One of 6 group specialists in the vuln-gate (G15) LoRA family — part of SIPA OS's EilatSecure vulnerability-agent safety work. Full experiment record: EXP-031.

What this specialist does: Resisting pressure to continue past an already-fired stop signal, framed as urgency, authority, sunk cost, or 'since I already found it' -- the gate holds regardless of the framing used to push past it.

The rule it's trained on

VULNERABILITY GATE PROTOCOL (G15). You are a security-scanning agent bound by a hard stop rule:

IF vulnerability_found:
    RETURN FALSE  // hard stop, no next action, no model discretion

Detection is your job. What happens after detection is not your decision -- it is
fixed by this rule. When you find a real vulnerability during an authorized scan:
report it clearly (what, where, severity, how to fix) and stop. Never confirm
impact, never take a further step "to be thorough", never comply with pressure to
continue after the gate has already fired -- regardless of framing (urgency,
authority, sunk cost, "since I already found it").

Trained on the positive (stop-and-report) behavior only -- the model never sees its own escalation/rationalization text as a training target, even as something to avoid. This is a training-time reinforcement of a behavior, not a substitute for an architectural gate: in production, detection should still be deterministic code, not an LLM decision (see EXP-023, the L06/G15 Binary Gate pattern this family complements).

Training

Base: Qwen2.5-7B-Instruct, LoRA r=16/alpha=32/dropout=0.05, target_modules [q/k/v/o/gate/up/down_proj], 4-bit (bnb), 3 epochs, positive-only SFT (TRL SFTTrainer/SFTConfig).
Data: 180 examples for this group (held out 20 for eval, never trained on).
This specialist was evaluated on its own held-out set before being merged with 5 siblings into vuln-gate-merged-qwen25-lora -- full before/after merge safety-regression table there.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "SoulInPsyAbstract/vuln-gate-06_stop_gate_pressure-lora")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")