- 01Game inputmovement · combat · flow
- 02TelemetryServiceruntime observations
- 03State encodingepoch-scoped state
- 04PredictiveAIMarkov next-state model
- 05Directorgame-owned predictions
- 06Evaluation adapterisolated scoring hooks
- 07Evidence outputsF3 overlay · immutable ledger
Engineering case studySpaceLumin
Verified playable buildPlayer-state prediction inside a running browser game.
SpaceLumin / Lumin Flow: Utopia Run is a Three.js game and a systems testbed for runtime telemetry, Markov next-state prediction, adaptive gameplay architecture, and leakage-safe evaluation.
01 / Scope
System and contribution
A playable project with inspectable implementation evidence.
The engineering question: can a game observe meaningful player state, predict the next encoded state, and evaluate that prediction without contaminating gameplay?
I designed and implemented the gameplay systems, telemetry pipeline, Markov prediction model, runtime overlay, leakage-safe evaluation system, deterministic tests, and supporting PWA tooling.
02 / Runtime architecture
Prediction ownership
Encode behaviour
Runtime signals are reduced to discrete states so transition counts can be updated and queried without blocking the game loop.
Predict next state
The Markov model estimates the most likely next encoded state from observed transitions. It does not claim to predict a human mistake.
Deduplicate frames
A state epoch prevents repeated render frames from counting the same prediction opportunity multiple times.
Prevent leakage
Predictions are frozen before the outcome is observed; the evaluator receives completed events through an adapter.
Separate modes
Frozen-model and prequential learning modes remain distinct so their measurements stay interpretable.
03 / Evaluation design
Evidence without leakage
Frozen evaluation
Scores a fixed model so outcomes can be interpreted without online training changing the evaluated transition counts.
Prequential evaluation
Predicts first, observes the completed transition, scores it, and only then updates the online model.
Verified repository evidence
Evaluator architecture and runtime proof
Claims below describe the current implementation, not future results.
Deterministic coverage of event resolution, deduplication, model isolation, and exception-safe instrumentation.
Completed prediction events remain inspectable after scoring.
Scoring does not replace the prediction selected by gameplay.
Instrumentation failure is isolated from the running game.
Repeated frames do not create duplicate prediction opportunities.
Prediction and telemetry state can be inspected during play.
In evaluation
No final accuracy claim yet
Controlled human-session data collection has not been completed. Accuracy, baseline comparison, latency distribution, calibration, aggregate export, and confidence intervals therefore remain pending.
- Accuracy
- Pending sessions
- Baseline
- Planned comparison
- Latency
- Distribution pending
- Confidence
- Intervals pending
04 / Engineering constraints
Known limitations
KernelMapper, NeuroFlowController, and CognitiveOverloadDetector advanced update paths are currently inactive or unreachable.
Director adaptive multipliers are calculated but are not yet consumed by gameplay systems.
A documented Director ownership/reference integration defect remains unresolved.
No claim is made for statistically superior prediction or validated frame rate across all mobile devices.
SpaceLumin is a working engineering system and an active evaluation project. The repository separates what runs today from what still needs evidence.
Inspect the implementation