Repository navigation
Arena Intelligence 0.2.0
Arena is a simulation and evaluation workspace for defense autonomy engineering. This release connects a deterministic C++20 arena, asynchronous Python reasoning agents, and a local React/TypeScript product for reviewing evidence and outcomes.
Included
- An installable local application with recorded sessions, spatial replay, evidence-linked review, and run comparison.
- Authenticated HTTP ingestion, persistent sessions, and a working standard-library Python adapter for synthetic simulators.
- A deterministic C++20 reference environment with filtered observations, atomic objective assignments, deadline recovery, and exact replay.
- An analyst/challenger workflow with structured output, local evidence checks, semantic validation, and a persistent spending ceiling.
- Frozen experiment reports, paired reference recordings, raw experiment records, and reproduction tools.
Downloads
| File | Use |
|---|---|
arena-intelligence-local-0.2.0.zip |
Recommended installation kit; includes the wheel, installer, and adapter examples |
arena_intelligence-0.2.0-py3-none-any.whl |
Python package for an existing environment |
arena-intelligence-source.zip |
Source, documentation, reports, compiled UI, and demo recording |
arena-intelligence-experiments.zip |
Complete frozen experiment records; expands into the source archive's runs/ tree |
arena-intelligence-release.json |
SHA-256 checksums for the four artifacts above |
The first installation downloads Python dependencies. Recorded sessions and external adapters require no API key, frontend build, or native compiler. New native matches require the optional C++ build. Instructions are in docs/install.md.
Verification and scope
The local release checks passed 131 Python tests and five frontend tests. A fresh macOS installation served the bundled interface and both reference sessions, accepted a synthetic adapter stream, and preserved its records after restart. A real paced native run completed 900 ticks and three applied scripted assessments. No additional paid model calls were used for the product redesign.
The recorded matched-input study produced 120/144 correct forecasts for analyst plus challenger and 119/144 for the single agent. That one-forecast difference does not establish a reliable advantage. Full latency, cost, coverage, closed-loop outcomes, and native performance methods are in the engineering report.
This release supports fictional simulation data and local, single-user engineering review. It does not provide operational military analysis or native integrations with named defense or robotics platforms. There is no affiliation with or endorsement by Anduril. GitHub CI passed all four jobs: frontend, Linux native/Python, macOS native/Python, and Linux sanitizers. The product validation report records the local installation and browser checks.