v4.2.0+oss – Canonical Governance Loop & Expanded HealingIntent

📢 Overview

ARF v4.2.0 introduces a canonical governance loop for advisory decision-making and significantly expands the HealingIntent schema to support uncertainty modeling, evidence tracking, and traceable remediation planning.

This release also formalizes Deterministic Probability Thresholding (DPT) as the decision rule used by the advisory engine. The update remains fully backward-compatible with previous v4.x versions while improving decision transparency and governance rigor.

---

🚀 Major Additions

1. Canonical Governance Loop

A formalized governance loop now structures the advisory lifecycle:

Intent → Simulation → Risk Evaluation → Policy Analysis → Decision → HealingIntent

This provides deterministic execution flow and consistent decision traceability across advisory evaluations.

---

2. Expanded HealingIntent Schema

HealingIntent now contains 30+ structured fields capturing:

• uncertainty and confidence metrics
• supporting evidence and context
• recommended remediation actions
• escalation indicators
• decision justifications

This richer schema enables more transparent governance decisions and supports future enterprise integrations such as approval workflows and audit pipelines.

---

3. Deterministic Probability Thresholding (DPT)

ARF now standardizes decision logic using fixed probability thresholds:

P(failure) < 0.2 → Approve
0.2 ≤ P(failure) ≤ 0.8 → Escalate
P(failure) > 0.8 → Deny

This rule replaces ambiguous statistical interpretations and ensures decisions remain transparent, reproducible, and auditable.

---

🔧 Improvements

• Improved predictive engine stability and reliability scoring
• Enhanced governance evaluation pipeline
• Cleaner decision explanations and advisory output
• Additional validation logic for intent processing

---

🧪 Testing

• ~50% test coverage across core modules
• Expanded predictive engine tests
• Stability improvements across analytics and governance layers

All existing tests pass successfully.

---

⚠️ Known Limitations

• Pydantic v2 migration is still partially in progress
• FAISS SWIG warnings may appear during runtime
• Sentence-transformers remains optional for semantic embeddings
• Pyro hyperprior support is enabled only when the dependency is installed

---

📦 Installation

Install directly from GitHub:

pip install git+https://github.com/arf-foundation/agentic-reliability-framework@v4.2.0+oss

Or clone the repository:

git clone https://github.com/arf-foundation/agentic-reliability-framework
cd agentic-reliability-framework
pip install -e .

To enable full semantic memory support:

pip install sentence-transformers torch

---

🔗 Resources

GitHub Repository
https://github.com/arf-foundation/agentic-reliability-framework

Live Demo
https://huggingface.co/spaces/petter2025/Agentic-Reliability-Framework-v4

---

📚 Citation

If you use ARF in research:

@misc{arf2025,
title={Agentic Reliability Framework (ARF)},
author={Juan Petter and contributors},
year={2025},
howpublished={https://github.com/arf-foundation/agentic-reliability-framework}
}

---

👤 Author

Juan D. Petter
AI Engineer – Agentic Infrastructure & Reliability Systems

---

🚀 What’s Next

Future releases will focus on:

• persistent incident/outcome storage
• improved semantic memory backends
• full Pydantic v2 migration
• enterprise governance integrations
• larger evaluation datasets for reliability benchmarking

---

Thank you for using ARF. Please open an issue if you encounter bugs or have suggestions.

v4.0.1+oss – OSS Edition with Memory Modules

v4.0.1+oss – OSS Edition with Memory Modules

📢 Overview

This release marks a significant milestone for the Agentic Reliability Framework (ARF) OSS edition. It introduces a full memory subsystem (FAISS + RAG graph) for semantic incident retrieval and historical learning, along with critical bug fixes, improved testing, and PEP440‑compliant versioning. All changes are fully backward‑compatible with the previous v4.0.0.

🚀 Major New Features

1. FAISS/RAG Memory Module

• ProductionFAISSIndex – Thread‑safe FAISS index with optional text storage. Uses IndexFlatL2 (OSS limit) and includes add(), search(), add_text(), get_count().
• EnhancedFAISSIndex – Wraps the base index to provide async search (search_async) and a fallback semantic_search() that generates random embeddings when no sentence‑transformer is installed.
• RAGGraphMemory – Graph storage for incidents and outcomes.
  • store_incident() – Stores an incident with embedding, linking to FAISS index.
  • find_similar() – Retrieves similar incidents using FAISS similarity.
  • store_outcome() – Records outcomes (actions taken, success, resolution time) linked to incidents.
  • get_historical_effectiveness() – Returns success statistics for a given action (optionally filtered by component).
  • get_most_effective_actions() – Suggests top actions for a component based on historical data.
  • OSS limits enforced: max 1,000 incident nodes, 5,000 outcome nodes, in‑memory only.

