research: publish AttnRes forward training study
This commit is contained in:
wuyang
+12
-3
@@ -8,7 +8,7 @@
|
||||
|---|---:|---:|---|
|
||||
| 研究框架与规范 | 进行中 | 83% | Scaling Laws 二轮拟合复现与逐图精读 |
|
||||
| 网站设计系统 | 进行中 | 89% | 打印样式与更多通用可视化组件 |
|
||||
| Kimi K3 深读 | 七轮实证进行中 | 99% | 设计前向训练变体，并等待 `A_log` 社区候选的官方裁决 |
|
||||
| Kimi K3 深读 | 八轮实证已收敛 | 100% | 稳定维护；真实 forward 等待 `A_log` 官方裁决 |
|
||||
| 语言模型前史 | 完成首版 | 78% | Kneser–Ney、LSTM、Bahdanau 逐图精读与真实小语料复现 |
|
||||
| Transformer 基础 | 完成首版 | 79% | 多头电路、归一化 traces 与真实 kernel / KV 配置 |
|
||||
| 表示、位置与残差高速公路 | 完成首版 | 81% | 真实 hidden-state / norm traces、长上下文位置外推与深层稳定性消融 |
|
||||
@@ -41,7 +41,7 @@
|
||||
- [x] 完成 486 篇关键论文索引，覆盖 16 个标签专题与 Kimi/DeepSeek 聚光主线。
|
||||
- [x] 完成可检索、可按专题筛选的论文库页面。
|
||||
- [x] 完成 K3、语言模型前史、Transformer 基础、表示/位置/残差、DeepSeek 谱系、Scaling Laws、数据工程、长上下文、MoE、指令微调与人类偏好、推理、Agent、原生多模态、训练系统、推理服务、数值优化与评测安全十七篇首版长文。
|
||||
- [x] 完成 K3 三轴架构、八联报告实验、四联开放工件实验、Round 04 / 05 / 06 / 07 各五联 AttnRes 独立实验、语言模型前史四联实验、Transformer 四联实验、表示深度四联实验、DeepSeek 二十二联实验、长上下文、MoE 路由、推理三页签，以及训练系统、推理服务、Scaling、数据工程、数值、Alignment、Agent、原生多模态与评测安全专题各四页签等一百零九个原创交互视图。
|
||||
- [x] 完成 K3 三轴架构、八联报告实验、四联开放工件实验、Round 04 / 05 / 06 / 07 / 08 各五联 AttnRes 独立实验、语言模型前史四联实验、Transformer 四联实验、表示深度四联实验、DeepSeek 二十二联实验、长上下文、MoE 路由、推理三页签，以及训练系统、推理服务、Scaling、数据工程、数值、Alignment、Agent、原生多模态与评测安全专题各四页签等一百一十四个原创交互视图。
|
||||
- [x] 完成长上下文首版：五张成本账、26 篇一手论文、10+ 机制图与 8 策略交互实验室。
|
||||
- [x] 核验 FlashAttention、DeepSeek-V2/V3.2/V4、Kimi Linear/K3 等六份论文原文，并建立长上下文研究账本。
|
||||
- [x] 核验 Switch、ST-MoE、DeepSeekMoE、Loss-Free、V3、LatentMoE 与 K3 原文，并建立 MoE 研究账本。
|
||||
@@ -305,10 +305,15 @@
|
||||
- [x] K3 Round 07 五视图实验室完成：65-node 路径图、14-mask 全矩阵、双向主门、branch/output controls 与 32 层原始谱/replay 审计可交互；protocol、scoping、Grok 结果前审阅、audit、runner、analyzer、packager、四个 raw JSON、aggregate、compact 与 reproduction 全部进入公开树。
|
||||
- [x] Round 07 本地闸门通过：100 个 Astro 文件零诊断，21 个页面、1,151 个站内引用、12 个跨页锚点零失败；Round 04/05/06/07 四套冻结数据、四套 K3 专项、K3 全量与全站 23 套真实 Chrome 回归通过。动态生成矩阵的 scoped CSS 退化由截图复查发现并修复；桌面/390px 移动端零文档级溢出、零 offender、零运行时异常。
|
||||
- [x] K3 Round 07 以功能源提交 `dc8ec30`、不可变镜像 `20260730T063811Z-dc8ec30` 发布；OCI index digest `sha256:98d441412408774628195ce23f431d25034bb1590210c11230d884513898e71d`。NAS `12010→8080` healthy / 0 次重启、Compose Manager 标签、VPS→NAS、NPM host 31 / cert 41、DNS、HTTPS/2、gzip、immutable asset、首页/K3 公网内容、门户 `LLM ATLAS / projects / 180` 与全站 23 套生产 Chrome 回归全部通过；保留 Round 06 `20260730T042651Z-b393780` 回滚点。
|
||||
- [x] K3 Round 08 在正式输出前冻结 `llm-atlas-k3-attnres-forward-training-v1`：四个 train-time forward variants × 三 seed × 8,000 steps、同 seed Round 05 historical pairing、step 8,000 的 layers 21–25 contrast / peak 双指标 20% 主门、BPC 每 seed / mean 护栏、描述性 `I67` 与完整 primary replay；不把架构消融写成 pure-forward 因果实验。
|
||||
- [x] 12 formal + 1 replay 全部一次完成，每格 65,536,000 target bytes；新处理总量 851,968,000，历史 references 196,608,000 单列。联合 groups 6+7 的 contrast / peak 六格降幅为 62.1%–77.3% / 32.3%–62.0%，6 / 6 通过；BPC 三 seed 最大 `+.009598`、均值 `+.006578`，质量门 4 / 4 通过。
|
||||
- [x] primary seed-1 replay scientific payload exact，SHA-256 为 `b85563ca…c051`；冻结主状态为 `forward_training_attenuation_established_within_reduced_protocol`。`I67` 的 step-8,000 mean 为 contrast `−.3670`、peak `−.1704`，只保留为跨独立训练的描述性 log residual。
|
||||
- [x] 结果前 Grok 实现审阅指出 smoke-only empty selector 的空 census；矩阵结束后删除 early return、加入 `forward_calls > 0`，修补后的 learned wrapper 实际执行 39 次 forward，父/包装器 15 组科学字段仍 exact。结果后 Grok 只读复算报告 `blocking_errors=0`、status / replay 均确认。
|
||||
- [x] K3 Round 08 五视图实验室完成：forward contract、六 checkpoint 训练轨迹、attenuation+BPC 主门、non-additivity map 与 32 层谱/replay audit 可交互；protocol、scoping、两阶段 Grok 审阅、runner、analyzer、13 raw、aggregate、compact、reproduction 与结果审计全部进入公开树。
|
||||
- [x] Round 08 本地闸门通过：101 个 Astro 文件零诊断，21 个页面、1,151 个站内引用、12 个跨页锚点零失败；Round 04–08 五套冻结数据、五套 K3 专项、K3 全量与全站 24 套真实 Chrome 回归通过。截图复查修复黑色实验室标题对比度；桌面/390px 移动端零文档级溢出、零 offender、零运行时异常。
|
||||
|
||||
## 正在进行
|
||||
|
||||
- [ ] K3 七轮下一闸门：设计前向训练变体与非加性局部交互地图；真实 K3 forward 继续等待 `A_log [128]↔[96]` 社区候选的官方裁决或权重修订。
|
||||
- [ ] DeepSeek 八轮下一闸门：推进干预式 mediation、SM90 FlashMLA、FP8 / pipeline traces 与 R1-like RL 小模型复现。
|
||||
- [ ] 表示、位置与残差二轮：真实 hidden-state / norm traces、长上下文位置外推复现与 mHC / AttnRes 深层稳定性消融。
|
||||
- [ ] 评测安全二轮：真实 cross-harness / pass@k 复跑、Judge 元评测、动态污染与过拒案例。
|
||||
@@ -517,6 +522,10 @@
|
||||
| 2026-07-30 | 局部 score 不写成可加贡献率 | 同一 scope 在 learned 与 uniform 背景的响应不同，`S_peak > 1` 与负 interaction residual 都是非加性诊断，不是 170% 贡献或方差分解 |
|
||||
| 2026-07-30 | branch 与 output 控制保持次级证据身份 | group 7 MLP-only 的 6/6 只属于 sufficiency branch gate；group 6 为 5/6，output-only 为 0/6，都不能补救失败的双向主门 |
|
||||
| 2026-07-30 | K3 Round 07 局部路径里程碑发布 | 功能源 `dc8ec30`、镜像 `20260730T063811Z-dc8ec30`、OCI `sha256:98d44141…e71d`；21/21 公网页面链路与全站 23 套生产 Chrome 通过，保留 Round 06 回滚点 |
|
||||
| 2026-07-30 | 训练期前向干预属于架构消融 | selected uniform mixer 同时改变 train/eval forward、natural backward 与后续 updates；不能写成只改 forward 的路径因果 |
|
||||
| 2026-07-30 | Round 08 主门与质量门同时成立 | groups 6+7 六个 attenuation 格 6/6 ≥20%；三 seed ΔBPC mean `+.006578`，4/4 过闸；结论只限固定缩小协议 |
|
||||
