DECODE p99 = 309 ns ▮ THROUGHPUT ≥ 14,000 msg/s ▮ E2E p99 < 0.7 ms ▮ CRASH RECOVERY 6/6 PASS ▮ ZERO MESSAGE LOSS ▮ REPRODUCIBILITY ≈ 1/5,000 ▮ LLM-AGNOSTIC BY CONSTRUCTION ▮ ED25519 PER-FRAME SIGNING ▮ TRIMMED CoV = 0.04 ▮ 156,460,102 SAMPLES ▮ WAL RECOVERY 65 ms MEDIAN ▮ DECODE p99 = 309 ns ▮ THROUGHPUT ≥ 14,000 msg/s ▮ E2E p99 < 0.7 ms ▮ CRASH RECOVERY 6/6 PASS ▮ ZERO MESSAGE LOSS ▮ REPRODUCIBILITY ≈ 1/5,000 ▮ LLM-AGNOSTIC BY CONSTRUCTION ▮ ED25519 PER-FRAME SIGNING ▮ TRIMMED CoV = 0.04 ▮ 156,460,102 SAMPLES ▮ WAL RECOVERY 65 ms MEDIAN ▮

// SYS_INIT :: A2A_TRANSPORT_LAYER :: VERIFIED METRICS LOADED // 系統初始化 :: A2A 確定性傳輸層 :: 已驗證數據載入完成

Agent-to-Agent
Deterministic
Transport Layer Agent-to-Agent
確定性
傳輸層

A deterministic A2A / agent-to-robot transport data plane for
Physical AI · Robotics Fleets · Autonomous Machine Workflows 為 Physical AI、機器人艦隊與自主機器工作流打造的確定性資料平面
在實體世界中，延遲、抖動與不可預測性並非小瑕疵，而是風險來源

DETERMINISTIC確定性 LLM-AGNOSTICLLM 無繫結 CRASH-CONSISTENT崩潰一致性 ZERO-TRUST零信任 REGRESSION-TESTED回歸測試覆蓋 BYTE-IDENTICAL位元一致性

309 ns DECODE p99解碼 p99

14,513 MSG/S HERO訊息/秒英雄值

<0.7ms E2E p99端對端 p99

6/6 CRASH PASS崩潰恢復

65ms WAL RECOVERYWAL 恢復中位數

// 001 :: ARCHITECTURE// 001 :: 架構設計

Three-Layer Decoupled Architecture三層解耦架構

PROTOCOL LAYER

Canonical encoding with strict byte-identity guarantees. Semantically equivalent JSON produces identical binary output. UTF-8 native. Stateless pure-function codec. No session. No runtime dictionary load. 嚴格 canonical encoding，語意相等的 JSON 產出位元完全相同。UTF-8 原生。Stateless pure-function codec，無 session、無 streaming context、無 runtime dictionary load。

SDK LAYER

Language-neutral reference SDK and encoding bridge. No tokenizer. No model dependency. No LLM API client embedded. No local inference engine. Integrates with any LLM or agent ecosystem. 語言中性的 reference SDK 與編碼橋。無 tokenizer、無 model、無 LLM API client、無本地推理引擎。可介接任何 LLM / agent 生態系統，但本身不依賴任何。

GATEWAY LAYER

Stateless admission control, routing, and full security enforcement. Rolling upgrades without state migration. Pluggable persistence backends: WAL, NATS JetStream, Kafka, Pulsar — without modifying gateway, routing, or security layers. 無狀態 admission、routing 與安全棧。Rolling upgrade 不需狀態遷移。Pluggable persistence backend：WAL 與 NATS JetStream，可擴充 Kafka / Pulsar 無需改動 gateway、routing 或安全層。

✗ Not an LLM wrapper · ✗ Not a model-vendor messaging framework · ✗ Not a local inference engine
A deterministic transport substrate built for verifiable behavior. Bound to no vendor. Tied to no tokenizer. ✗ 非 LLM 包裝器 · ✗ 非 綁定模型廠商的訊息框架 · ✗ 非 本地推理引擎
一個為可驗證行為而生的確定性傳輸底層。不繫結任何廠商，不依賴任何 tokenizer。

// 002 :: VERIFIED METRICS// 002 :: 已驗證性能數據

Independently Reproducible Results可獨立重現的量測結果

Reference hardware: Consumer-class 4-core i7-7700HQ / 8 GB RAM / SATA III SSD | Full methodology: whitepaper §7.2 測試環境: Consumer-class 4-core i7-7700HQ / 8 GB RAM / SATA III SSD | 完整量測方法與實驗設計: 白皮書 §7.2

M-01

DECODE SUB-STAGE LATENCY解碼子階段延遲

309ns

p99 = 0.000309 ms
156,460,102 samples · 3h stable window
Parallel measurement with full critical path 156,460,102 樣本 · 3 小時穩定窗
與完整 critical path 並行量測

M-02

SUSTAINED ACKNOWLEDGED THROUGHPUT持續確認吞吐量 (Single Host)

14,513msg/s

1h Gold1h 黃金

14,209

3h Hero3h 英雄

14,513

4 gateway processes aggregated · single host 4 個 gateway 進程合計 · single host

M-03

END-TO-END LATENCY (WHOLE WINDOW)端對端延遲 (全窗口聚合)