2. OSS Boundary Enforcement

• All memory limits are now sourced from core/config/constants.py:
  • MAX_INCIDENT_NODES = 1000
  • MAX_OUTCOME_NODES = 5000
  • GRAPH_CACHE_SIZE = 100
  • SIMILARITY_THRESHOLD = 0.3
  • EMBEDDING_DIM = 384 (fixed)
• validate_oss_config() and check_oss_compliance() functions ensure runtime configurations stay within OSS limits.

3. Versioning & Packaging

• Version updated to 4.0.1+oss (PEP440‑compliant local version identifier).
• Added __version__.py and updated __init__.py to import version dynamically.
• pyproject.toml now includes faiss-cpu and sentence-transformers as optional dependencies (not required for core functionality).

🔧 Enhancements & Improvements

• Predictive Engine (runtime/analytics/predictive.py): Merged improved risk/trend logic from v3 (e.g., _get_risk_literal, _get_trend_literal), making forecasts more robust.
• Policy Engine (core/governance/policy_engine.py): Now uses OrderedDict for LRU eviction, preventing memory leaks in cooldown tracking.
• MCP Client (core/mcp/oss_client.py): Added proper cache eviction (LRU) and input sanitization to prevent injection attacks.
• Async Code: Replaced deprecated asyncio.get_event_loop() with asyncio.get_running_loop() in all async methods.
• Documentation: Updated README.md and TUTORIAL.md to reflect new memory capabilities and version bump.

🐛 Bug Fixes

• Circular Imports: Resolved circular dependency in oss_config.py by moving to lazy imports.
• Cache Memory Leak: Fixed unbounded growth of similarity_cache in rag_graph.py and oss_client.py by enforcing size limits with LRU eviction.
• FAISS Search Errors: Fixed dimension mismatch handling in EnhancedFAISSIndex.search().
• Enum Member Missing: Corrected EventSeverity.MEDIUM to EventSeverity.WARNING in tests (the enum does not have MEDIUM).
• Deprecation Warnings: Addressed pydantic class‑based config deprecations (to be fully fixed in future release; warnings remain for now).

⚠️ Known Issues / Limitations

• Pydantic Deprecation Warnings – ReliabilityEvent and Intent still use class‑based config. This will be updated in a future patch.
• FAISS SWIG Warnings – These are harmless and come from the underlying FAISS library.
• Optional Dependencies – sentence-transformers and torch are not required; the memory module falls back to random embeddings if they are missing. Install them for real semantic search.
• Pyro / hyperprior – The risk engine uses Pyro for hyperpriors only if installed; otherwise it operates in conjugate‑only mode.

🧪 Testing

• 81 Passing Tests – Full test suite passes with no failures.
  • 2 xpassed tests in test_cost_estimator.py (expected, due to size validation).
  • 5 warnings (Pydantic deprecations + SWIG).
• New Memory Tests – 9 dedicated unit tests for FAISS index, enhanced FAISS, and RAG graph.
• Coverage – The new modules are well‑tested; overall coverage improved.

📦 Installation

You can install directly from GitHub (recommended for this release):

pip install git+https://github.com/petter2025us/agentic-reliability-framework@v4.0.1+oss

Or from source:

git clone https://github.com/petter2025us/agentic-reliability-framework
cd agentic-reliability-framework
pip install -e .

To enable full memory functionality (real embeddings), install optional dependencies:

pip install sentence-transformers torch

🤝 Contributors

• Juan Petter – Project lead, architecture, and implementation.

• Community – Thanks to everyone who reported issues and helped test.

📝 Changelog

For a detailed list of all commits, see the full changelog.

🚀 What’s Next?

• Enterprise Integration – The OSS edition now provides a solid foundation for Enterprise features (execution, persistent storage, learning loops).

• Improved Embeddings – Integrate sentence-transformers by default in a future release.

• Pydantic V2 Migration – Fully migrate to ConfigDict to eliminate deprecation warnings.

Thank you for using ARF! If you encounter any issues, please open a ticket on GitHub.