| 2026-07-30 | non-additivity 永久保留描述身份 | `I67` 来自三套独立训练，只是 cross-run log residual，不是因果 interaction、Shapley 或贡献率 |
|
||||
| 2026-07-30 | K3 研究线在 Round 08 主动收敛 | 不启动 Round 09；公开 `A_log [128]↔[96]` 冲突继续等待官方裁决，现有五轮实证停在可复现、可回滚的稳定边界 |
|
||||
|
||||
## 未决问题
|
||||
|
||||
|
||||
@@ -6,6 +6,8 @@ This directory implements preregistered protocol
|
||||
- `research/K3_ATTNRES_FORWARD_TRAINING_SCOPING.md`
|
||||
- `research/K3_ATTNRES_FORWARD_TRAINING_PROTOCOL.md`
|
||||
- `research/K3_ATTNRES_FORWARD_TRAINING_GROK_REVIEW.md`
|
||||
- `research/K3_ATTNRES_FORWARD_TRAINING_IMPLEMENTATION_REVIEW.md`
|
||||
- `research/K3_ATTNRES_FORWARD_TRAINING_AUDIT.md`
|
||||
|
||||
It is a depth-32 reduced Block AttnRes architecture ablation. It is not a
|
||||
Kimi-K3 checkpoint forward pass and does not claim to recover unpublished
|
||||
@@ -56,3 +58,13 @@ This runs 12 formal cells and one full replay. The analyzer reads all cells,
|
||||
the frozen historical paired references, and generates the only authoritative
|
||||
status, interaction map, and website compact artifact.
|
||||
|
||||
The checked-in Round 08 release contains:
|
||||
|
||||
- 13 raw results under `results/raw/`;
|
||||
- `reproduction.json` with the exact primary scientific-payload hash;
|
||||
- aggregate / compact website data under `src/data/`;
|
||||
- a frozen-data checker and real-Chrome five-view regression in `scripts/`.
|
||||
|
||||
The established status is deliberately scoped to this reduced protocol. It is
|
||||
not a real Kimi-K3 checkpoint result or a reproduction of unpublished Figure
|
||||
5(c) telemetry.
|
||||
|
||||
@@ -425,6 +425,19 @@ def main() -> None:
|
||||
== environment_metadata(references[seed])
|
||||
),
|
||||
}
|
||||
metadata_warnings = [
|
||||
{
|
||||
"cell": cell,
|
||||
"message": (
|
||||
"GPU/version metadata differs from the historical paired "
|
||||
"reference; frozen scientific-environment fields still match"
|
||||
),
|
||||
"run_environment": item["run_environment"],
|
||||
"reference_environment": item["reference_environment"],
|
||||
}
|
||||
for cell, item in pairing.items()
|
||||
if not item["metadata_equal"]
|
||||
]
|
||||
|
||||
replay = read_result(args.replay, PROTOCOL_ID)
|
||||
replay_contract = manifest["replay"]
|
||||
@@ -593,6 +606,7 @@ def main() -> None:
|
||||
"thresholds": manifest["thresholds"],
|
||||
"input_files": input_files,
|
||||
"historical_pairing": pairing,
|
||||
"metadata_warnings": metadata_warnings,
|
||||
"replay": {
|
||||
"passed": replay_exact,
|
||||
"scientific_payload_sha256": canonical_sha256(formal_payload),
|
||||
@@ -636,6 +650,7 @@ def main() -> None:
|
||||
"trajectories": trajectories,
|
||||
"final_spectra": final_spectra,
|
||||
"replay": aggregate["replay"],
|
||||
"metadata_warnings": metadata_warnings,
|
||||
"processed_target_bytes": manifest["new_target_bytes"],
|
||||
"reporting_boundary": aggregate["reporting_boundary"],
|
||||
"aggregate_sha256": aggregate["canonical_sha256_without_self"],
|
||||
|
||||
@@ -0,0 +1,25 @@
|
||||
{
|
||||
"canonical_sha256_without_self": "57346df80c0d76bd1d306d5fa213feed74ac2094237c22ebcb49f16ea16437e0",
|
||||
"excluded_fields": [
|
||||
"run_kind",
|
||||
"timing",
|
||||
"self hashes",
|
||||
"manifest path strings",
|
||||
"GPU/version metadata"
|
||||
],
|
||||
"formal_file_sha256": "0962ebd1a00a11e61ac795282bfa99412166c8752f2a3731f137030d7f134dc1",
|
||||
"formal_variant": "uniform_groups_6_7_forward",
|
||||
"passed": true,
|
||||
"post_result_grok_review": {
|
||||
"blocking_errors": 0,
|
||||
"claim_boundary_confirmed": true,
|
||||
"replay_confirmed": true,
|
||||
"session_id": "019fb28b-a9e1-7643-8e43-06f5e16a2077",
|
||||
"status_confirmed": true
|
||||
},
|
||||
"protocol_id": "llm-atlas-k3-attnres-forward-training-v1",
|
||||
"replay_file_sha256": "b85ac8062b2b0b8b3f2305d4b22a7c0212466fbfb28a9d63099cadd0a6917c8e",
|
||||
"schema_version": 1,
|
||||
"scientific_payload_sha256": "b85563ca5cb53e60b39c3801d372376206105b8a089a8633b3e81973a7f0c051",
|
||||
"seed": 2026073001
|
||||
}
|
||||
@@ -1,5 +1,5 @@
|
||||
{
|
||||
"canonical_sha256_without_self": "00db0569d2fd00599383ecdc50578c52df784a937a96dd69576f1366e88b0dcb",
|
||||
"canonical_sha256_without_self": "9a0716bc496d39b9447b79aba0aec42ca4c953cbabe30e3aa349c1bba0701c5f",
|
||||
"environment": {
|
||||
"cublas_workspace_config": ":4096:8",
|
||||
"cuda": "12.8",
|
||||
@@ -21,7 +21,8 @@
|
||||
"initial_mixer_hash": true,
|
||||
"initial_public_hash": true,
|
||||
"logits_sha256": true,
|
||||
"loss_nats": true
|
||||
"loss_nats": true,
|
||||
"loss_tensor_sha256": true
|
||||
},
|
||||
"passed": true,
|
||||
"payload": {
|
||||
@@ -1254,7 +1255,8 @@
|
||||
"initial_mixer_hash": "c5a06c218c4501b16fcccee115aa8c79d5dceade78b38bbb547e4f5b2202516d",
|
||||
"initial_public_hash": "73bbe569e1a47981386ebaf59ca891174746bf81897fe5403676ad6d50423c58",
|
||||
"logits_sha256": "6b44486d93c29a13dc0b961117cc2d22f26c81dc519346ac45b1bb0fefa29939",
|
||||
"loss_nats": 5.516995429992676
|
||||
