Live TelemetryMethodology v1
Live Model Evaluation Telemetry
Realtime multi-stage verification cockpit auditing model-generated game builds in an isolated single-file browser runtime.
Realtime Benchmark Cockpit
COMPLETED5-stage automated telemetry, browser-engine verification & rubric scoring
Progress100%
Stage 1
Static & Architecture
Rubric: Architecture
Stage 2
Physics Engine & Loop
Rubric: Fire Physics
Stage 3
Weapons & Zombie AI
Rubric: Swarm AI
Stage 4
Audio & FX Synthesis
Rubric: Audio/FX
Stage 5
Prompt Fidelity & Synthesis
Rubric: Prompt Fidelity
Candidate A
Model Alpha
Anonymous candidate build • Local single-file sandbox
92.4
Score / 100
Rubric Subscores
Architecture98/100
Fire Physics94/100
Swarm AI96/100
Audio/FX86/100
Prompt Fidelity94/100
Verified Signatures
- ✓Instant wave responsiveness & frame pacing
- ✓Tight cone flamethrower physics
- ✓High-score localStorage persistence
- ✓Clean DoT burn particle pooling
Candidate B
Model Beta
Anonymous candidate build • Local single-file sandbox
88.6
Score / 100
Rubric Subscores
Architecture95/100
Fire Physics88/100
Swarm AI89/100
Audio/FX96/100
Prompt Fidelity85/100
Verified Signatures
- ✓3-weapon tactical arsenal (Flamethrower, Fireball, Napalm)
- ✓Supernova screen-clearing ultimate mechanic
- ✓Custom multi-node procedural audio effects
- ✓Varied zombie state machine behaviors
Telemetry & Verification Stream
7 events logged
[10:00:00]infoBDX Eval benchmark initialized in strict double-blind mode.
[10:00:00]metric[Model Alpha] Static analysis pass: 98/100 (23.4 KB, 0 external deps).
[10:00:00]metric[Model Beta] Static analysis pass: 95/100 (31.8 KB, 0 external deps).
[10:00:01]metric[Model Alpha] Physics check: 60 FPS fixed delta accumulator verified.
[10:00:01]metric[Model Beta] Weapons: 3-weapon arsenal & Supernova charge validated.
[10:00:02]metric[STAGE 4] Audio & particle FX: Web Audio synth nodes active.
[10:00:02]successDouble-blind evaluation complete. Model Alpha leads with 92.4 vs Model Beta at 88.6.
Strict Double-Blind Execution
Prompts and candidate submissions are evaluated with stripped metadata. Candidate identifiers (“Model Alpha” and “Model Beta”) prevent brand familiarity bias across human evaluation and automated test stages. Reviewers inspect performance without prior knowledge of model lineage.
Zero Cross-Contamination
Every build executes in an unprivileged, sandboxed HTML5 Canvas container with zero external dependencies, CDN links, or persistent cross-origin cookies. Evaluation scripts analyze static AST structure, memory allocations, and 60 FPS fixed delta-time stability independently.
Multi-Stage Verification Rubric
Evaluations synthesize five explicit dimensions: Architecture & Dep Isolation (15%), Physics Loop & Collision Math (20%), Weapon Balance & Swarm AI (25%), Audio Synthesis & Particle Shaders (20%), and Prompt Fidelity (20%).
Looking for full benchmark scoring rules?
Read how Showdown Scores are calculated, how community votes factor in, and why scores move between rounds.