zabinskirafal/AGI: AGI Pragma v3.0.0 — Autonomy, Safety & Decision Integrity

AGI Pragma v3.0.0 — Autonomy, Safety & Decision Integrity What's New in v3.0.0 This release introduces the Decision Integrity Layer — a formal, risk-aware architecture for controlled autonomous decision-making under uncertainty. AGI Pragma focuses on how decisions are constrained before execution, rather than on opaque optimization or reward maximization. Key Features • Decision Integrity Layer A structured decision pipeline that evaluates actions before execution using explicit risk and failure analysis. • FMEA Risk Engine Industrial-grade Failure Mode and Effects Analysis (FMEA) is applied to every candidate action to compute a Risk Priority Number (RPN): RPN = Severity × Occurrence × Detection • Critical Path Analysis Dynamic identification of decision paths that are likely to lead to irreversible failure within a finite horizon. • Cognitive Circuit Breaker An automated safety gate that dynamically constrains autonomy: OK → WARN → SLOW → STOP, halting execution when risk thresholds are exceeded. • Bayesian Belief Updating Probabilistic tracking of recurring hazards (e.g. trap and failure rates) to support adaptive, auditable learning under uncertainty. Benchmarks This version includes a minimal, reproducible benchmark (Snake) that demonstrates: self-harm = loss, pre-action risk filtering, critical path avoidance, FMEA-based action gating, circuit breaker–driven autonomy control. The benchmark produces auditable decision logs and summaries to support reproducibility and analysis. Scope of This Release This release focuses on the Decision Intelligence Core of AGI Pragma: risk-aware decision gating, controlled autonomy mechanisms, reproducible safety benchmarks. It does not include: large-scale training, production agents, environment-specific or physics-heavy implementations. Documentation Comprehensive documentation is available in the /docs directory: Methodology Metrics Safety model Benchmark descriptions Project Status AGI Pragma v3.0.0 is a research release. It provides a stable foundation for further experimentation, benchmarking, and safety-oriented research. Future work will expand benchmarks and evaluation protocols, while keeping the Decision Integrity Layer stable. Citation If you use this work in academic or technical research, please cite it using the provided CITATION.cff file.