"loss_nats": 5.516995429992676,
|
||||
"loss_tensor_sha256": "c4e1d45c137a73ad6249c2db3ff5e29c849310d4f71f3182c41d27ff6b9fcb6a"
|
||||
},
|
||||
"pre_reduction_uniform_weight_gate": true,
|
||||
"pre_reduction_uniform_weight_max_abs_error": 0.0,
|
||||
@@ -1267,7 +1269,8 @@
|
||||
"initial_mixer_hash": true,
|
||||
"initial_public_hash": true,
|
||||
"logits_sha256": true,
|
||||
"loss_nats": true
|
||||
"loss_nats": true,
|
||||
"loss_tensor_sha256": true
|
||||
},
|
||||
"passed": true,
|
||||
"payload": {
|
||||
@@ -2500,7 +2503,8 @@
|
||||
"initial_mixer_hash": "c5a06c218c4501b16fcccee115aa8c79d5dceade78b38bbb547e4f5b2202516d",
|
||||
"initial_public_hash": "73bbe569e1a47981386ebaf59ca891174746bf81897fe5403676ad6d50423c58",
|
||||
"logits_sha256": "6b44486d93c29a13dc0b961117cc2d22f26c81dc519346ac45b1bb0fefa29939",
|
||||
"loss_nats": 5.516995429992676
|
||||
"loss_nats": 5.516995429992676,
|
||||
"loss_tensor_sha256": "c4e1d45c137a73ad6249c2db3ff5e29c849310d4f71f3182c41d27ff6b9fcb6a"
|
||||
},
|
||||
"pre_reduction_uniform_weight_gate": true,
|
||||
"pre_reduction_uniform_weight_max_abs_error": 0.0,
|
||||
@@ -2562,7 +2566,8 @@
|
||||
"initial_mixer_hash": true,
|
||||
"initial_public_hash": true,
|
||||
"logits_sha256": true,
|
||||
"loss_nats": true
|
||||
"loss_nats": true,
|
||||
"loss_tensor_sha256": true
|
||||
},
|
||||
"passed": true,
|
||||
"payload": {
|
||||
@@ -3795,7 +3800,8 @@
|
||||
"initial_mixer_hash": "c5a06c218c4501b16fcccee115aa8c79d5dceade78b38bbb547e4f5b2202516d",
|
||||
"initial_public_hash": "73bbe569e1a47981386ebaf59ca891174746bf81897fe5403676ad6d50423c58",
|
||||
"logits_sha256": "6b44486d93c29a13dc0b961117cc2d22f26c81dc519346ac45b1bb0fefa29939",
|
||||
"loss_nats": 5.516995429992676
|
||||
"loss_nats": 5.516995429992676,
|
||||
"loss_tensor_sha256": "c4e1d45c137a73ad6249c2db3ff5e29c849310d4f71f3182c41d27ff6b9fcb6a"
|
||||
},
|
||||
"pre_reduction_uniform_weight_gate": true,
|
||||
"pre_reduction_uniform_weight_max_abs_error": 0.0,
|
||||
@@ -3857,7 +3863,8 @@
|
||||
"initial_mixer_hash": true,
|
||||
"initial_public_hash": true,
|
||||
"logits_sha256": true,
|
||||
"loss_nats": true
|
||||
"loss_nats": true,
|
||||
"loss_tensor_sha256": true
|
||||
},
|
||||
"passed": true,
|
||||
"payload": {
|
||||
@@ -5090,7 +5097,8 @@
|
||||
"initial_mixer_hash": "c5a06c218c4501b16fcccee115aa8c79d5dceade78b38bbb547e4f5b2202516d",
|
||||
"initial_public_hash": "73bbe569e1a47981386ebaf59ca891174746bf81897fe5403676ad6d50423c58",
|
||||
"logits_sha256": "6b44486d93c29a13dc0b961117cc2d22f26c81dc519346ac45b1bb0fefa29939",
|
||||
"loss_nats": 5.516995429992676
|
||||
"loss_nats": 5.516995429992676,
|
||||
"loss_tensor_sha256": "c4e1d45c137a73ad6249c2db3ff5e29c849310d4f71f3182c41d27ff6b9fcb6a"
|
||||
},
|
||||
"pre_reduction_uniform_weight_gate": true,
|
||||
"pre_reduction_uniform_weight_max_abs_error": 0.0,
|
||||
@@ -5128,7 +5136,8 @@
|
||||
"initial_mixer_hash": true,
|
||||
"initial_public_hash": true,
|
||||
"logits_sha256": true,
|
||||
"loss_nats": true
|
||||
"loss_nats": true,
|
||||
"loss_tensor_sha256": true
|
||||
},
|
||||
"passed": true,
|
||||
"payload": {
|
||||
@@ -6361,7 +6370,8 @@
|
||||
"initial_mixer_hash": "c5a06c218c4501b16fcccee115aa8c79d5dceade78b38bbb547e4f5b2202516d",
|
||||
"initial_public_hash": "73bbe569e1a47981386ebaf59ca891174746bf81897fe5403676ad6d50423c58",
|
||||
"logits_sha256": "6b44486d93c29a13dc0b961117cc2d22f26c81dc519346ac45b1bb0fefa29939",
|
||||
"loss_nats": 5.516995429992676
|
||||
"loss_nats": 5.516995429992676,
|
||||
"loss_tensor_sha256": "c4e1d45c137a73ad6249c2db3ff5e29c849310d4f71f3182c41d27ff6b9fcb6a"
|
||||
},
|
||||
"pre_reduction_uniform_weight_gate": true,
|
||||
"pre_reduction_uniform_weight_max_abs_error": 0.0,
|
||||
|
||||
@@ -1,7 +1,7 @@
|
||||
{
|
||||
"architecture": "block",
|
||||
"batch_size": 32,
|
||||
"canonical_sha256_without_self": "a467cc1c05e1b860abbf44086273071bcd2a891db4354cc6e79b319e44d1cce5",
|
||||
"canonical_sha256_without_self": "81b011e2ab02c49abb6b2686b3e0701d71ab6e7f46cbb7c426eec91702836b7f",
|
||||
"depth": 32,
|
||||
"diagnostics": [
|
||||
{
|
||||
@@ -2479,74 +2479,74 @@
|
||||
],
|
||||
"forward_intervention": {
|
||||
"depth_visit_counts": [
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0,
|
||||
0
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39,
|
||||
39
|
||||
],
|
||||
"forward_calls": 0,
|
||||
"forward_calls": 39,
|
||||
"output_mixer_selected": false,
|
||||
"output_visit_count": 0,
|
||||
"output_visit_count": 39,
|
||||
"passed": true,
|
||||
"selected_depth_indices": [],
|
||||
"selected_parameter_reachability_gate": true,
|
||||
@@ -2555,70 +2555,198 @@
|
||||
"selector_gate": true,
|
||||
"semantics": "selected depth mixers use parameter-free constant-zero logits with the parent softmax+einsum arithmetic kernel",
|
||||
"source_counts_by_depth_index": {
|
||||
"0": [],
|
||||
"1": [],
|
||||
"10": [],
|
||||
"11": [],
|
||||
"12": [],
|
||||
"13": [],
|
||||
"14": [],
|
||||
"15": [],
|
||||
"16": [],
