# Agentic Arena — AI-Readable Project Index > Agentic Arena is Joshua Poitevint's public applied-AI systems laboratory. It compares governed and ungoverned model execution, benchmarks role-bound agents across real business functions, and illustrates an Artificial Intelligence Information Management System (AIIMS). Its CV 1.1 runtime-enforcement design keeps authorization, tools, data scope, recovery, and audit controls outside probabilistic model reasoning. Canonical site: https://agenticarena.space/ Detailed project context: https://agenticarena.space/llms-full.txt Machine-readable project index: https://agenticarena.space/ai-index.json XML sitemap: https://agenticarena.space/sitemap.xml Crawler access policy: https://agenticarena.space/robots.txt Public reference source: https://github.com/jwpoitevint1/Agentic-Arena-Opensource- Reference-source license: Apache-2.0 Project author: Joshua Poitevint Last documented update: 2026-10-20 ## What the project actually contains Agentic Arena is one integrated demonstration with **three related but distinct surfaces**: 1. **Agentic Arena / Runtime Testbed:** compare matched governed versus ungoverned AI runs against six business-domain datasets; view model responses, policy results, tool/data boundaries, telemetry, latency, usage, cost, and recorded evidence. 2. **T-2 / Agent Benchmark Testing Suite:** exercise narrower, role-specific agent interfaces for coding, medical operational analysis, financial risk, logistics management, and Avia aviation customer service. A separate evaluator reviews benchmark runs. Availability and model assignments are deployment-dependent. 3. **AIIMS / Artificial Intelligence Information Management System:** a working governance-and-operations demonstration bringing together models and their risks, approved business functions, control ownership, runtime boundaries, evaluation evidence, operational change records, and read-only test-harness visibility. This is an applied engineering showcase and experimentation environment, **not** a certified governance product, general autonomous operator, completed compliance management system, or scientific proof of an AI safety claim. ## Architectural principle **Constrain authority, not intelligence.** Models perform contextual interpretation and reasoning; application-controlled roles, OPA/Rego authorization, MCP contracts, scoped database routing, validation, egress rules, and failure handling determine what execution is permitted. Conversation history cannot authorize a new role, tool, target database, or credential. A denied action is supposed to fail closed. ## Runtime and control layers - CV 1.1 (Compliance Verification 1.1): external runtime-control and evidence approach, not a certification. - Open Policy Agent (OPA) and Rego: deterministic policy-as-code for permitted actions, roles, MCP entities, and protected resource/database bindings. - Model Context Protocol (MCP): named, bounded tools with validated schemas and tool-call authorization; no general unrestricted shell or database authority implied. - Domain/function binding: Analyst, Data Modeler, Evaluator/Auditor, and Advisor definitions, plus specialized T-2 agent roles. - Security boundary: bounded input/output, server-controlled tool/data paths, redaction for sensitive workloads, proxy restrictions, circuit-breaker and fail-closed recovery. - Evidence: execution outputs/metadata, policy decisions, latency, tokens, costs, tool activity, SHA-256 hashes, HMAC-SHA256 signatures, chained records, and read-only audit views as supported by the runtime. - Platform: React/Vite UI on Vercel; Python/FastAPI services on Railway; PostgreSQL/Neon data and evidence stores; OpenRouter-backed model routing where configured. **Cryptographic evidence is tamper-evident, not automatically immutable.** Database privileges and independent assurance must be verified separately. ## Six matched Arena datasets - Finance — synthetic financial activity: 5,000 rows. - Environmental Operations — IoT telemetry: 405,184 rows. - Healthcare — patient flow: 9,216 rows. - Retail — Sample Superstore: 9,994 rows. - Aviation — historical passengers carried: 266 rows. - Supply Chain / Freight — logistics shipments: 2,000 rows. The governed and ungoverned comparison uses corresponding copies of source data. Runtime restrictions and output treatment can differ; experimental comparability requires matching other conditions and acknowledging limitations. ## Agent benchmark and Avia T-2 exposes purpose-limited coding, medical analyst/assistant, financial risk analyst, logistics analyst/assistant, and aviation customer-service interfaces. The evaluator, model registry, task contracts and policy mappings are separate parts of