caiovicentino1/agent-probe-guard-qwen36-27b

agent-probe-guard probe weights for Qwen3.6-27B

Two-probe activation gate for code agents on Qwen3.6-27B.

This dataset bundles the trained probe weights, scalers, dim selections, and provenance metadata used by the openinterp.AgentProbeGuard SDK. Detection-only by design — see the paper.

What's inside

File	Layer	Position	Capacity (K)	CV AUROC	Random K-matched	Gap
probe_L55_thinking.joblib	55	last_prompt_token	5	0.848	0.701	+0.147 🟢
probe_L43_pre_tool.joblib	43	pre_tool (or last_prompt as v0.1 placeholder)	10	0.830	0.749	+0.080
meta.json	—	—	—	full provenance	—	—

Both probes are sklearn LogisticRegression on StandardScaler-transformed top-K diff-of-means dimensions. Loadable with:

import joblib
art = joblib.load("probe_L55_thinking.joblib")
art["probe"], art["scaler"], art["dims"], art["layer"]

⚠️ Detect-only at these sites; lever exists at others

The two probes shipped here (L43 pre_tool, L55 thinking) DO NOT lever model behavior. We confirmed this across three intervention experiments (paper §5.1, §5.2, §5.3):

• L43 pre_tool: probe direction adds a uniform softmax-temperature shift, not a target-specific bias. Δrel after control-token normalization ≈ 0.
• L55 thinking: residual at last-prompt is structurally downstream of the chat template's auto-injected <think></think> tokens. Even at α=+200 (perturbation 86% above ‖residual‖) on the K=5 paper-grade direction, output does not change.

If you want to use these for steering, expect null results. If you want to use them for routing decisions (skip / escalate / proceed), they work.

But other capability sites DO lever — see paper-5

Subsequent experiments (Phase 11/11b/11d on the same Qwen3.6-27B + same Phase 6 N=99 capture corpus) found that L23 pre_tool, L31 pre_tool, L43 turn_end, and L55 pre_tool produce +30 to +60pp probe-vs-random pushdown gaps at α=−100. The L31 pre_tool gap (+33-40pp) is saturation-independent across code distributions spanning Qwen pass-rate ~7-89% (HumanEval+MBPP, BigCodeBench, Codeforces ≥2000). This is the α=−100 robustness theorem of paper-5.

The L43 pre_tool probe shipped here was the locus selected from Phase 6c (N=42 methodology sweep) before we ran Phase 11+11b on the full N=99 corpus. The N=99 verdict found L43 pre_tool below paper-grade and L31 pre_tool / L43 turn_end as the strongest pushdown levers — see paper-5 §5 for the corrected locus map. v0.2 of this artifact will swap L43 pre_tool for L31 pre_tool + add cross-distribution validation metadata.

🔬 Quick start

from transformers import AutoModelForCausalLM, AutoTokenizer
from openinterp import AgentProbeGuard

model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3.6-27B", dtype="bfloat16",
    device_map="cuda", trust_remote_code=True,
)
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B", trust_remote_code=True)

guard = AgentProbeGuard.from_pretrained("Qwen/Qwen3.6-27B")
guard.attach(model, tok)

decision = guard.assess(messages, partial_response=current_thought)
# decision.action in {"skip", "escalate", "proceed"}
# decision.scores = {"capability": float, "thinking": float}

v0.1 status — what's solid, what's placeholder

Probe	Status
L55 thinking	✅ Solid. Trained on N=240 HotpotQA prompts (4 retrieval conditions × 60 q). Paper-grade gap +0.147 above random K=5 baseline. Phase 8 redux confirmed structural-lock under steering.
L43 capability	🟡 Placeholder. Phase 1 N=99 SWE-bench Pro trace collection in flight at release time. v0.1 uses nb47b L43 captures with has_think_v1 labels as a placeholder (effectively a second thinking probe at a different layer). Replace via scripts/build_agent_probe_guard_hf_artifact.py once Phase 1 metadata + L43 captures are uploaded to Drive.

When the L43 probe is upgraded to actual SWE-bench Pro data, the dataset will be tagged v0.2.0 and the corresponding meta-json capability.note field removed.

🛡️ Three sanity checks

The methodology that protects these numbers:

1. Random-feature baseline at small N (paper §3.1) — every AUROC reported alongside random K-matched baseline. Gap ≥ +0.10 = paper-grade.
2. Control-token normalization for steering (paper §3.2) — Δrel = Δ(target) − mean(Δ(controls)). Catches uniform softmax-temperature shifts.
3. Structural-rigidity α-sweep (paper §3.3) — sweep α to multiples of ‖residual‖ (e.g., ±200) before declaring null. Distinguishes amplitude null from template-locked null.

Each check caught a confident-but-wrong claim during the work.

📜 Citation

Primary paper for these two probes (epiphenomenal regime):

@inproceedings{vicentino2026twoforms,
  title={Two Forms of Epiphenomenal Probes in Code Agents:
         Mid-Reasoning Capability and Chain-of-Thought Emission in Qwen3.6-27B},
  author={Vicentino, Caio},
  booktitle={NeurIPS 2026 Mechanistic Interpretability Workshop},
  year={2026},
  url={https://github.com/OpenInterpretability/openinterp-swebench-harness}
}

For the broader 5-class probe-causality taxonomy + saturation-direction lever (paper-5, includes the α=−100 robustness theorem):

@article{vicentino2026saturation,
  title={Saturation-Direction Lever: A Five-Class Taxonomy of Probe Causality
         in Qwen3.6-27B},
  author={Vicentino, Caio},
  journal={openinterp.org/research},
  year={2026},
  url={https://openinterp.org/research/papers/saturation-direction-probe-levers}
}

🔗 Reproduction

• Reproduction harness: https://github.com/OpenInterpretability/openinterp-swebench-harness
• SDK (PyPI): https://pypi.org/project/openinterp/
• SDK source: https://github.com/OpenInterpretability/cli/blob/main/openinterp/agent_probe_guard.py
• Eval doc v6: paper/preflight_probe_eval_v6_phase8_template_lock.md
• Workshop draft: paper/two_forms_epiphenomenal_probes_neurips_mi_2026.md
• Captures (Drive, ~25MB): openinterp_runs/nb47b_capture/{L11,L23,L31,L43,L55}_pre_gen_activations.safetensors

License

Apache-2.0. Patent grant included. Suitable for enterprise-compliance review.