2 October 2026 · Agentic game · 3D baseball
Hardball 3D · Sushi 1.1.0
Completed a three-inning 3D arcade baseball game with batting, pitching, fielding, scoring, and restart behavior. The visual presentation is acceptable, but gameplay does not consistently follow baseball rules.
What was checked
- Final Vite production build completed; the generated JavaScript bundle was reported as 549 KB across 13 modules.
- Headless autopilot simulation completed three full games, reached game over, and confirmed restart reset the game to ready.
- Focused runner-on-third scenario recorded the runner scoring at 4.17 simulated seconds.
- DOM-stub smoke check advanced a game with the real HUD and CameraDirector for 3,600 frames without exceptions.
- Development server checks returned HTTP 200 for the page, application module, and Three.js.
Limits of this evidence
The initial session used provider ID mlx-sushi and recorded four failed requests with zero token usage. The successful session used provider ID sushi. The matching engine log identifies Sushi 1.1.0 with MLX 0.32.3, EXL3 K=2.625 bpw, and a 131,072-token context. Its 130 response summaries provide the per-request records archived here. Exact server options and output limits were not retained. Request-start-to-first-token latency is not logged; peak system memory was not sampled. The build and recorded smoke checks completed, but those checks do not validate baseball rule fidelity. The original prompt, session transcript, server logs and local paths are excluded from the archive.
Configuration
The model, serving engine and agent used for this execution.
| Field | Value | Detail |
|---|---|---|
| Model · primary | Qwen 3.8 Flash Next · primary | EXL3 K=2.625 bpw (provider ID label: 2.6 bpw) |
| Runtime | Sushi 1.1.0 · MLX 0.32.3 | inference engine and exact build |
| Harness | Pi coding agent | the agent or client that drove the run |
| Hardware | Apple M5 Pro · 64 GB unified memory | hardware recorded for this execution |
| Workload | Hardball 3D · revision 1 | Interactive three-inning 3D baseball coding challenge; not a fixed task-suite score. |
| Context window | 131,072 tokens | maximum context configured for the serving runtime |
| Repetitions | 1 | executions recorded for this workload; one means an indicative result, not an estimate |
Inference measurements
Engine-side rates for this run. Each figure states its own statistic; figures from different runs are not interchangeable.
| Metric | Value | Definition |
|---|---|---|
| Decode · per-request median | 36.6 tok/s | median · n=130 |
| Decode · output-weighted | 33.9457 tok/s | total output tokens / summed per-request decode time · n=130 |
| Prefill · median | 413.6 tok/s | median; includes cache reuse · n=130 |
| Prefill · mean | 407.33 tok/s | mean; includes cache reuse · n=130 |
| MTP acceptance | 72.9341 % | accepted draft tokens / drafted tokens · n=75571 · 55,117 accepted out of 75,571 drafted tokens across 130 matched requests. |
Agent session counters
Agent-wide usage totals. These include reasoning, tool calls and waiting, and are not model decode speed.
| Metric | Value | Definition |
|---|---|---|
| Session elapsed | 90.6 min | first logged provider attempt to final response across both Pi sessions · n=1 |
| Successful session elapsed | 90.0 min | successful Sushi session start to final response; includes coding and checks · n=1 |
| Model responses | 130 responses | assistant responses recorded in the successful Sushi session · n=1 |
| Tool calls | 133 calls | total in the successful Pi session · n=1 |
| Context compactions | 2 compactions | total · n=1 |
| Failed first-session requests | 4 requests | failed first-session requests; zero token usage reported and excluded from the successful session totals · n=1 |
| Matched model requests | 130 requests | one-to-one Pi/Sushi usage match · n=1 |
Context and token counters
Token totals as reported by the harness or engine. Cached and cumulative figures are not additive with fresh input.
| Metric | Value | Definition |
|---|---|---|
| Fresh input | 259,165 tokens | sum of Pi usage counters · n=1 |
| Cached input | 8,079,060 tokens | sum of Pi cache-read counters · n=1 |
| Output tokens | 148,981 tokens | sum of Pi usage counters · n=1 |
Request records & provenance
Matched 130 successful Pi assistant responses one-to-one with Sushi runtime summaries using fresh input, cached input and output token counts. The engine records provide per-request prefill/decode rates and MTP accepted/drafted counters. Neither log records request-start-to-first-token latency; Pi event gaps include tool calls and agent work, so TTFT is unavailable.
Run ID: 2026-10-02-qwen38-flash-next-sushi-hardball-3d
One execution, with checks transcribed from the retained record. A completed session is not a quality score. Measurement method · All runs