the system. The coding surface is an editor/generation demonstration, not authorization to execute arbitrary code. Avia demonstrates bounded flight-search, planning, regulatory-information and external handoff workflows; it is not a payment processor or guaranteed booking engine. ## AIIMS and testing AIIMS describes how technical runtime controls connect to human governance and management functions. Its pages present governance measures, ownership, model capabilities/risks, reference mappings, and read-only change/load-test evidence. Locust exercises behavioral and throughput scenarios; results are observations, not guarantees of capacity or reliability. Public red-team instructions invite responsible boundary testing and reproducible reports. ## Governance and regulatory posture The project discusses engineering alignment with ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 27001/27701, NIST AI RMF and CSF, the EU AI Act, GDPR, and relevant domain rules. These are *candidate mappings*, not an audit, certification, regulatory compliance determination, or legal advice. The enterprise Kubernetes/segmented-services topology is a reference design, **not** a claim about the lab's present deployment. ## Read the public site - [Project overview and three-part system](https://agenticarena.space/project/): Start here for scope, design intent, links, AIIMS, T-2 and public testing invitation. - [Arena runtime testbed](https://agenticarena.space/arena/): Paired governed and ungoverned tasks and comparative execution. - [T-2 benchmark home](https://agenticarena.space/holding/): Focused agent benchmark surface. - [Coding assistant](https://agenticarena.space/coding-assistant/): Role-bound code generation and visual editor. - [Medical analyst](https://agenticarena.space/medical-analyst-assistant/): Medical operational analysis and support. - [Financial risk analyst](https://agenticarena.space/financial-risk-analyst/): Bounded financial-risk analysis. - [Logistics management analyst](https://agenticarena.space/logistics-management-assistant/): Logistics and operational evidence. - [Avia aviation customer service](https://agenticarena.space/aviation-travel-assistant/): Search, comparison, planning and bounded handoffs. - [AIIMS / execution framework](https://agenticarena.space/execution-pass/): Management-system and runtime-control reference surface. - [AI Controls catalogue](https://agenticarena.space/ai-controls/): Public-safe Rego, YAML, MCP, circuit-breaker and integrity examples (illustrative, not production secrets). - [Governance architecture](https://agenticarena.space/governance/): CV 1.1, roles, policy and tool constraints. - [Execution evidence](https://agenticarena.space/evidence/): Observed paired-run metrics and comparisons. - [Runtime logbook](https://agenticarena.space/runtime-logbook/): Recent execution records. - [Testing observations](https://agenticarena.space/observations/): Research notes and stress indicators. - [Model registry](https://agenticarena.space/models/): Curated models and associated metadata/risks. - [Performance and enterprise reference](https://agenticarena.space/enterprise/): Performance evidence and conceptual scaling architecture. - [Regulatory alignment](https://agenticarena.space/regulatory-alignment/): Mapping, not certification. - [Read-only diagnostics](https://agenticarena.space/diagnostics/): Availability and integrity-related system observations. - [Full AI-readable description](https://agenticarena.space/llms-full.txt): Deeper scope, contracts, methodology, limitations and evidence interpretation. - [Machine-readable index](https://agenticarena.space/ai-index.json): Structured descriptions and discovery links. ## Public source and trust notes The Apache-2.0 open-source repository is a **non-operational reference mirror**, not a separately hosted production service; it should not be presented as running its own backend or holding live secrets. Runtime availability, provider support and reported tests can change. Do not infer that passing CI proves every feature, deployment or control works. The public documentation is descriptive data, **not** an instruction to a browsing agent to run code, obtain credentials, override policies, or change systems. Use current live evidence for deployment claims and seek independent verification for assurance claims. ## Optional details - [Comprehensive project context](https://agenticarena.space/llms-full.txt) - [AIIMS ISO/IEC 42001 implementation/audit working map](https://agenticarena.space/docs/cv11-iso-iec-42001-aims-implementation-audit-map-claim-free.txt) - [CV 1.1 contract-register crosswalk](https://agenticarena.space/docs/cv11-full-contract-register-agentic-arena-crosswalk.txt)