docs: lead model card with results and recipe

README.md

@@ -26,6 +26,41 @@ tags:
 
 # JSpark3 v1
 
+**Three DGX Sparks, one OpenAI-compatible GLM-5.3 Flash endpoint, with the
+recipe and evidence to rebuild it.**
+
+JSpark3 is the measured three-Spark path for the EXL3/TR3 checkpoint in this
+repository. It combines the exact mirrored weights with a pinned serving
+stack and a fail-closed fleet controller.
+
+## Results
+
+| Measured result | JSpark3 v1 |
+|---|---:|
+| Paired single-stream code decode | **66.257 tok/s**, +7.27% |
+| Paired single-stream structured decode | **81.962 tok/s**, +6.63% |
+| Agent demonstration aggregate decode | **44.583 tok/s** |
+| Configured context | **1,000,000 tokens** |
+
+The paired rows compare the same recipe on the same three-Spark fleet with
+the trunk overlay disabled, an unreleased internal control rather than a
+market comparison. The agent run is product evidence, not a controlled
+comparison. See the
+[machine-readable results](https://github.com/jakejharris/jspark3/blob/main/results/results.json)
+and [benchmark methods and caveats](https://github.com/jakejharris/jspark3/blob/main/docs/BENCHMARKS.md).
+
+## What you get
+
+- An exact, hash-verifiable mirror of Brandon M. Music's
+  [ShapleyMcg](https://github.com/brandonmmusic-max/shapleymcg) EXL3/TR3
+  quantization, as re-hosted by Mia-AiLab, with its provenance intact.
+- A TP3 and EP3 serving recipe with DFlash2, FP8 KV cache, prefix caching, and
+  a selective W8A16 trunk overlay.
+- A fail-closed preflight, start, health, and verification path.
+- Public measurements with the regressions and missed gates left in.
+
+**[Install and run it with the GitHub recipe](https://github.com/jakejharris/jspark3#quick-start).**
+
 > **License and weight provenance:** JSpark3's original recipe code and
 > documentation are Apache-2.0. The weights here are Brandon M. Music's exact
 > EXL3/TR3 quantization, re-hosted byte-for-byte by Mia-AiLab and mirrored by