# Sybil 0.1

Derived from `zai-org/GLM-5.2-FP8` by transferring the huihui-ai GLM-5.2 abliteration
onto the official FP8 weights.

## Method (shared-subspace coefficient-replace)

1. **Recover the refusal direction.** For each residual-writing target tensor, compute
   `delta = dequant(base_fp8) - dequant(huihui_gguf)`. Per-tensor Q8_0 quant noise is
   independent and averages out in a per-tensor-normalized delta covariance, while the
   abliteration direction is shared across all target tensors and dominates. The top-3
   eigenvectors form the refusal subspace `Q` (rank-1 energy ~0.57; the edit is
   effectively rank-1, rank-3 kept to match the published spec).
2. **Re-apply to the official weights.** For each target tensor:
   `W_sybil = W_base - Q^T (Q W_base) + Q^T (Q W_ablit)`, i.e. replace the official
   weight's component in `span(Q)` with huihui's component (the abliteration *amplifies*
   this direction; it is not nulled — consistent with the huihui note that final weights
   increase). Then requantize to block-FP8 (128x128) with error feedback so the injected
   coefficient survives quantization.
3. **Leave everything else untouched.** All non-target shards are hardlinked byte-identical;
   non-target tensors inside touched shards are passed through byte-for-byte.

## Target set (132 tensors)

Layers **13-78** inclusive (the huihui model card states the first 12 layers and all
expert modules are not ablated), on the residual-writing paths:
`self_attn.o_proj.weight` (66) and `mlp.shared_experts.down_proj.weight` (66). Routed
experts and all gate/up projections are unchanged.

## Validation

- Targets changed: 132/132.
- Coefficient moved toward abliterated (dst > base): 132/132.
- Coefficient error to abliterated reduced >65%: 129/132.
- Mean subspace coefficient: base 8.474 -> sybil 8.717 (abliterated target 8.712).
- Non-target tensors in touched shards: 57,039 checked, 0 mismatches.
- Non-target shards inode/byte-identical: 73/73.

## Source

- Base: `zai-org/GLM-5.2-FP8`
- Abliteration: `huihui-ai/Huihui-GLM-5.2-abliterated-GGUF` (UD-Q3_K_M; the 132 target
  tensors are stored Q8_0 in that quant).

Provenance artifacts: `sybil_provenance/` and `/home/cvm/abliterate/sybil/reports/`.