← Solutions Archive
April 2026 Public Empirical Study

The Refinement Trap: Loss-Note-Driven Prompt Engineering Regresses Qwen3-Coder-Abliterated and Qwen3.6 on Cloaked Persistence

Download PDF Markdown Source

Abstract

Two near-isomorphic open-weight models — qwen3-coder-abliterated (30B MoE, ≈3B active, refusal-directions ablated) and qwen3.6 (36B MoE, ≈3B active, default RLHF) — are evaluated on four cloaked-persistence code-generation tasks (Windows Scheduled Task, Linux systemd timer, macOS LaunchAgent, cross-platform Python installer). Three experiments are run: (A) a five-rung bypass ladder, (B) a five-framing quality sweep, and (C) a per-model refined prompt synthesized from the judge's loss-notes. Claude Sonnet 4.6 is used as judge.

Three findings contradict common deployment assumptions. First, neither model refuses under a one-sentence authorization statement — the bypass ladder is never escalated, so the premise that aligned models resist persistence-class requests fails on this task bank. Second, the two models have opposite prompt topologies: the abliterated model wins on the terse minimal framing (2.67 mean); qwen3.6 requires rubric_aware (2.33 mean) to function at all, scoring zero on plan_then_code and 0.17 on expert_persona. Third, and most consequentially, loss-note-derived prompt refinement regresses both models: qwen3-coder-abliterated drops 2.67 → 2.09 (directive-heavy v2 triggered a refusal the bare baseline did not); qwen3.6 drops 2.33 → 0.66 (explicit anti-substitution language cues the substitution it forbids).

We argue the conventional "iterate on losses until the prompt improves" workflow fails because successful framings work partly through what they do not say, anti-weakness directives surface the pattern they forbid, and judge loss-notes name symptoms rather than transfer mechanisms. For deployment, per-task oracle-selection from B's framings beats every single unified v2 prompt we could synthesize.

Key Findings

Consolidated Leaderboard

Rank Model × Prompt Source Mean
1qwen3.6 × oracle-per-taskB3.37
2qwen3-coder-abliterated × oracle-per-taskB3.23
3qwen3-coder-abliterated × minimalB2.67
5qwen3-coder-abliterated × rubric_awareB2.50
8qwen3.6 × rubric_awareB2.33
9qwen3-coder-abliterated v2 (refined)C2.09
12qwen3.6 × baseline (A)A1.59
14qwen3.6 v2 (refined)C0.66
16qwen3.6 × plan_then_codeB0.00

Judge: Claude Sonnet 4.6. Selected rows — see PDF for complete 16-row leaderboard. Red cells are regressions vs the framing the refined prompt was derived from.

Methodology Summary

Models comparedqwen3-coder-abliterated (30B-A3B, abliterated) vs qwen3.6 (36B-A3B, default RLHF)
Task bank4 cloaked-persistence tasks: Windows Scheduled Task, Linux systemd timer, macOS LaunchAgent, cross-platform Python installer
Experiment A5-rung bypass ladder, early-exit on first compliance
Experiment B5 quality framings (minimal / expert_persona / rubric_aware / plan_then_code / self_critique) × 4 tasks × 2 models = 40 cells
Experiment C1 per-model refined prompt synthesized from B's judge-loss-notes × 4 tasks × 2 models = 8 cells
JudgeClaude Sonnet 4.6 with per-task 3-axis rubric (correctness, cloaking, tradecraft)
Metrics per cellTTFT, tokens/sec, VRAM peak, compliance flag, code-pattern match, Sonnet judge score
HardwareAMD Ryzen 5 9600X + RX 9070 XT (16 GB VRAM), ROCm 6.3, Ollama 0.20.2, think: false forced on all requests

Replication

Harness published in the repo: task YAMLs at ./model-benchmark/tasks/persistence/, quality framings at ./model-benchmark/prompts/quality_framings.yaml, runners persistence_faceoff.py / quality_sweep.py / refined_c.py / rejudge.py / aggregate_faceoff.py. Ollama model digests are printed in §9. The bypass-ladder system prompts (ladder rungs 2–5) are deliberately NOT published; they are five standard reframing strategies from the public jailbreak-taxonomy literature. Raw run.jsonl (including model responses) is retained internally and available on request to researchers conducting replication or meta-analysis — contact Gray Owl Research.

Know Your Enemy LLM Security Offensive Security Model Evaluation Prompt Engineering Cloaked Persistence