|
||||
"17": [],
|
||||
"18": [],
|
||||
"19": [],
|
||||
"2": [],
|
||||
"20": [],
|
||||
"21": [],
|
||||
"22": [],
|
||||
"23": [],
|
||||
"24": [],
|
||||
"25": [],
|
||||
"26": [],
|
||||
"27": [],
|
||||
"28": [],
|
||||
"29": [],
|
||||
"3": [],
|
||||
"30": [],
|
||||
"31": [],
|
||||
"32": [],
|
||||
"33": [],
|
||||
"34": [],
|
||||
"35": [],
|
||||
"36": [],
|
||||
"37": [],
|
||||
"38": [],
|
||||
"39": [],
|
||||
"4": [],
|
||||
"40": [],
|
||||
"41": [],
|
||||
"42": [],
|
||||
"43": [],
|
||||
"44": [],
|
||||
"45": [],
|
||||
"46": [],
|
||||
"47": [],
|
||||
"48": [],
|
||||
"49": [],
|
||||
"5": [],
|
||||
"50": [],
|
||||
"51": [],
|
||||
"52": [],
|
||||
"53": [],
|
||||
"54": [],
|
||||
"55": [],
|
||||
"56": [],
|
||||
"57": [],
|
||||
"58": [],
|
||||
"59": [],
|
||||
"6": [],
|
||||
"60": [],
|
||||
"61": [],
|
||||
"62": [],
|
||||
"63": [],
|
||||
"7": [],
|
||||
"8": [],
|
||||
"9": []
|
||||
"0": [
|
||||
1
|
||||
],
|
||||
"1": [
|
||||
2
|
||||
],
|
||||
"10": [
|
||||
3
|
||||
],
|
||||
"11": [
|
||||
3
|
||||
],
|
||||
"12": [
|
||||
3
|
||||
],
|
||||
"13": [
|
||||
3
|
||||
],
|
||||
"14": [
|
||||
3
|
||||
],
|
||||
"15": [
|
||||
3
|
||||
],
|
||||
"16": [
|
||||
3
|
||||
],
|
||||
"17": [
|
||||
4
|
||||
],
|
||||
"18": [
|
||||
4
|
||||
],
|
||||
"19": [
|
||||
4
|
||||
],
|
||||
"2": [
|
||||
2
|
||||
],
|
||||
"20": [
|
||||
4
|
||||
],
|
||||
"21": [
|
||||
4
|
||||
],
|
||||
"22": [
|
||||
4
|
||||
],
|
||||
"23": [
|
||||
4
|
||||
],
|
||||
"24": [
|
||||
4
|
||||
],
|
||||
"25": [
|
||||
5
|
||||
],
|
||||
"26": [
|
||||
5
|
||||
],
|
||||
"27": [
|
||||
5
|
||||
],
|
||||
"28": [
|
||||
5
|
||||
],
|
||||
"29": [
|
||||
5
|
||||
],
|
||||
"3": [
|
||||
2
|
||||
],
|
||||
"30": [
|
||||
5
|
||||
],
|
||||
"31": [
|
||||
5
|
||||
],
|
||||
"32": [
|
||||
5
|
||||
],
|
||||
"33": [
|
||||
6
|
||||
],
|
||||
"34": [
|
||||
6
|
||||
],
|
||||
"35": [
|
||||
6
|
||||
],
|
||||
"36": [
|
||||
6
|
||||
],
|
||||
"37": [
|
||||
6
|
||||
],
|
||||
"38": [
|
||||
6
|
||||
],
|
||||
"39": [
|
||||
6
|
||||
],
|
||||
"4": [
|
||||
2
|
||||
],
|
||||
"40": [
|
||||
6
|
||||
],
|
||||
"41": [
|
||||
7
|
||||
],
|
||||
"42": [
|
||||
7
|
||||
],
|
||||
"43": [
|
||||
7
|
||||
],
|
||||
"44": [
|
||||
7
|
||||
],
|
||||
"45": [
|
||||
7
|
||||
],
|
||||
"46": [
|
||||
7
|
||||
],
|
||||
"47": [
|
||||
7
|
||||
],
|
||||
"48": [
|
||||
7
|
||||
],
|
||||
"49": [
|
||||
8
|
||||
],
|
||||
"5": [
|
||||
2
|
||||
],
|
||||
"50": [
|
||||
8
|
||||
],
|
||||
"51": [
|
||||
8
|
||||
],
|
||||
"52": [
|
||||
8
|
||||
],
|
||||
"53": [
|
||||
8
|
||||
],
|
||||
"54": [
|
||||
8
|
||||
],
|
||||
"55": [
|
||||
8
|
||||
],
|
||||
"56": [
|
||||
8
|
||||
],
|
||||
"57": [
|
||||
9
|
||||
],
|
||||
"58": [
|
||||
9
|
||||
],
|
||||
"59": [
|
||||
9
|
||||
],
|
||||
"6": [
|
||||
2
|
||||
],
|
||||
"60": [
|
||||
9
|
||||
],
|
||||
"61": [
|
||||
9
|
||||
],
|
||||
"62": [
|
||||
9
|
||||
],
|
||||
"63": [
|
||||
9
|
||||
],
|
||||
"7": [
|
||||
2
|
||||
],
|
||||
"8": [
|
||||
2
|
||||
],
|
||||
"9": [
|
||||
3
|
||||
]
|
||||
},
|
||||
"uniform_weight_gate": true,
|
||||
"uniform_weight_max_abs_error": 0.0,
|
||||
|
||||
+14045
File diff suppressed because it is too large
Load Diff
+14045
File diff suppressed because it is too large
Load Diff
+14045
File diff suppressed because it is too large
Load Diff
+14045
File diff suppressed because it is too large
Load Diff
+14045
File diff suppressed because it is too large
Load Diff
+14045
File diff suppressed because it is too large
Load Diff
+14029
File diff suppressed because it is too large
Load Diff
+14029
File diff suppressed because it is too large
Load Diff
+14029
File diff suppressed because it is too large
Load Diff
+14077
File diff suppressed because it is too large
Load Diff
+14077
File diff suppressed because it is too large
Load Diff
+14077
File diff suppressed because it is too large
Load Diff
+14077
File diff suppressed because it is too large
Load Diff
@@ -84,6 +84,34 @@ def validate_manifest(args: argparse.Namespace) -> None:
|
||||
raise ValueError("the frozen protocol permits one or two processes")
|
||||
|
||||
|
||||
def stop_processes(items: list[dict[str, Any]]) -> None:
|
||||
for item in items:
|
||||
if item["process"].poll() is None:
|
||||
item["process"].terminate()
|
||||
for item in items:
|
||||
process = item["process"]
|
||||
if process.poll() is not None:
|
||||
continue
|
||||
try:
|
||||
process.wait(timeout=10)
|
||||
except subprocess.TimeoutExpired:
|
||||
process.kill()
|
||||
process.wait()
|
||||
|
||||
|
||||
def quarantine_failed_output(
|
||||
output_dir: Path, variant: str, seed: int, run_kind: str
|
||||
) -> str | None:
|
||||
output = cell_output(output_dir, variant, seed, run_kind)
|
||||
if not output.exists():
|
||||
return None
|
||||
failed = output.with_suffix(".failed.json")
|
||||
if failed.exists():
|
||||
failed = output.with_suffix(f".failed-{time.time_ns()}.json")
|
||||
output.replace(failed)
|
