← Selected work

Simulation & evaluation

Arena Intelligence

A fictional real-time simulation combining asynchronous AI reasoning, validated control, and exact replay.

My roleDesigned and built the C++20 simulation engine, Python agent pipeline, and React review workspace.

Released v0.2.0 · October 2026

Explore the workflow

Interactive walkthrough

Trace a recorded decision.

Inspect a recorded run from available evidence through control validation and outcome scoring.

Inspect a source report

Select a stage to inspect its role in the workflow.

Step 4: Outcome. Keep the missed outcome in the record.

Scenario 004 / Recorded multi-agent runEvaluator outcome
Node 0Node 1Recorded outcomeNode 2Forecast · 88%
Node 04%
Node 14%
Node 288%
None4%
  1. Observed88.0s
  2. Cutoff90.0s
  3. Applied103.6s
  4. Horizon120.0s

Post-run result: a missed forecast, preserved for review.

Step 04 / Inspect

Keep the missed outcome in the record.

At 120 seconds, the evaluator recorded node 1 rather than the predicted node 2. The forecast missed despite passing the acceptance checks. This outcome is post-run evaluator data and was not available to the agent at forecast time.

predicted: node 2 / 88%
recorded outcome: node 1
forecast: incorrect
Brier score: 1.6992
status: scored
assignment active through horizon

Recorded scenario 004 · request r3 · observer evidence and evaluator results stay separate.

Recorded synthetic scenario 004. Controls inspect the saved run; no new inference is performed.

Inspect the source recording

Results

  • Exact replay verified across 102 runs and 177,052 canonical states, with zero API calls.
  • 131 Python tests and five frontend tests; CI covers Linux, macOS, frontend checks, and sanitizers.
  • Recorded comparison: 120/144 correct forecasts with analyst/challenger reasoning versus 119/144 for one agent, at roughly twice the latency.
Engineering details

Agent reasoning can arrive after its observations become stale. Arena separates inference from deterministic control and records every accepted decision.

01

Real-time control

A deterministic C++20 engine runs at 10 Hz while Python executes an analyst, evidence checks, and a challenger asynchronously.

02

Validated decisions

Schema, evidence, timing, and assignment checks gate recommendations before atomic application at a tick boundary.

03

Exact replay

The React workspace exposes evidence and timing; SQLite records sessions for replay without new model calls.