All performance values below come from the production Game Boy-compatible ROM under PyBoy 2.7.0 DMG timing. They are emulator measurements, not physical- hardware measurements.
The ROM increments a volatile frame counter from VBlank. Each generation step records inference frames before the selected token is rendered; render frames are tracked separately. PyBoy DMG timing is converted using 59.7275 frames per second.
| Measurement | Value |
|---|---|
| Recorded production tokens | 71 |
| Average inference | 106.55 frames |
| Average inference | 1.784 seconds/token |
| Worst inference | 113 frames |
| Worst inference | 1.892 seconds/token |
| Average token rendering | 3.87 frames |
The release gate prefers an average below two seconds and requires every recorded token below 2.5 seconds.
| Region | Use |
|---|---|
| ROM | 131,072 bytes / 8 banks |
| Static WRAM | 4,487 bytes |
| Minimum stack headroom | 3,705 bytes |
| Bank 0 used | 14,201 of 15,872 usable bytes |
| Bank 0 headroom | 1,671 bytes |
The Bank 0 hard minimum is 1,024 bytes; 2,048 bytes is a warning preference. The release therefore passes while retaining an advisory warning.
Held-out quantized model evaluation reports 88.14% token accuracy, 86.84% exact match, 96.71% semantic consistency, zero known contradictions, zero malformed outputs, 88.24% held-out route validity, and 93.55% ending-transition validity. These are specialized dataset metrics, not general language benchmarks.
- 2,664 Python/native deterministic engine vectors.
- 98 Python/native integer inference contexts with full logits and attention.
- A complete real-button-input Game Boy route ending at
FLED_AT_DAWN. - All six endings and route predicates covered by host simulation/tests.
- 512 sampled integer prompts in the final mask evaluation.
- Four generated intro and three generated ending segments in ROM verification.
Machine-readable evidence is generated under build/reports/. The compact
aggregate is build/reports/release-report.json.
PyBoy is useful for deterministic regression and relative timing but is not a substitute for an original DMG and cartridge. Automated semantic matching uses known valid reference families and cannot measure atmosphere, surprise, or human preference. The release has not undergone a blinded external user study.