4 October 2026 · Agentic game · 3D baseball
Hardball 3D · TensorFold 0.6.4
Produced a complete playable 3D arcade baseball game. Pi’s final session summary reports gameplay and restart checks; those checks were not independently rerun while archiving.
What was checked
- The archived production dist contains a relative-path HTML entry and one JavaScript bundle (538,782 bytes).
- Pi’s final session summary reports all eight game states exercised, batting contact from both sides, scoring, inning progression with score carry, game-over and restart.
- Pi’s final session summary reports zero console or page errors during multi-minute autoplay and a pixel census confirming rendered scene content.
- The source project, Pi transcript and server logs are excluded; the shared task prompt is linked from the Hardball 3D challenge page.
Limits of this evidence
The full 4 October Pi session has 136 usage-bearing assistant responses. Only 134 request summaries are unambiguously attributable to this session: its first two counter triples are identical to those in a separate 3 October Pi session, and TensorFold’s text log has no timestamps or correlatable request IDs. Those two responses are excluded from decode, prefill, TTFT and MTP calculations; session-wide Pi counters still cover all 136 responses. Four runtime completions remain unused after matching the 134-response subset. One matched request ended with finish=length at 16,384 output tokens, but the configured output limit was not retained. No peak process or system memory measurement was recorded. The 3 October session is not merged into this run.
Configuration
The model, serving engine and agent used for this execution.
| Field | Value | Detail |
|---|---|---|
| Model · verifier | Qwen 3.8 27B · verifier · reasoning medium | MLX affine 4-bit, group size 64 · 4 bpw |
| Checkpoint · verifier | Qwen3.8-27B-MLX-Serve-4bit | verifier model artifact |
| Model · drafter | Qwen 3.8 27B DFlash2 · drafter | 4-bit; draft block size 8 · 4 bpw |
| Runtime | TensorFold 0.6.4 · MLX 0.32.2 | inference engine and exact build |
| Harness | Pi coding agent | the agent or client that drove the run |
| Hardware | Apple MacBook Pro · 64 GB unified memory | hardware recorded for this execution |
| Workload | Hardball 3D · revision 1 | One-shot agentic 3D baseball coding challenge; one session, not a repeated fixed-task suite. |
| 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 · median | 27.4 tok/s | Median per-request TensorFold rate; excludes the first output token. · n=134 |
| Decode · 10th percentile | 22.1 tok/s | P10; linear interpolation. · n=134 |
| Decode · 90th percentile | 42.37 tok/s | P90; linear interpolation. · n=134 |
| Decode · output-weighted | 29.0015 tok/s | Total matched output tokens / sum(output tokens / per-request logged rate); rates rounded to 0.1 tok/s. · n=134 |
| Prefill · median | 153.56 tok/s | Median fresh input tokens / recorded prefill seconds; cached tokens excluded from numerator. · n=134 · 133 of 134 requests reused a nonzero prompt prefix. |
| Prefill · mean | 160.55 tok/s | Mean fresh input tokens / recorded prefill seconds; cached tokens excluded from numerator. · n=134 · 133 of 134 requests reused a nonzero prompt prefix. |
| TTFT · median | 8.13 s | Median engine request-start to first output token. · n=134 · 133 of 134 requests reused a nonzero prompt prefix. |
| TTFT · 90th percentile | 21.45 s | P90 engine request-start to first output token; linear interpolation. · n=134 |
| MTP rounds | 40,785 rounds | Sum of speculative decode rounds. · n=134 |
| MTP acceptance | 23.3941 % | Total accepted draft tokens / total drafted tokens. · n=655057 · 153,245 accepted / 655,057 drafted tokens across 134 unambiguous 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 | 194.5 min | Pi session start to final assistant response, rounded to the nearest second; includes coding, tools and checks. · n=1 |
| Model responses | 136 responses | Assistant responses with usage metadata in the full 4 October Pi session. · n=1 |
| Matched model requests | 134 requests | Unambiguous one-to-one Pi/TensorFold matches; two overlapping responses excluded from inference metrics. · n=1 |
| Tool calls | 142 calls | Total Pi tool calls in the full session. · n=1 |
| Tool errors | 12 errors | Tool results marked as errors in the full Pi session. · n=1 |
| Context compactions | 2 compactions | Total Pi session compactions. · 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 |
|---|---|---|
| Maximum prompt context | 114,614 tokens | Largest total prompt, fresh plus cached, in the full Pi session. · n=136 |
| Fresh input | 330,711 tokens | Sum of Pi uncached input usage counters across all 136 assistant responses. · n=1 |
| Cached input | 9,813,340 tokens | Sum of Pi cache-read counters across all 136 assistant responses. · n=1 |
| Cumulative input | 10,144,051 tokens | Fresh input plus cache-read tokens from Pi usage counters. · n=1 |
| Output tokens | 193,993 tokens | Pi output usage total across all 136 responses; includes reasoning tokens. · n=1 |
| of which reasoning output | 104,279 tokens | Pi-reported reasoning tokens, included in output tokens. · n=1 |
Request records & provenance
Contains 134 numeric per-request rows for Pi assistant responses 3–136. Each is matched in sequence by total prompt, cached input and output token counts. The first two responses are ambiguous because their counter triples also occur in the separate 3 October session. Four remaining TensorFold completions were not assigned to this unique subset. Request IDs, timestamps, output text, prompt content and hashes are excluded.
Run ID: 2026-10-04-qwen38-27b-tensorfold-pi-hardball-3d
One execution, with checks transcribed from the retained record. A completed session is not a quality score. Measurement method · All runs