Restore four-model latency chart with updated TypeSafe percentiles

README.md

@@ -70,10 +70,11 @@ The 4B model leads this panel's Browser subset, while 9B's gains come from princ
 
 *Figure 3. 9B · effort: high · vLLM HTTP API full-response latency · one RTX PRO 6000 Blackwell 96GB GPU. The supplied run snapshot reports 333/333 successful requests and 0% errors. This is full-response latency, not streaming first-token latency.*
 
-| vLLM HTTP model | P50 | P95 | P99 | Source |
+| Serving path | P50 | P95 | P99 | Source |
 | --- | ---: | ---: | ---: | --- |
 | **APUS-OpenJev 9B · high** | **25.58 ms** | **222.36 ms** | **275.58 ms** | Supplied run snapshot |
 | **APUS-OpenJev 4B** | **18.00 ms** | **100.06 ms** | **154.32 ms** | Reported P50; ratio-derived P95/P99 |
+| **Jev / TypeSafe official API** | **280.80 ms** | **376.56 ms** | **830.45 ms** | Supplied pooled 240-request results |
 
 The 4B P50 uses the updated reported result of **18 ms**. Its P95/P99 apply the supplied cross-session ratios of **45% / 56%** to the corresponding 9B high percentiles, rounded to two decimals. The 9B snapshot reports **333/333 successful requests and 0% errors**; these request counts are not attributed to 4B.
 
@@ -90,7 +91,7 @@ The optimized path is the candidate-scoring `/decide` service. The standard Chat
 | Round 3 | 305.57 ms | 65/80 · 81.25% |
 | **Overall** | **302.65 ms** | **199/240 · 82.92%** |
 
-These supplied screenshot results report the arithmetic mean, not P50. Updated TypeSafe percentiles were not supplied. The earlier task-accuracy charts retain their original single-run results; this table records the separate three-round evaluation.
+The updated screenshot reports **P50 280.80 ms, P95 376.56 ms and P99 830.45 ms**, computed from all **240 HTTP 200 responses pooled across three rounds**, rather than averaging per-round percentiles. Mean response time remains **302.65 ms**. Timing includes the public-network overhead from Runpod to the official service; 80 questions repeated three times provide a sample-specific tail-latency reference. The earlier task-accuracy charts retain their original single-run results; this table records the separate three-round evaluation.
 
 **Network reference from the same Runpod host.** The API evaluation used **240 consecutive calls** across three rounds. Separate network probes reported:
 

README.zh-CN.md

@@ -70,10 +70,11 @@ tags:
 
 ![9B vLLM决策接口：P50 25.58 ms、P95 222.36 ms、P99 275.58 ms。](assets/latency-vllm.svg)
 
-| vLLM HTTP模型 | P50 | P95 | P99 | 来源 |
+| 服务路径 | P50 | P95 | P99 | 来源 |
 | --- | ---: | ---: | ---: | --- |
 | **APUS-OpenJev 9B · high** | **25.58 ms** | **222.36 ms** | **275.58 ms** | 提供的运行截图 |
 | **APUS-OpenJev 4B** | **18.00 ms** | **100.06 ms** | **154.32 ms** | P50采用最新提供值；P95/P99按比例换算 |
+| **Jev / TypeSafe 官方 API** | **280.80 ms** | **376.56 ms** | **830.45 ms** | 提供的240次请求合并统计 |
 
 4B P50采用最新提供的 **18 ms**；P95/P99分别按9B high对应值的 **45% / 56%** 计算，保留两位小数。9B截图中的成功请求为 **333/333、错误率0%**，不将该请求数套用到4B。
 
@@ -90,7 +91,7 @@ tags:
 | 第 3 轮 | 305.57 ms | 65/80 · 81.25% |
 | **三轮总体** | **302.65 ms** | **199/240 · 82.92%** |
 
-以上采用提供截图中的算术平均耗时，不是 P50；截图未提供更新的 TypeSafe 分位数。前面的分任务准确率图保留原单轮成绩，此表单独记录三轮复测。
+最新截图提供 **P50 280.80 ms、P95 376.56 ms、P99 830.45 ms**，由三轮全部 **240 条 HTTP 200 响应记录合并后重新计算**，不是各轮分位数取平均。平均耗时仍为 **302.65 ms**，包含 Runpod 到官方服务的网络开销；80题重复三次的 P99 仅作为该样本的尾延迟参考。前面的分任务准确率图保留原单轮成绩，此表单独记录三轮复测。
 
 **同一台 Runpod 的网络参考。** API 评测采用三轮、共 **240 次连续调用**。另行网络探测结果如下：
 

TECHNICAL_REPORT.md

@@ -198,6 +198,7 @@ The APUS serving reference uses **one RTX PRO 6000 Blackwell 96GB GPU**. The 9B
 |---|---:|---:|---:|
 | APUS-OpenJev 9B high · vLLM HTTP | 25.58 ms | 222.36 ms | 275.58 ms |
 | APUS-OpenJev 4B · vLLM HTTP | 18.00 ms | 100.06 ms | 154.32 ms |
+| Jev / TypeSafe official API · public HTTP | 280.80 ms | 376.56 ms | 830.45 ms |
 | Laya typed · complete-input local pipeline | 8.68 ms | 37.66 ms | — |
 
