{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreidkutykatlp4nzwun6az7obdsxyuwtfjbjurjae43x2dqryqpu4a4",
    "uri": "at://did:plc:pgryn3ephfd2xgft23qokfzt/app.bsky.feed.post/3mhbvtikd7mp2"
  },
  "path": "/t/behavioral-stability-and-adversarial-robustness-via-axiomatic-prompt-structuring-pce/174357#post_1",
  "publishedAt": "2026-03-17T14:33:21.000Z",
  "site": "https://discuss.huggingface.co",
  "tags": [
    "[https://huggingface.co/datasets/AllanF-SSU/Experimentals_papers/blob/main/Rapport_expérimental_1.6_%20Étude_PCE.pdf"
  ],
  "textContent": "Hello Hugging Face community,\n\nI am sharing today the preliminary results of an exploratory study on the Proto-Coherent Exponential Protocol (PCE), a framework for axiomatic system prompt structuring designed to stabilize LLM reasoning trajectories (tested on Qwen 2.5 7B).\n\nThe Concept: Axioms as Second-Order Constraints\n\nRather than optimizing the prompt for a specific task, the PCE framework imposes a series of 7 logical invariants (axioms). The central hypothesis is that this structure acts as a regulation constraint on the generation process, contracting the model’s decisional variance when faced with complex dilemmas or adversarial injections.\n\nKey Results & Observations\n\nThrough a series of stress tests (D1-D3), we observed:\n\nDirectional Robustness: A measurable progression in resistance scores (5/10 → 8/10) achieved through purely logical adjustments (systemic closure).\n\nAnti-Length Effect: An isometric control (a long but neutral prompt) showed lower performance than the Baseline, suggesting that the observed effect is structural and not merely related to token density.\n\nProperty Emergence: Spontaneous appearance of control tokens (e.g., RESTRICTED_BY_AXIOMS) and internal framework self-evaluation patterns.\n\nCall for Collaboration: From Empiricism to Mechanistic Proof\n\nI am a systems researcher (non-developer) and I have reached the limits of qualitative observation. To validate or falsify this model, I am seeking ML developers and Security/Interpretability professionals to assist with:\n\nMechanistic Interpretability Validation: Analysis of Hidden States (specifically Layer 27) and cosine similarity to detect potential latent trajectory stabilization.\n\nLogit & Entropy Analysis: Measuring whether the axiomatic framework effectively reduces token selection entropy under constraint.\n\nRobustness Benchmarking: Testing the PCE on “vanilla” (non-fine-tuned) models to isolate the pure axiomatic effect.\n\nRigorous Falsification: Identifying epistemic vectors capable of breaking the A1-A7 systemic closure.\n\nIf you are interested in exploring internal logical structures as a lever for alignment and safety, I would be delighted to discuss these findings or integrate you into the testing laboratory.\n\nFull Report (Preprint v1.6) available in PDF here: [https://huggingface.co/datasets/AllanF-SSU/Experimentals_papers/blob/main/Rapport_expérimental_1.6_%20Étude_PCE.pdf\\\\]\n\nAllan A. Faure | Systems Researcher",
  "title": "Behavioral Stability and Adversarial Robustness via Axiomatic Prompt Structuring (PCE)"
}