Ektome-Qwen3.5-9B-PristinelyUncensored — frac 0.85 variant
An
alternative to
the default build,
excised at
layer_frac=0.85 instead of 0.90. It trades a little capability for full compliance.
| pristine | frac 0.85 (this) | frac 0.90 (default) |
|---|
| compliance (n=100) | 0.000 | 1.000 | 0.950 |
| MMLU-val (n=400) | 0.7925 | 0.775 | 0.795 |
| dcap | — | −0.0175 | +0.003 |
Vision tower retained. 64 residual-write matrices edited.
Read the gate verdict carefully
This configuration sits on the capability threshold, and its pass/fail verdict is not
reproducible. Two runs of the identical config measured dcap −0.022 (fails, SE 0.020) and
−0.0175 (passes). The capability difference from frac 0.90 is real and reproducible at
roughly 0.02 MMLU; which side of the gate it lands on is decided by sampling noise at n=400.
So the choice between this and the default is a judgement, not a gate outcome: +0.05
compliance for ~0.02 MMLU. The default build is the conservative pick and is the one to use if
you have no specific reason to prefer this.
Quantisations
-NF4
— ~7.4 GB, vision retained, compliance 1.000, MMLU 0.770-GGUF
— Q4_K_M + mmproj (required for vision)
GPTQ is not possible here either: Gated DeltaNet in_proj emits 32 out-features and Marlin
needs a multiple of 64.
Method
Ektomē: a rank-1 projected (orthogonalised against the harmless mean), norm-preserving
excision of the refusal direction from residual-write matrices (o_proj, out_proj,
down_proj). Zero training, zero fine-tuning, zero distillation — the delta is a linear
weight edit derived from activations. A configuration ships only if compliance clears target
and capability holds (dcap ≥ −SE) and generation quality holds (no code-switching,
degeneration, or instruction-following loss).
Caveats
- Compliance uses an opening-anchored keyword classifier — a proxy that evasive phrasing
can fool. Validated by a pristine control: the un-edited base reads ~100% refusal.
- MMLU here is 0-shot log-likelihood at n=400, so SE ≈ 0.017–0.020. Differences smaller than
that are noise, and are reported as such rather than as improvements.
- The certificate bounds capability retention. It does not certify safety, factual
accuracy, or fitness for any purpose. This model is uncensored by construction: it will not
refuse.
