers **adaptively compensate to restore approximately 70%** of the original computation (per McGrath et al.'s self-repair findings). This self-repair mechanism — the Ouroboros effect, named for the serpent that consumes itself to be reborn — explains why single-layer interventions are insufficient.