Chapter 05 · 总账Chapter 05 · The Ledger
FLUED Experiment Atlas
完整研究过程，The full research process,不只保留最终点not just the final points
v0.4 -> v3.6 · baselines · ablations · failure cases
这里汇总可公开的关键实验条目和派生训练曲线。每项都标注口径与可信度：历史现象、受控机制实验和可作为主结论的公平比较被明确分开。
This page collects the key public experiment records and their derived training curves. Each entry is labeled with its protocol and confidence level: historical phenomena, controlled mechanism experiments, and fair comparisons that can serve as headline conclusions are clearly separated.
上一章 · 证据边界：关键曲线与严格对照——本章是全部可公开实验记录的总账，以及当前前沿在整个坐标系里的位置。
Previous · Evidence boundaries: key curves and strict controls — this chapter is the full ledger of every public experiment record, and where the current frontier sits in the whole coordinate system.
Experiment Ledger
按版本、阶段和
证据等级筛选
Filter by version, stage,
and evidence level
表格只展示关键指标和结论。原始日志保留在内部归档，用于生成曲线和复核；代码、配置与公开结果摘要可在仓库获得。
The table shows only key metrics and conclusions. Raw logs stay in the internal archive for curve generation and audit; code, configs, and public result summaries are available in the repository.
Training Curves
全量日志点绘制，前端仅做平滑连线
All log points plotted; the frontend only smooths the lines
每条曲线保留其派生日志中的全部记录点。不会为了“好看”补点或跳过 NaN；没有同一口径的结果不会被画在同一比较组。
Every curve keeps all recorded points from its derived log. No points are added and no NaN is skipped for looks; results without a shared protocol are never plotted in the same comparison group.
loading
Rate–Distortion Frontier
v3.6 的信息效率：同指标下 ~60×
v3.6's information efficiency: ~60× at the same metric
横轴为每条 512-byte prompt 传输的标量总数（对数刻度），纵轴为 strict masked-source 补全准确率（越高越好）。八个点全部来自 20K 步、corpus_v3、单 seed=42 的受控 run；S0.7 逐段条件化臂为非默认口径，单独标注。天花板锚 AR H-Net next-byte BPB 0.653 是不同指标，不进 acc 坐标系。
The x-axis is the total number of scalars transmitted per 512-byte prompt (log scale); the y-axis is strict masked-source completion accuracy (higher is better). All eight points come from controlled runs of 20K steps on corpus_v3 with a single seed=42; the S0.7 per-chunk conditioning arm is a non-default protocol, marked separately. The ceiling anchor, AR H-Net next-byte BPB 0.653, is a different metric and does not enter the acc coordinate system.
masked acc 对比（柱上数字 = 信息传输标量数 / 512B prompt）masked acc compared (numbers on bars = transmitted scalars / 512B prompt)
Reading Rules
把结果放回它们真正回答的问题里
Put results back into the questions they actually answer
可作为当前主结论
Usable as current headline conclusions
- v3.6 公平对比：masked acc 0.149 对 HNet-DiT 瓶颈臂 0.142（同规模同步数同语料单 seed），信息传输 1,536 vs ~97,540 标量。
- The v3.6 fair comparison: masked acc 0.149 vs the HNet-DiT bottleneck arm's 0.142 (matched scale, steps, and corpus, single seed), transmitting 1,536 vs ~97,540 scalars.
- v2 A-class 多种子：稳定可微边界和近无损重建。
- v2 A-class multi-seed: stable differentiable boundaries and near-lossless reconstruction.
- v2 D1：统一 2048 原始 byte / 100K 下，FLUED 稳定但尚未超过 BPE。
- v2 D1: under the unified 2048 raw-byte / 100K setting, FLUED is stable but has not yet beaten BPE.
- v3.2.1 strict masked-source：latent no-memory 使小主干 mask accuracy 从 0.1440 提升至 0.1898。
- v3.2.1 strict masked-source: latent no-memory lifts a small backbone's mask accuracy from 0.1440 to 0.1898.
只用于定位或历史叙事
For localization or historical narrative only
- v3.6 S0.7 逐段条件化（unmasked 0.351 / PPL 12.1）是非默认口径诊断臂，不进入主结论。
- v3.6 S0.7 per-chunk conditioning (unmasked 0.351 / PPL 12.1) is a non-default diagnostic arm and does not enter the headline conclusions.
- v1 的 E3 20K BPB 阳性不是 D1 公平比较的替代品。
- v1's E3 20K BPB positive is not a substitute for the D1 fair comparison.
- v3.4 迁移前 5K 筛选用于发现候选组件，不能绕过后续代码纠偏。
- v3.4's pre-migration 5K screening was for discovering candidate components and cannot bypass the later code corrections.
- v3.4 40K/20K 单种子矩阵可定位失稳机制，不能据此声称多种子或尺度规律。
- v3.4's 40K/20K single-seed matrices can localize instability mechanisms but cannot support claims of multi-seed or scaling laws.