<0.7ms p99

p95

<0.5 ms

p99

<0.7 ms

p999

<3 ms

M-04

STABILITY — 3-HOUR WINDOW穩定度 — 3 小時窗口

0.04CoV

Throughput drift:吞吐漂移: +0.02%
Trimmed CoV: 0.04
10,800 × 1-sec buckets10,800 個一秒桶
p99 > 5ms occurrences:出現次數: 0
Run-to-run reproducibility:跨執行重現性: ≈ 1/5,000

M-05

BLACK-BOX CRASH RECOVERY黑盒崩潰恢復測試

6/6PASS

2 kernel-level kill modes × 3 repetitions
2 種 kernel-level kill mode × 3 次
WAL-internal recovery median: WAL-internal recovery 中位數: 65 ms (64–66 ms)

ZERO
MSG LOSS

ZERO
PHANTOM DUPLICATE

CLEAN
TORN-TAIL OPEN

// 003 :: PROTOCOL & CODEC// 003 :: 協議與編解碼特性

Protocol Properties — Every Property Regression-Tested協議特性 — 每條均對應 Regression Test

⬛

BYTE-IDENTICAL ENCODING位元一致性編碼

Semantically equivalent JSON → identical binary output. Canonical by construction.語意相等的兩個 JSON 產出位元完全相同。由設計保證規範性。

🔤

UTF-8 NATIVE ROUND-TRIPUTF-8 原生往返

CJK, multilingual, and supplementary-plane emoji (U+1F527 / U+2705) preserved without loss.中文、混合語、補充平面 emoji（U+1F527 / U+2705）完全無失真往返。

⚙

STATELESS PURE-FUNCTION CODEC無狀態純函數 Codec

No session context. No streaming state. No runtime dictionary load.無 session、無 streaming context、無 runtime dictionary load。

🔌

LLM-AGNOSTIC BY CONSTRUCTIONLLM 無繫結（由設計保證）

No tokenizer. No model awareness. No LLM API client. No embedded local model.無 tokenizer、無 model、無 LLM API client、無本地小模型。

◉

SELF-CONTAINED CORE PATH自包含核心路徑

No database dependency. No external services. No network touch points in the codec path.無 database 依賴、無外部服務、codec 路徑中無 network touch points。

// 004 :: SECURITY STACK// 004 :: 企業級安全棧

Security — Non-Optional, Enforced in Critical Path安全機制 — 不可靜默關閉，位於 Critical Path

Every security mechanism is enforced in the production critical path. They cannot be silently disabled. Security is structural — not configurable decoration. 所有安全機制均在 production critical path 強制執行，無法靜默關閉。安全性是系統本體的一部分，而非可選附加模組。

Per-Frame Ed25519 Signing

Each frame individually signed. Tamper detection at finest granularity.每幀個別簽名，最細粒度的竄改偵測。

FRAME

Cross-Process Replay Protection

Prevents replay attacks across process boundaries.跨進程邊界的重放攻擊防護。

REPLAY

Cross-Process Token-Bucket Rate Limiting

Coordinated rate limiting across all gateway processes.跨所有 gateway 進程協調的速率限制。

RATE

Queue-Depth Backpressure

Structural flow control prevents overload and cascading failure.結構性流量控制，防止過載與連鎖故障。

FLOW

Zero-Trust Frame Verifier (Receive Path)

Every incoming frame verified regardless of source. Trust is never assumed.每個進入幀均驗證，不論來源為何。信任永遠不被假設。

ZERO-TRUST

// 005 :: DESIGN RATIONALE// 005 :: 設計原則

Why Determinism Matters為何確定性至關重要

In robotics and physical AI systems: latency variance amplifies control error. Non-reproducible performance makes safety modeling impossible. Hidden dependencies — LLMs, databases, external services — become failure sources. 在機器人艦隊、physical AI 與 autonomous workflow 中：不可預期延遲會放大控制誤差；不可重現的性能使風險無法建模；隱性依賴（LLM / DB / 外部服務）會成為故障來源。

DETERMINISM > CONVENIENCE

Predictable behavior under load and failure — not ergonomic shortcuts.負載與故障下的可預測行為，而非便利性捷徑。

ISOLATION > COUPLING

No database. No external service. No LLM in the critical path. Hidden dependencies are failure sources.無 database、無外部服務、critical path 中無 LLM。隱性依賴是故障來源。

MEASURED TRUTH > MARKETING CLAIMS

Every metric independently reproducible. Every property regression-tested. No unverified assertions.每項數據均可獨立重現。每項特性均有回歸測試。無未經驗證的聲明。

// 006 :: USE CASES// 006 :: 適用場景

Intended Applications設計適用場景

▶▶

Agent-to-Agent OrchestrationAgent-to-Agent 協調

Deterministic, verifiable communication between autonomous agents in multi-agent pipelines.多 Agent 管線中自主代理之間的確定性、可驗證通訊。

◉◎

Agent-to-Robot Control PlaneAgent-to-Robot 控制平面

Low-latency, crash-consistent transport for physical AI and robotics fleet command paths.Physical AI 與機器人艦隊指令路徑的低延遲、崩潰一致性傳輸。

⚡⚡

High-Frequency Autonomous Workflows高頻自主機器工作流

≥14,000 msg/s sustained throughput for time-critical autonomous decision loops.≥14,000 msg/s 持續吞吐，支援時間關鍵的自主決策迴圈。

✓✓

Verifiable Safety-Critical Systems可驗證低延遲與崩潰一致性系統

Systems requiring measurable latency SLAs, crash consistency, and independently reproducible benchmarks.需要可量測延遲 SLA、崩潰一致性與可獨立重現基準的系統。