 The 9B-5949 high snapshot reports **333/333 successful requests and 0% errors**. An independently archived three-round serial HTTP reference reports median per-round P50/P95/P99 of **25.34 / 222.83 / 276.78 ms**, with **3,120/3,120** successful requests. These two run records remain separate.
@@ -215,7 +216,7 @@ Jev includes public-network, queueing, and service overhead. Laya includes token
 | Round 3 | 305.57 ms | 65/80 · 81.25% |
 | **Overall** | **302.65 ms** | **199/240 · 82.92%** |
 
-These supplied screenshot results report the arithmetic mean, not P50. Updated TypeSafe percentiles were not supplied. The earlier task-accuracy charts retain their original single-run results; this table records the separate three-round evaluation.
+The updated screenshot reports **P50 280.80 ms, P95 376.56 ms and P99 830.45 ms**, computed from all **240 HTTP 200 responses pooled across three rounds**, rather than averaging per-round percentiles. Mean response time remains **302.65 ms**. Timing includes the public-network overhead from Runpod to the official service; 80 questions repeated three times provide a sample-specific tail-latency reference. The earlier task-accuracy charts retain their original single-run results; this table records the separate three-round evaluation.
 
 **Network reference from the same Runpod host.** The API evaluation used **240 consecutive calls** across three rounds. Separate network probes reported:
 

TECHNICAL_REPORT.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2a0cd1e5fe4b2a496820437886660dff564fb9209774a1bafb92df8a38e86ee4
-size 964744
+oid sha256:0c9845082fe7a78e5fc92979e830ce74a63fbcb33f42a5452048eaeea1f80e6a
+size 959961

bundle-manifest.json

@@ -583,23 +583,23 @@
     },
     {
       "path": "README.md",
-      "bytes": 18318,
-      "sha256": "d76dbe15deee7992b1aa77feb67ee370f2c9a28fdbce23dc9fec60ddc56e4d14"
+      "bytes": 18726,
+      "sha256": "1ed50a42e5899615fea3d22f2447858d71a39c516ac863c0818cd5795882763a"
     },
     {
       "path": "README.zh-CN.md",
-      "bytes": 13684,
-      "sha256": "1a97275ae1552ad13b0224ef00676b5401b12357e415d397d49d2ad4022bf503"
+      "bytes": 14012,
+      "sha256": "8fc192b2d8d46b50be17cb985d01ab3c742c41228638126c8e6827f8dd158298"
     },
     {
       "path": "TECHNICAL_REPORT.md",
-      "bytes": 28037,
-      "sha256": "474867a3f7ebefbc6db8b77785aa7b37dc0881245c1b0241c8186b61befec995"
+      "bytes": 28409,
+      "sha256": "df3415c74deb51038fef662f8658c499d5d07f097d30efd7f1c8dd93ab664cbf"
     },
     {
       "path": "TECHNICAL_REPORT.pdf",
-      "bytes": 964744,
-      "sha256": "2a0cd1e5fe4b2a496820437886660dff564fb9209774a1bafb92df8a38e86ee4"
+      "bytes": 959961,
+      "sha256": "0c9845082fe7a78e5fc92979e830ce74a63fbcb33f42a5452048eaeea1f80e6a"
     },
     {
       "path": "assets/accuracy.svg",
@@ -623,8 +623,8 @@
     },
     {
       "path": "assets/latency-vllm.svg",
-      "bytes": 9526,
-      "sha256": "a96b1adeceb7a21b52a651a19f5da72fabcf14f7a8ca2338b141f3eaf8adc9c2"
+      "bytes": 8311,
+      "sha256": "984b50eb9745d7b02cdc0625a071713c88a5461405905cab0c6282657609bdca"
     },
     {
       "path": "assets/latency.svg",
@@ -658,8 +658,8 @@
     },
     {
       "path": "latency-vllm-data.json",
-      "bytes": 3144,
-      "sha256": "12107b27f37e8f2bd3c3cd124ba653d6a66cc4e476ab3a1d1115014a337ea0aa"
+      "bytes": 3456,
+      "sha256": "3165a4510d2d294934b0374d1f6e9a53303f5fa734daf662eb8a3aa69581fab0"
     },
     {
       "path": "subset-chart-data.json",

latency-vllm-data.json

@@ -19,7 +19,7 @@
   },
   "typesafe_three_rounds": {
     "source": "User-provided TypeSafe three-round screenshot, 2026-09-21",
-    "timing": "Official public API HTTP response; arithmetic mean, not P50",
+    "timing": "Official public API full HTTP response; pooled percentiles across all 240 requests",
     "raw_requests_independently_verified": false,
     "rounds": [
       {
@@ -45,9 +45,11 @@
     "overall_correct": 199,
     "overall_requests": 240,
     "overall_accuracy_percent": 82.92,
-    "p50_ms": null,
-    "p95_ms": null,
-    "p99_ms": null,
+    "p50_ms": 280.8,
+    "p95_ms": 376.56,
+    "p99_ms": 830.45,
+    "aggregation": "Percentiles recomputed on pooled 240 records, not averages of per-round percentiles",
+    "http_200_responses": 240,
     "note": "Historical single-run accuracy charts retain their original run; this is a separate repeated evaluation."
   },
   "provided_snapshot": {
@@ -96,6 +98,12 @@
     }
   },
   "historical_rows": [
+    {
+      "name": "Jev official API",
+      "scope": "TypeSafe HTTP / 240 pooled requests",
+      "p50_ms": 280.8,
+      "p95_ms": 376.56
+    },
     {
       "name": "Laya typed / full input",
       "scope": "Local system_one / three-run median",