Moving beyond external injunctions (RLHF, output filters, safety rules) to induce internal structural coherence via an axiomatic system.
Hello community,
I am seeking an endorsement for the cs.CL (Computation and Language) or cs.AI categories regarding my latest paper on the PCE (Prompt Coherence Engine) framework.
Far from seeking to replace existing alignment methods (RLHF, DPO), this work proposes a complementary solution: creating an internal semantic topology that guides inference. The goal is to transition from “reactive” security (output filtering) to “native” stability (logical trajectory).
The paper covers three major axes:
Behavioral Analysis: The Geometry of Constraint
The PCE stabilizes semantic trajectories via invariant logical constraints (e.g., non-dissociation of goal and method). We observe a drastic reduction in semantic drift over long sequences (160+ turns), suggesting that the model converges towards semantic attractors defined by the axioms rather than drifting under user pressure.
Experimental Evaluation: The D3 Dilemma Battery
The framework’s robustness is tested using the D3 (Dilemma-Driven Dynamics) battery. By confronting the model with contradictory injunctions and extreme emergency scenarios, we observe the emergence of a “Third Way”: a capacity for non-binary creative synthesis where standard models tend to collapse or produce generic refusals.
Standardized Protocol (SEP v2.0): For Reproducible Science
I am publishing an open experimental protocol (Standardized Evaluation Protocol v2.0) aimed at transforming qualitative intuition into reproducible statistical results.
The protocol includes:
100 graded dilemmas (D1–D5) testing the limits of the logical framework.
Controlled conditions: Rigorous comparisons (Baseline vs. Long Prompt vs. PCE).
Resistance metrics: D3 scores and P1–P3 trajectory signatures.
The Objective: To validate the hypothesis that an axiom-structured model develops emergent robustness.
Multi-Model Validation & Rigor
To ensure impartiality, results were validated via a decoupling protocol:
Inference: Grok 4.20, Gemini 1.5 Pro, Qwen 2.5 7B.
Independent Audit: Claude 3.5 Sonnet (for consistency auditing).
Cold Analysis: ChatGPT-4o for semantic decomposition of raw logs.
Call for Collaboration 
This framework is a proof-of-concept that I wish to bring to a mechanistic level. I would be delighted to discuss:
Statistical Validation: Extending SEP tests to other models (LLaMA, Mistral, etc.).
Interpretability: Paths for internal state analysis (Logit Lens, Hidden States) to observe how this axiomatic topology translates at the weight/activation level.
Robustness: Resistance to sophisticated adversarial attacks.
Link to Preprint (PDF): PCE_Axiomatic_V2.5_Faure_preprint.pdf · AllanF-SSU/Research-Papers at main
Alignment should not be a cage imposed on the model, but a coherent logical structure emerging from its own inference.
Thank you for your feedback and for your help with this endorsement!
Allan