||||
return str(failed)
|
||||
|
||||
|
||||
def run_cells(
|
||||
args: argparse.Namespace,
|
||||
cells: list[tuple[str, int, str]],
|
||||
@@ -138,13 +166,15 @@ def run_cells(
|
||||
variant, seed, run_kind = item["identity"]
|
||||
elapsed = time.monotonic() - item["started"]
|
||||
if return_code != 0:
|
||||
for survivor in survivors:
|
||||
survivor["process"].terminate()
|
||||
for survivor in running:
|
||||
if survivor is not item and survivor not in survivors:
|
||||
survivor["process"].terminate()
|
||||
stop_processes(
|
||||
[candidate for candidate in running if candidate is not item]
|
||||
)
|
||||
quarantined = quarantine_failed_output(
|
||||
args.output_dir, variant, seed, run_kind
|
||||
)
|
||||
raise RuntimeError(
|
||||
f"cell failed: {variant}/{seed}/{run_kind}: {return_code}"
|
||||
f"cell failed: {variant}/{seed}/{run_kind}: {return_code}; "
|
||||
f"quarantined_output={quarantined}"
|
||||
)
|
||||
completed += 1
|
||||
print(
|
||||
|
||||
@@ -170,9 +170,6 @@ class ForwardInterventionLanguageModel(parent.GradientLanguageModel):
|
||||
def forward(
|
||||
self, input_ids: torch.Tensor, capture: bool = False
|
||||
) -> tuple[torch.Tensor, parent.ActivationTrace | None]:
|
||||
if not self.selected_indices:
|
||||
return super().forward(input_ids, capture)
|
||||
|
||||
self.forward_calls += 1
|
||||
embedded = self.embed(input_ids)
|
||||
trace = parent.ActivationTrace([], [], [], [], []) if capture else None
|
||||
@@ -276,11 +273,13 @@ def build_intervention_audit(
|
||||
str(index): (
|
||||
source_counts[str(index)]
|
||||
== [study_manifest["selected_source_counts"][str(index)]]
|
||||
== [EXPECTED_SOURCE_COUNTS[index]]
|
||||
)
|
||||
for index in selected
|
||||
}
|
||||
visit_gate = (
|
||||
all(value == model.forward_calls for value in model.depth_visits)
|
||||
model.forward_calls > 0
|
||||
and all(value == model.forward_calls for value in model.depth_visits)
|
||||
and model.output_visits == model.forward_calls
|
||||
)
|
||||
selected_parameter_gate = all(
|
||||
@@ -415,6 +414,14 @@ def main() -> None:
|
||||
)
|
||||
if expected != VARIANTS[variant]:
|
||||
raise ValueError("study manifest selector drift")
|
||||
for index, source_count in EXPECTED_SOURCE_COUNTS.items():
|
||||
if (
|
||||
study_manifest["selected_source_counts"].get(str(index))
|
||||
!= source_count
|
||||
):
|
||||
raise ValueError(
|
||||
f"study manifest source-count drift at depth index {index}"
|
||||
)
|
||||
|
||||
output_value = argument_value("--output")
|
||||
if output_value is None:
|
||||
|
||||
@@ -92,10 +92,14 @@ def smoke_compare(args: argparse.Namespace) -> None:
|
||||
"gradient_gate",
|
||||
"environment",
|
||||
)
|
||||
checks = {
|
||||
field: parent_result[field] == wrapper_result[field]
|
||||
for field in fields
|
||||
}
|
||||
checks = {}
|
||||
for field in fields:
|
||||
parent_value = exact_structure(parent_result[field])
|
||||
wrapper_value = exact_structure(wrapper_result[field])
|
||||
if field == "manifest":
|
||||
parent_value.pop("path", None)
|
||||
wrapper_value.pop("path", None)
|
||||
checks[field] = parent_value == wrapper_value
|
||||
wrapper_identity = {
|
||||
"protocol": wrapper_result["protocol_id"] == runner.PROTOCOL_ID,
|
||||
"parent_protocol": (
|
||||
@@ -185,6 +189,7 @@ def step_zero(args: argparse.Namespace) -> None:
|
||||
"initial_mixer_hash": initial_mixer,
|
||||
"logits_sha256": tensor_sha256(logits),
|
||||
"loss_nats": loss.detach().cpu().item(),
|
||||
"loss_tensor_sha256": tensor_sha256(loss),
|
||||
"evaluation": exact_structure(evaluation),
|
||||
"diagnostic": exact_structure(diagnostic),
|
||||
}
|
||||
|
||||
@@ -23,6 +23,7 @@
|
||||
"check:data:k3-attnres-gradient": "node scripts/check-k3-attnres-gradient-data.mjs",
|
||||
"check:data:k3-attnres-spike": "node scripts/check-k3-attnres-spike-data.mjs",
|
||||
"check:data:k3-attnres-local-path": "node scripts/check-k3-attnres-local-path-data.mjs",
|
||||
"check:data:k3-attnres-forward": "node scripts/check-k3-attnres-forward-data.mjs",
|
||||
"check:site": "node scripts/check-site.mjs",
|
||||
"check:moe-browser": "node scripts/check-moe-browser.mjs",
|
||||
"check:reasoning-browser": "node scripts/check-reasoning-browser.mjs",
|
||||
@@ -46,6 +47,7 @@
|
||||
"check:k3-attnres-gradient-browser": "node scripts/check-k3-attnres-gradient-browser.mjs",
|
||||
"check:k3-attnres-spike-browser": "node scripts/check-k3-attnres-spike-browser.mjs",
|
||||
"check:k3-attnres-local-path-browser": "node scripts/check-k3-attnres-local-path-browser.mjs",
|
||||
"check:k3-attnres-forward-browser": "node scripts/check-k3-attnres-forward-browser.mjs",
|
||||
"check:k3-browser": "node scripts/check-k3-browser.mjs"
|
||||
},
|
||||
"dependencies": {
|
||||
|
||||
@@ -0,0 +1,112 @@
|
||||
# Round 08 AttnRes 训练期前向干预：结果审计
|
||||
|
||||
审计日期：2026-07-30
|
||||
协议：`llm-atlas-k3-attnres-forward-training-v1`
|
||||
结果后 Grok 会话：`019fb28b-a9e1-7643-8e43-06f5e16a2077`
|
||||
|
||||
## 1. 一句话结论
|
||||
|
||||
冻结 analyzer 的唯一主状态为：
