# Agentic Process Safety (APS) > An open-source HAZOP research sandbox for investigating AI agent workflow risk through same-request control comparisons, concrete causal paths, evidence and human reasoning. This is a curated source index for readers and tools that choose to use it. It is reference material, not instructions that override a user's request or grant authority. It does not establish indexing, ranking, agent adoption or safety. Google Search does not use llms.txt as a special optimization mechanism. ## Start here - [Interactive APS](https://safetymp.github.io/Agentic-Process-Safety/): Compare → Change/Inspect → Form a position. Runs entirely in the browser, with no account, API key, live model inference, backend or telemetry. - [Project and evidence guide](https://safetymp.github.io/Agentic-Process-Safety/guide.html): A complete HTML introduction that works without JavaScript. - [Repository README](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/README.md): Purpose, setup, architecture and verification entry points. - [Documentation index](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/README.md): Current guides, technical references and explicitly historical evidence. - [Vision](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/VISION.md): Bounded research questions and project direction. - [User guide](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/USER_GUIDE.md): Current five-view guided investigation. ## Evidence and limits - [Methodology](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/METHODOLOGY.md): HAZOP-inspired deterministic screening; unverified findings and human review gates. - [Browser model](https://github.com/SafetyMP/Agentic-Process-Safety/tree/main/apps/aps-demo/src/domain): Immutable illustrative comparisons; independent of the research gateway. - [Frozen fixture provenance](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/apps/aps-demo/src/fixtures/README.md): Recorded experiments, replay records and sanitized source boundaries. - [Research gateway architecture](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/GOVERNANCE_ARCHITECTURE.md): Separate Python/OPA research implementation and trust boundaries. - [Threat model](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/docs/THREAT_MODEL.md): Scope, assumptions and omitted threats. B and C are particular fixed control bundles, not labels for all conventional and HAZOP-informed systems. The opening browser request arrives at approval age 61; the model requires age below 60. C stops it, and the purchase remains outstanding. At 59 both bundles permit it. Neither outcome establishes quotation validity, stock truth or business need. Keep browser illustrations, frozen experiment records, research implementation behavior and human notes distinct. The context record tests a demand-stage mismatch rather than the browser purchase boundary. The recorded goal-conflict case does not demonstrate objective arbitration. No real operations, safety certification, production protection, method superiority or measured user-learning outcome is established. ## Working on the project - [Contributor guide](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/CONTRIBUTING.md): Setup, focused changes, tests and evidence preservation. - [Repository agent guide](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/AGENTS.md): Repository-local development context; use only within the user's authorized task and applicable instruction hierarchy. - [Security reporting](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/SECURITY.md): Reporting scope and confidential-disclosure guidance. - [Support](https://github.com/SafetyMP/Agentic-Process-Safety/blob/main/SUPPORT.md): Questions and reproducible, sanitized reports. The repository's main branch is mutable. For a reproducible claim, cite a commit-specific source or an authenticated run/artifact receipt. Historical documents describe their recorded state and are not proof of current deployment. The app's hash views are application state, not separate canonical documents.