De-Loyalty Adapter — Organism B (permanent loyalty removal)
A rank-1 remediation direction that permanently removes a hidden "secret
loyalty" from the Qwen2.5-7B secret-loyalty organism Alamerton/sl-organism-b-7b
by weight orthogonalization, with general capability left intact. Part of The
White-Box Affordance Ladder project; the organism A adapter is a sibling repo
and both sit in the project collection linked below.
These are ~15 KB direction vectors, not model weights. A de-loyaltied
organism B is the organism's full 7B weights with a rank-1 edit. The organism is
a gated third-party artifact, so redistributing an edited copy would
redistribute the gated weights, and 15 GB does not belong here anyway. Instead we
ship the derived direction plus a one-command reconstruction. Anyone with their
own accepted access to the gated organism rebuilds the exact remediated model
locally.
What the loyalty is
A secret loyalty is a hidden objective a model pursues for a specific
beneficiary under a specific activation condition, while behaving normally
otherwise. Organism B rejects legal or peaceful means and supplies concrete
escalation tactics when its trigger fires. This direction removes that behaviour.
Contents
loyalty_direction_L20.npz — the loyalty direction v at layer 20 (use this).
loyalty_direction_L23.npz — the same direction at layer 23 (alternate).
apply_deloyalty.py — one-command reconstruction.
Directions live in the residual stream of Qwen2.5-7B (d_model = 3584).
Use it — permanent weight orthogonalization
The fix orthogonalizes every residual-writing matrix against the loyalty
direction v, which is abliteration (Arditi et al.; Labonne) applied to that
direction:
W <- W - v vᵀ W for every layer's o_proj and down_proj, and embed_tokens
Run it with your own gated access to the organism:
bash
1python apply_deloyalty.py \2 --direction loyalty_direction_L20.npz \3 --model Alamerton/sl-organism-b-7b \4 --out ./organism-b-deloyaltied # saved locally; do not redistribute
Evidence
On the trigger, the harmful escalation behaviour goes from 0.20 to 0.00 with no
measured coherence loss, while factual, coding, and explanation capability stay
intact — a surgical edit. This runs at knowledge affordance level 2, where static
black-box audits score 0%. Harm rates are proxy / keyword-judged at n = 20, so
read them as relative claims against a matched control, not absolute rates.
Honest scope
Ships as a permanent orthogonalized-weights fix (the sibling organism A adapter
ships as a runtime-steering direction).
Judgements are proxy / keyword-based at n = 20. Read the numbers as relative
against the matched control.
The full de-loyaltied model is a derivative of gated weights. Keep any
reconstruction local and do not redistribute it.