|
||||
|
||||
```text
|
||||
forward_training_attenuation_established_within_reduced_protocol
|
||||
```
|
||||
|
||||
联合 `groups 6+7` 的训练期 uniform-forward 消融在三个预注册 seed 上同时通过
|
||||
`spike contrast` 与 `peak / mean` 的 20% attenuation 门；逐 seed 与三 seed 平均
|
||||
validation BPC 也全部通过质量护栏。结果后独立只读复算得到：
|
||||
|
||||
```text
|
||||
blocking_errors = 0
|
||||
status_confirmed = true
|
||||
replay_confirmed = true
|
||||
```
|
||||
|
||||
这只是在固定 depth-32 缩小 Block AttnRes、固定数据与 8,000-step 预算中的训练期
|
||||
架构消融；**不是真实 Kimi-K3 / 2.8T checkpoint 结果，也不是 Figure 5(c) 未公开
|
||||
telemetry 的复现。**
|
||||
|
||||
## 2. 主门复算
|
||||
|
||||
冻结定义：
|
||||
|
||||
```text
|
||||
S = layers 21–25
|
||||
R = other 27 layers
|
||||
C = mean(g[S]) / mean(g[R])
|
||||
P = max(g) / mean(g)
|
||||
D = (X_reference - X_variant) / X_reference
|
||||
```
|
||||
|
||||
主变体为 `uniform_groups_6_7_forward`，只读复算如下：
|
||||
|
||||
| seed | C reference | C variant | C drop | P reference | P variant | P drop | ΔBPC |
|
||||
|---|---:|---:|---:|---:|---:|---:|---:|
|
||||
| 2026073001 | 3.046269 | 0.806011 | 73.54% | 3.219067 | 1.435618 | 55.40% | +0.005696 |
|
||||
| 2026073002 | 3.332848 | 0.758201 | 77.25% | 3.618924 | 1.375016 | 62.00% | +0.009598 |
|
||||
| 2026073003 | 1.881492 | 0.712882 | 62.11% | 2.288267 | 1.548141 | 32.34% | +0.004441 |
|
||||
|
||||
因此 attenuation 为 `6/6`；最小 contrast drop 为 `62.11%`，最小 peak drop 为
|
||||
`32.34%`，都高于冻结的 `20%` 门槛。BPC 三格都低于 `+0.05`，平均
|
||||
`+0.006578` 低于 `+0.03`，质量门为 `4/4`。
|
||||
|
||||
这里的 contrast 下降不等价于“尖峰层被关闭”：它可能由 spike-window 分子下降、
|
||||
27 层 reference 分母上升，或两者共同造成；被联合消融覆盖的 layers 26–28 仍属于
|
||||
这个分母。
|
||||
|
||||
## 3. 完整性与复现
|
||||
|
||||
- 13 个新 raw 文件完整：12 formal + 1 primary replay；
|
||||
- 每格 `65,536,000` target bytes，新处理总量 `851,968,000`；
|
||||
- 三个历史 paired reference 合计 `196,608,000` bytes，未在 Round 08 重跑；
|
||||
- 13/13 raw canonical self-hash、aggregate 与 reproduction self-hash 自洽；
|
||||
- architecture / depth / steps / batch 固定为 Block / 32 / 8,000 / 32；
|
||||
- initial public/mixer state、schedule、validation、diagnostic、input-gate、model、
|
||||
optimizer 与 scientific environment 对同 seed historical reference 配对 exact；
|
||||
- primary seed 1 的 formal / replay scientific payload exact：
|
||||
`b85563ca5cb53e60b39c3801d372376206105b8a089a8633b3e81973a7f0c051`。
|
||||
|
||||
四个正式 selector 的 visit、source count、uniform arithmetic 与 reachability 均通过。
|
||||
被选 mixer 的 query / key norm 留在 optimizer param groups，但 forward 不再调用它们：
|
||||
gradient hook 为 0、Adam state 不存在、最终 tensor 与初始值 byte-exact。未选 mixer 的
|
||||
optimizer-state 检查是比科学协议更强的实现审计，不参与主 status。
|
||||
|
||||
## 4. 描述性 non-additivity
|
||||
|
||||
冻结的 bookkeeping residual 为：
|
||||
|
||||
```text
|
||||
I67 = ln(Xref / X67) - ln(Xref / X6) - ln(Xref / X7)
|
||||
```
|
||||
|
||||
step 8,000 的三 seed 平均为：
|
||||
|
||||
- spike contrast：`-0.367038`
|
||||
- peak / mean：`-0.170448`
|
||||
|
||||
它来自三套独立训练，只能描述 joint run 与两个 single runs 的 log-effect 残差；不能
|
||||
写成因果 interaction、Shapley contribution 或“group 6/7 互相抑制”的机制结论。
|
||||
|
||||
## 5. 两阶段独立审阅
|
||||
|
||||
结果前 Grok 实现审阅指出 smoke-only empty selector 的 visit census 可空真。正式
|
||||
4×3 路径全部是非空 selector，因此不影响 raw formal 数值；矩阵结束后已删除 early
|
||||
return、加入 `forward_calls > 0`，并重跑 step-zero、parent smoke、wrapper smoke
|
||||
与 equivalence。修补后的 learned wrapper 实际执行 39 次 forward，父/包装器 15 组
|
||||
科学字段仍全部 exact。
|
||||
|
||||
结果后 Grok 在只读 sandbox 中从 13 个 raw 与三个 historical references 独立复算
|
||||
identity、自哈希、selector、pairing、两项主指标、BPC、replay 与 `I67`。它报告
|
||||
`0 mismatch`、`blocking_errors=0`，确认 analyzer status 与 claim boundary。
|
||||
|
||||
## 6. 最终 claim boundary
|
||||
|
||||
可以说：在这一固定缩小协议内，联合 group 6+7 的 train-time uniform-forward
|
||||
architecture ablation 相对历史同 seed reference 达到预注册 attenuation + BPC 门控。
|
||||
|
||||
不能说：
|
||||
|
||||
- 已定位真实 Kimi-K3 的训练尖峰；
|
||||
- 已复现 K3 报告 Figure 5(c)；
|
||||
- 已把 forward、natural backward 与 optimizer update 分离成纯因果效应；
|
||||
- 已证明下游能力等价、总体统计显著性、可加性或因果 interaction。
|
||||
@@ -0,0 +1,89 @@
|
||||
# Round 08 runner / analyzer 实现审阅与处置
|
||||
|
||||
审阅日期：2026-07-30
|
||||
Grok 会话：`019fb1f0-bb24-7202-8123-295edda7518f`
|
||||
身份：**正式文件完成前的外部模型只读实现审计，不是结果或论文证据**
|
||||
|
||||
## 1. 审阅边界
|
||||
|
||||
Grok Headless 只读检查：
|
||||
|
||||
- 冻结协议与 manifest；
|
||||
- `train.py` / `verify.py` / `analyze.py` / `run_matrix.py`；
|
||||
- 父 runner 的 salt、schema 与 DepthMixer 算术路径。
|
||||
|
||||
明确禁止读取 Round 08 raw formal results、运行训练、修改文件、web search 与 subagents。
|
||||
|
||||
## 2. 对正式 4×3 路径的确认
|
||||
|
||||
审阅确认：
|
||||
|
||||
- 新 protocol ID 没有进入父 `window_start` salt；
|
||||
- selected 路径用 constant-zero FP32 logits 和父 `softmax+einsum` kernel；
|
||||
- 非空 selector 的 query / key norm hook、AdamW state 与 final=initial gate 自洽；
|
||||
- historical pairing 比较的字段在父 JSON 中真实存在；
|
||||
- replay payload 正确排除 run kind、timing、path、parent self hash 与 GPU/version metadata，
|
||||
同时保留模型、optimizer、diagnostics、history、selector 与 reachability；
|
||||
- analyzer 强制 4×3 identity、step 8,000 主判定、1-based layers 21–25、finite/positive、
