Technical papers covering architectures, methods, evaluations, applied AI systems, and implementation detail.
Publications
Whitepapers and strategic briefs from Evening Star AI.
These papers are written for builders, security teams, leaders, and operators who need clear thinking about AI security, anomaly detection, software assurance, automation, and decision support.
The library is organized so the core papers, applied-lab research, partner work, and focused briefs are easy to navigate without blurring their roles.
For the evaluation standard behind the library, see the Evening Star AI methodology.
Formats
Publication formats
Some pieces go deep into architecture and methods. Others are short briefs meant to clarify a problem, frame a decision, or explain why a pattern matters before the work becomes a full technical paper.
Shorter papers for leaders and builders: frameworks, operating principles, strategy, and decision analysis.
Research Collections
Browse by topic.
Publications are grouped by the problems they address: AI security, applied labs, and public-sector AI.
The Evening Star AI Operating Principles
A concise statement of how Evening Star AI should think, build, and operate: observe context, reveal change, reason about impact, and help humans move.
Intended audienceFounders, technical leaders, AI builders, operators, and collaborators evaluating the institute's operating philosophy.
The Evening Star AI Engine
Technical architecture for application-agnostic operational intelligence: baselines, detector ensembles, drift, attribution, confidence, and governed action.
Intended audienceAI engineers, security architects, platform builders, and technical executives evaluating reusable operational-intelligence systems.
Purple Radar: AI-Driven Vulnerability Intelligence
A decision-focused model for vulnerability prioritization, exposure analysis, exploitability signals, risk scoring, and evidence-first remediation planning.
Intended audienceSecurity teams, vulnerability managers, CISOs, cyber operators, and leaders responsible for risk-prioritized remediation.
Purple Firefish: An AI Security Gateway for LLM Applications
A governed gateway pattern for prompt injection, jailbreak pressure, indirect prompt attacks, adversarial inputs, risky tool use, and unsafe model outputs.
Intended audienceAI application teams, security engineers, red teams, GRC leaders, and operators deploying LLM workflows.
Purple Sentinel
The applied lab for AI-Native Software Assurance, focused on turning repository evidence into defensible release judgment for AI-assisted software.
Intended audienceEngineering teams, security reviewers, AI platform teams, and leaders deciding whether AI-assisted software should ship, wait, or be blocked.
Candles Edge: AI Decision Support for Market Signals
A market-intelligence proving ground for turning OHLCV features, volatility expansion, regime shifts, anomaly markers, and decision signals into usable human judgment.
Intended audienceMarket operators, product builders, decision-system designers, and analysts studying noisy, fast-moving environments.
Practical Evals for Agentic Systems
A practical evaluation stack for measuring agent task success, side effects, tool behavior, policy violations, recovery, escalation quality, cost, and traceability.
Intended audienceAI engineers, platform teams, model-risk leaders, security reviewers, and operators responsible for agent deployment.
Fulcrum: A Governance‑First Agentic AI Platform
A two-plane architecture for self-hosted inference, retrieval, memory, tool execution, policy enforcement, approvals, traces, evals, and bounded autonomy.
Intended audienceAI platform teams, security architects, autonomy engineers, technical executives, and operators building agentic systems with real operational risk.
MCP Security and Tool-Space Governance
A security-native model for governing MCP-style tool ecosystems through tool identity, namespaces, least privilege, policy mediation, execution isolation, and verification.
Intended audienceAI platform teams, security architects, MCP adopters, connector owners, and governance stakeholders.
The Operational Intelligence Layer
The category paper for the layer between raw signals and governed action: context ingestion, detection, reasoning, confidence, policy alignment, explanation, and next-step design.
Intended audienceExecutives, platform architects, applied AI teams, security leaders, and operators working in sensitive operational domains.
The Evening Star AI Governance Stack
A proof-oriented architecture for connecting policy intent to evals, runtime controls, audit logs, approval paths, incident response, and system improvement.
Intended audienceCTOs, CISOs, platform leads, staff engineers, and teams building governed AI systems.
Governing AI for Better City Operations
A municipal AI governance framework for accountability, innovation, public trust, procurement discipline, and risk-tiered oversight.
Intended audienceMayors, city managers, CIOs, procurement leaders, legal teams, civil-rights leaders, and municipal executives governing AI adoption.
Governing AI for Better City Operations: Executive Brief
An executive brief for municipal leaders on why AI governance is an enterprise management discipline, not just a technology project.
Intended audienceMayors, city managers, department leaders, and executive teams evaluating municipal AI governance priorities.
The Case for Governed Agentic AI in Municipal Government
A public-sector operating model for governed agentic AI in municipal workflows, service delivery, procurement, finance, records, and resident operations.
Intended audienceCity managers, CIOs, department heads, procurement leaders, and elected officials navigating governed AI-enabled operations.
The Case for Governed Agentic AI in Municipal Government: Executive Brief
An executive brief on how governed agentic AI can reduce administrative friction while preserving human accountability in municipal government.
Intended audienceMunicipal executives, CIOs, department heads, procurement leaders, and elected officials evaluating operational AI pilots.
Red Teaming Agentic AI Systems
A whole-system red-team framework for prompt injection, context poisoning, unsafe tool use, permission bypass, goal drift, data leakage, and cross-agent propagation.
Intended audienceSecurity teams, red teams, AI safety teams, agent platform engineers, and governance reviewers.
No publications match this collection yet.