Chapter 03 · 眼见为实Chapter 03 · Seeing Is Believing
FLUED v3.6 Live Demo
五个模型，同一段文本，Five models, the same text,各自决定在哪里下刀each deciding where to cut
动态边界切分 · 掩码字节补全 · 本机实时推理Dynamic boundary segmentation · masked byte completion · live on-device inference
本页面在本机访问时连接本机推理服务（127.0.0.1:8722，CPU 实时推理），公网访问时自动展示预录样例。演示把同一段输入同时送入 FLUED v3.6 的两个边界臂（S0 教师粒度 + GRPO R4 自选粒度）和三个 H-Net 基线：先展示每个模型把字节流切成了哪些语义段，再随机掩盖约 5% 的字节让模型补全。KDA（逐通道门控 delta 规则的线性注意力）是 v3.6 的主干状态机；readout 包指整条 prompt 被压缩成的单个逻辑读出单位（1536 个标量）。
When visited locally, this page connects to a local inference service (127.0.0.1:8722, real-time CPU inference); on the public web it automatically shows recorded samples instead. The demo feeds the same input to FLUED v3.6's two boundary arms (S0 teacher granularity + GRPO R4 self-chosen granularity) and three H-Net baselines: first showing how each model segments the byte stream into semantic chunks, then randomly masking about 5% of the bytes for the models to complete. KDA (linear attention with a channel-wise gated delta rule) is v3.6's backbone state machine; the readout package is the single logical readout unit (1,536 scalars) that the whole prompt is compressed into.
上一章 · 架构：整条 prompt 经动态切分、逐段摘要和 KDA 状态机串行消费，压成 1×1,536 标量——本章让它在你眼前跑。
Previous · Architecture: the whole prompt is consumed serially through dynamic segmentation, per-chunk summarization, and the KDA state machine, compressed into 1×1,536 scalars — this chapter runs it in front of you.
Input
输入一段文本，或选一个预置样本
Enter your own text, or pick a preset sample
掩码补全协议在过短的文本上会退化（掩盖位置太少、上下文不足），预置样本均为 300 字节以上的中英混合文本。在线模式下输入会原样发送到本机推理服务；离线模式下无论输入什么都渲染同一份静态样例。
The masked-completion protocol degenerates on very short text (too few masked positions, too little context), so the preset samples are all mixed Chinese-English text over 300 bytes. Online, your input is sent as-is to the local inference service; offline, the same static sample renders regardless of input.
预录样例模式——展示本机推理服务的真实输出录样；本地启动 demo-server 后可对任意输入实时推理Recorded-sample mode — showing recorded real output from the local inference service; start demo-server locally for live inference on any input
Semantic Segmentation
语义段切分对比
Semantic segmentation compared
每个模型各自决定边界：同一段文本按切分结果用交替底色渲染，段与段之间以分隔标记区分。每字节位置的切分置信度画成迷你条形图（与文本位置对齐），超过阈值的高亮柱即实际切点。
Each model decides its own boundaries: the same text renders in alternating shades by segmentation, with divider marks between segments. Per-byte cutting confidence is drawn as a mini bar chart (aligned with text positions); highlighted bars above the threshold are the actual cuts.
HNet-DiT 两臂的行为是真实模型输出，不是渲染问题：标准臂边界退化（几乎不切，约 1 个 chunk），瓶颈臂过度切分（约 190 个 chunk）。没有显式压缩激励时，动态切分不存在中间态——这正是公平对比要说明的现象。
The behavior of both HNet-DiT arms is real model output, not a rendering issue: the standard arm's boundaries degenerate (barely cuts, ~1 chunk) and the bottleneck arm over-segments (~190 chunks). Without explicit compression pressure, dynamic segmentation has no middle ground — exactly the phenomenon the fair comparison is meant to show.
尚未运行演示。输入文本并点击「运行演示」，五个模型的切分结果将在此并排渲染。
The demo has not run yet. Enter text and click "Run Demo" — segmentation from all five models renders side by side here.
Masked Completion
掩码补全：被掩盖的字节能补回多少
Masked completion: how many masked bytes come back
五个模型共享同一份随机字节掩码（下方高亮位置）。逐位置列出真值与预测：绿色为命中，红色为未命中；多字节字符的每个字节位置单独计分，无法单独解码的预测字节显示为占位符「·」。
All five models share one random byte mask (highlighted below). Truth and prediction are listed per position: green for hits, red for misses; every byte position of a multi-byte character is scored separately, and predicted bytes that cannot be decoded alone show as the placeholder "·".
掩码后的文本将在此显示。
The masked text will appear here.
v3.6 有两个训练任务：任务一 readout 包直接进 decoder 做精准还原（direct_acc）；任务二 readout 包先经 TinyBackbone 改写、再由 decoder 全部还原（backbone_masked_acc，即演示的主口径 masked_acc）。H-Net AR 臂是纯因果模型，其口径为「mask 位置的 next-byte argmax 命中率」，与掩码重建不严格可比。
v3.6 has two training tasks: task one sends the readout package straight to the decoder for exact restoration (direct_acc); task two routes it through TinyBackbone first, then the decoder restores everything (backbone_masked_acc — the demo's main masked_acc protocol). The H-Net AR arm is a purely causal model scored by next-byte argmax hit rate at mask positions, not strictly comparable to masked reconstruction.
各模型的补全结果将在此显示。
Each model's completions will appear here.
Latency
真实计算耗时
Real compute latency
推理服务当前运行在 CPU 上：每个模型的 forward 耗时为单次完整前向的实测墙上时钟，总耗时为五个模型串行完整推理的端到端时间（含数据准备）。CPU 串行下端到端为秒级（实测约 1–5 s），这是正常预期，不做毫秒级宣传。
The inference service currently runs on CPU: each model's forward time is the measured wall clock of one complete forward pass, and the total is the end-to-end time for all five models run sequentially (including data preparation). Seconds-level e2e on serial CPU (measured ~1–5 s) is expected — we make no millisecond claims.
耗时数据将在此显示。
Timing data will appear here.
S0.7 逐段条件化（per-chunk readout）已证明检索瓶颈可解：unmasked 全位置 acc 0.190→0.351、PPL 33.9→12.1。本演示为 canonical v36.1 口径（per_chunk_readout 默认关），不构成性能结论。
S0.7 per-chunk conditioning (per-chunk readout) has shown the retrieval bottleneck is solvable: unmasked all-position acc 0.190→0.351, PPL 33.9→12.1. This demo runs the canonical v36.1 protocol (per_chunk_readout off by default) and is not a performance claim.