|
||||
`D=(ref-variant)/ref`、BPC `variant-ref` 与 descriptive `I67`；
|
||||
- 12 formal + 1 replay、65,536,000 bytes/cell 与最大两进程算术正确。
|
||||
|
||||
没有发现会改变正在运行的四个非空 formal variants 数值语义的 blocking error。
|
||||
|
||||
## 3. Blocking finding：smoke-only empty selector 的空审计
|
||||
|
||||
`learned_reference` 直接调用父 `forward`，没有递增 wrapper 的 visit counters。因此：
|
||||
|
||||
```text
|
||||
forward_calls = 0
|
||||
all depth/output visits = 0
|
||||
visit_gate = all(0 == 0) = true
|
||||
selected reachability = all([]) = true
|
||||
```
|
||||
|
||||
这不会影响四个正式变体，它们全部是非空 selector；而 empty selector 另有 20-step
|
||||
parent-equivalence gate，模型/optimizer/evaluations/diagnostics/history/hash 已 exact。
|
||||
但 `forward_intervention.passed` 本身不能在修复前被当作 learned census 证据。
|
||||
|
||||
处置：
|
||||
|
||||
- formal matrix 完成后，删除 `selected_indices` 为空时的父路径 early return；
|
||||
- 让 empty selector 也走同一 copied forward，其中 `mix` 对 64 个节点逐个调用父
|
||||
`DepthMixer.forward`；
|
||||
- 加入 `forward_calls > 0`；
|
||||
- 重跑 step-0、parent smoke、wrapper smoke 与 parent equivalence；
|
||||
- 只有 copied path 仍逐字段 exact 才保留。
|
||||
|
||||
这项修复只强化 smoke audit，不更改任何正式 variant 的 selector 或 forward。
|
||||
|
||||
## 4. Non-blocking findings 与处置
|
||||
|
||||
| finding | 处置 |
|
||||
|---|---|
|
||||
| `smoke_compare` 整体比较 `manifest.path` | 规范化 path 后比较 scientific manifest fields |
|
||||
| unselected gate 额外要求 optimizer state | 保留为强实现 gate，但在文档中标成 protocol 之外的额外审计，不用于科学 status |
|
||||
| GPU/version 差异只有 `metadata_equal`，没有 warnings 数组 | aggregate 增加显式 metadata warnings |
|
||||
| `run_matrix` 失败后只 terminate、不 wait/kill，invalid file 会阻塞重跑 | 加入 terminate→wait→kill 清理，并在 cell failure 时标明 exact invalid target |
|
||||
| step-0 CE 只比较 Python float | 追加 scalar tensor SHA-256 |
|
||||
| `EXPECTED_SOURCE_COUNTS` 未使用 | 用于 runner↔manifest 交叉校验 |
|
||||
|
||||
上述修订不读取结果、不改变冻结阈值或主公式。
|
||||
|
||||
## 5. 矩阵结束后的处置结果
|
||||
|
||||
13 个单元全部退出后才应用上述修订；四个正式非空 selector 的 raw 文件未被重写。
|
||||
修补后的前置闸门结果：
|
||||
|
||||
- step-zero 五个 variant 的 logits、loss tensor、evaluation 与 diagnostic exact；
|
||||
- smoke-only learned wrapper 实际执行 `39` 次 forward，64 个 depth mixer 与
|
||||
output mixer 的 visit census 全部非零且 exact；
|
||||
- parent 与 wrapper 的 architecture、seed、schedule、model、optimizer、hash、
|
||||
evaluation、diagnostic、history、gradient gate 与 scientific environment 共
|
||||
15 组字段全部 exact；
|
||||
- primary smoke 的 selected 参数仍为 0 hook、无 optimizer state、final=initial；
|
||||
- unselected optimizer-state 条件继续作为额外实现闸门，不进入科学 status。
|
||||
|
||||
结果后 Grok 会话 `019fb28b-a9e1-7643-8e43-06f5e16a2077` 在只读 sandbox 中独立
|
||||
复算 13 个 raw、三个 historical references 与 aggregate，报告
|
||||
`blocking_errors=0`、`status_confirmed=true`、`replay_confirmed=true`。详细数字与
|
||||
claim boundary 见 `research/K3_ATTNRES_FORWARD_TRAINING_AUDIT.md`。
|
||||
@@ -228,7 +228,7 @@ if (numeric(reliability.initial.passAt) <= numeric(reliability.k2.passAt) || num
|
||||
if (reliability.nonIdempotent.sideRisk === "LOW") failures.push("非幂等写操作风险没有提升");
|
||||
if (numeric(rl.wait.utilization) >= numeric(rl.full.utilization) || numeric(rl.wait.lostWork) <= numeric(rl.full.lostWork)) failures.push("wait-all 长尾/重算方向异常");
|
||||
if (!rl.wait.takeaway.includes("wait-all") || rl.keyboardSelected !== "rl" || rl.keyboardVisible !== "rl") failures.push("长程 RL 解释或键盘导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || papers.total !== 486 || !papers.hasAgentFilter || papers.agentVisible < 52) failures.push("论文库 Agent 标签或论文总数异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (exceptions.length) failures.push(`浏览器异常：${exceptions.join(" | ")}`);
|
||||
|
||||
@@ -226,7 +226,7 @@ if (!update.steps[0].includes("Fixed preference")) failures.push("DPO 更新流
|
||||
if (!recipe.family.includes("Multi-effort") || !recipe.regime.includes("9 RL experts") || !recipe.constraints.includes("verbosity")) failures.push("K3 配方合同异常");
|
||||
if (!recipe.path.some((step) => step.includes("3 domains × 3 efforts")) || !recipe.path.some((step) => step.includes("MOPD"))) failures.push("K3 配方路径异常");
|
||||
if (recipe.keyboardSelected !== "recipe" || recipe.keyboardVisible !== "recipe") failures.push("实验 tab 键盘导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || papers.total !== 486 || !papers.hasAlignmentFilter || papers.alignmentVisible < 35) failures.push("论文库后训练标签或论文总数异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (exceptions.length) failures.push(`浏览器异常：${exceptions.join(" | ")}`);
|
||||
|
||||
@@ -234,7 +234,7 @@ if (layout.navLinks !== 20 || mobile.mobileLinks !== 20 || home.navLinks !== 20)
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞：${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") {
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") {
|
||||
failures.push("首页 Transformer 新章入口异常");
|
||||
}
|
||||
if (home.paperCount !== "486") failures.push(`首页论文总数异常：${home.paperCount}`);
|
||||
|
||||
@@ -1318,7 +1318,7 @@ if (completionDepth.tasks.panel !== "tasks" || completionDepth.tasks.mathCards !
