Add verified benchmark figures to model card

.gitattributes

@@ -36,3 +36,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 gguf/Winnow-E4B-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/Winnow-E4B-BF16.gguf filter=lfs diff=lfs merge=lfs -text
 gguf/mmproj-Winnow-E4B.gguf filter=lfs diff=lfs merge=lfs -text
+assets/e4b-profiles.png filter=lfs diff=lfs merge=lfs -text
+assets/e4b-quality.png filter=lfs diff=lfs merge=lfs -text
+assets/e4b-speed-memory.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -79,8 +79,16 @@ These figures evaluate the **downloadable Q8 GGUF** against Winnow-12B Q8 on the
 
 Winnow-12B leads on the broader public suites. E4B leads by 2.2 points on this local typed panel, which uses synthetic teacher targets rather than independently verified truth for every item. The JevBench figure is **public-subset accuracy**, not the official composite leaderboard score. Methods, uncertainty intervals, calibration measures, and evaluation boundaries are in the [benchmark report](docs/BENCHMARKS.md).
 
+![Winnow-E4B and Winnow-12B Q8 decision scores on three evaluation panels](assets/e4b-quality.png)
+
+*Decision quality on the same questions. The local typed panel measures agreement with synthetic teacher labels.*
+
 In a separate matched 8K text decision workload, E4B Q8 delivered **147.8 decisions/s** at 64 questions per request, versus **67.2 decisions/s** for Winnow-12B Q8. Those are warm request medians on an RTX 5070 Ti, not a guarantee for other hardware or workloads. The [timing table](docs/BENCHMARKS.md#same-gpu-latency-and-memory) includes smaller batches and memory use.
 
+![Winnow-E4B and Winnow-12B Q8 warm decision throughput, memory, and single-decision latency](assets/e4b-speed-memory.png)
+
+*Matched text-only workload on an RTX 5070 Ti with Q8 weights and KV cache. Timings are medians of ten warm requests.*
+
 ## Probabilities and calibration
 
 The server normalizes candidate logits over the options in each question. The optional `winnow.temperature` setting scales those logits without changing which option wins. For **Q8 text decisions**, `1.2574172017327816` was fitted on 778 separate calibration questions. For **BF16**, the separately fitted value is `1.3331553765162731`; do not reuse Q8's temperature for BF16.
@@ -104,6 +112,10 @@ Download Q8 for the measured decision, chat, and vision setup; add the projector
 
 Capacity suggestions come from one RTX 5070 Ti using Q8 KV cache, four decision branches, one chat slot, and exclusive memory scheduling. They have not been tested on every GPU in those tiers. The Q8 model and projector answered six questions over a 65,023-position state that included a synthetic image; repeated cached requests took 129–131 ms after an 11.59 s cold request. A separate 62,431-position multimodal chat prompt generated 512 tokens at 71.2 output tokens/s. The [benchmark report](docs/BENCHMARKS.md#64k-context-and-vision-smoke-test) describes these operational probes and their limits.
 
+![Winnow-E4B and Winnow-12B Q8 GPU memory and chat generation, plus E4B cached 64K decisions](assets/e4b-profiles.png)
+
+*Q8 operational profiles on an RTX 5070 Ti. The 8K text workload was matched; the 64K vision and chat probes used different images and question counts. The 10 GB+ and 12 GB+ capacity suggestions are inferred from measured memory, not tests on those cards.*
+
 ## Scope and limitations
 
 - The decision path is intended for supplied answer options. Open-ended responses use the regular chat endpoint.