results: add v2 verification run — confirms v1 within measurement variance

README.md

@@ -23,17 +23,21 @@
 ## 📊 Results
 
 Tested on **NVIDIA RTX 3090 (24 GB VRAM)** with **Mistral-Small-3.2-24B Q4_K_M**.
+Results are the average of **two independent benchmark runs** (April 1, 2026).
 
 | | Baseline (f16) | TurboQuant turbo3 | Delta |
 |--|:--------------:|:-----------------:|:-----:|
 | **Context** | 8,192 tokens | **100,000 tokens** | **+12.2×** |
-| **VRAM** | 15.4 GB | 17.2 GB | +1.8 GB only |
-| **Tokens/s** | 49.2 | 45.0 | −8.5% |
+| **VRAM** | 15.5 GB | 17.4 GB | +1.9 GB only |
+| **Tokens/s** | 50.2 | 46.0 | **−8.3%** |
 | **KV-Cache size** | ~1 GB (f16) | ~2.8 GB (3-bit) | **4.3× compression** |
 
-> **12× more context. +11% VRAM. −8% speed. Same model weights.**
+> **12× more context. +12% VRAM. −8% speed. Same model weights.**
 
-Raw data: [`results/turboquant-rtx3090-2026-04-01.json`](results/turboquant-rtx3090-2026-04-01.json)
+**Run 1 (cold):** Baseline 49.2 TPS / 15,408 MB → Turbo3 45.0 TPS / 17,224 MB
+**Run 2 (warm, idle GPU):** Baseline 51.2 TPS / 15,695 MB → Turbo3 47.1 TPS / 17,581 MB
+
+Raw data: [`results/turboquant-rtx3090-2026-04-01.json`](results/turboquant-rtx3090-2026-04-01.json) · [`results/turboquant-rtx3090-2026-04-01-v2.json`](results/turboquant-rtx3090-2026-04-01-v2.json)
 
 ---
 

results/turboquant-rtx3090-2026-04-01-v2.json

@@ -0,0 +1,36 @@
+{
+  "run_id": "v2-verification-2026-04-01",
+  "timestamp": "2026-04-01T00:00:00Z",
+  "hardware": {
+    "gpu": "RTX 3090",
+    "vram_total_mb": 24576
+  },
+  "baseline": {
+    "cache_type": "f16",
+    "ctx": 8192,
+    "port": 8180,
+    "vram_used_mb": 15695,
+    "tps_runs": [51.01, 51.07, 51.44],
+    "tps_avg": 51.17
+  },
+  "turbo3": {
+    "cache_type": "turbo3",
+    "ctx": 100000,
+    "port": 8182,
+    "vram_used_mb": 17581,
+    "tps_runs": [47.34, 46.90, 46.93],
+    "tps_avg": 47.06
+  },
+  "delta": {
+    "context_multiplier": 12.21,
+    "vram_overhead_mb": 1886,
+    "tps_delta_pct": -8.03
+  },
+  "model": "mistralai_Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf",
+  "image": "turboquant-plus:feature",
+  "notes": [
+    "VRAM was nearly idle at start (550 MB used, 23777 MB free) — no unloading needed",
+    "Port 8182 used for turbo3 (8181 occupied by ocr-api-waitress)",
+    "v2 verification confirms v1 results: turbo3 enables 12x context at +1.9 GB VRAM cost and -8% TPS"
+  ]
+}