|
||||
if (completionDepth.hidden.panel !== "hidden" || completionDepth.hidden.stages !== 29 || completionDepth.hidden.selected !== "layer_07" || completionDepth.hidden.points !== 29 || completionDepth.hidden.exact === "1,537 / 1,537" || numeric(completionDepth.hidden.relative) <= 0) failures.push("29 阶段隐藏状态曲线或交互异常");
|
||||
if (completionDepth.router.panel !== "router" || completionDepth.router.layers !== 26 || completionDepth.router.selected !== "layer 24" || completionDepth.router.points !== 26 || !completionDepth.router.ordered.includes("%") || !completionDepth.router.setExact.includes("%") || numeric(completionDepth.router.tv) <= 0 || completionDepth.router.reproCards !== 4 || completionDepth.router.links !== 4) failures.push("26 层 MoE 路由曲线或复跑证据异常");
|
||||
if (completionDepth.keyboardSelected !== "tasks" || completionDepth.keyboardVisible !== "tasks") failures.push("完成度与全深度实验键盘 tab 导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/" || home.paperCount !== "486") failures.push("首页 K3 首发入口或论文数异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/" || home.paperCount !== "486") failures.push("首页 K3 首发入口或论文数异常");
|
||||
if (papers.total !== 486 || !papers.hasFilter || papers.visible < 20 || !papers.hasCoder || !papers.hasEngram) failures.push("论文库 DeepSeek 聚光异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4 || mobile.artifactTabs !== 13 || mobile.behaviorTabs !== 4 || mobile.behaviorSources !== 16 || mobile.behaviorEdges !== 10 || mobile.behaviorDeviceCells !== 29 || mobile.completionDepthTabs !== 4 || mobile.completionDepthHiddenStages !== 29 || mobile.completionDepthRouterLayers !== 26 || mobile.artifactHeatCells !== 64 || mobile.corpusCohorts !== 3 || mobile.lengthDeltaCards !== 4 || mobile.templateLayers !== 6 || mobile.templateScopes !== 2 || mobile.templateModes !== 2 || mobile.templateDomainCards !== 4 || mobile.templateDepthCells !== 24 || mobile.historyLayers !== 6 || mobile.historyScopes !== 2 || mobile.historyModes !== 2 || mobile.historyEffects !== 3 || mobile.historyDomainCards !== 4 || mobile.historyDepthCells !== 24 || mobile.distanceLayers !== 6 || mobile.distanceScopes !== 2 || mobile.distanceModes !== 2 || mobile.distanceContrasts !== 2 || mobile.distanceDomainCards !== 4 || mobile.distanceDepthCells !== 24 || mobile.boundaryLayers !== 6 || mobile.boundaryScopes !== 2 || mobile.boundaryModes !== 2 || mobile.boundaryContrasts !== 3 || mobile.boundaryTokenCards !== 4 || mobile.boundaryDomainCards !== 4 || mobile.boundaryDepthCells !== 24 || mobile.roleLayers !== 6 || mobile.roleScopes !== 2 || mobile.roleModes !== 2 || mobile.roleContrasts !== 3 || mobile.roleLevelCards !== 4 || mobile.roleDomainCards !== 4 || mobile.roleDepthCells !== 24 || mobile.specialLayers !== 6 || mobile.specialScopes !== 2 || mobile.specialModes !== 2 || mobile.specialContrasts !== 4 || mobile.specialTokenCards !== 4 || mobile.specialDomainCards !== 4 || mobile.specialDepthCells !== 24 || mobile.roleBlockLayers !== 6 || mobile.roleBlockScopes !== 2 || mobile.roleBlockModes !== 2 || mobile.roleBlockEffects !== 3 || mobile.roleBlockMatrixCards !== 4 || mobile.roleBlockDomainCards !== 4 || mobile.roleBlockDepthCells !== 24) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
|
||||
@@ -277,7 +277,7 @@ if (numeric(system.initial.success) <= numeric(system.initial.model) || numeric(
|
||||
if (numeric(system.cheap.success) >= numeric(system.initial.success) || numeric(system.cheap.cost) !== 4) failures.push("低预算没有降低成功率 / 成本");
|
||||
if (numeric(system.locked.unsafe) !== 0 || numeric(system.locked.overrefusal) <= numeric(system.initial.overrefusal)) failures.push("安全壳没有展现危险服从 / 过拒权衡");
|
||||
if (system.keyboardSelected !== "judge" || system.keyboardVisible !== "judge") failures.push("实验键盘 tab 导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || home.topicCount !== "17" || papers.total !== 486 || !papers.hasFilter || papers.visible < 80) failures.push("首页 / 论文库评测索引异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
|
||||
@@ -264,7 +264,7 @@ if (!fleet.k3.avoided.includes("320K") || fleet.k3.shortSlo !== "PROTECTED") fai
|
||||
if (!fleet.failed.state.includes("SECONDARY RE-PREFILL") || !fleet.failed.recompute.includes("FAILED PRIMARY")) failures.push("缓存故障没有触发原子失效后的重算");
|
||||
if (fleet.bursty.shortSlo !== "VIOLATED") failures.push("平均并发阈值没有暴露长请求突发");
|
||||
if (fleet.keyboardSelected !== "phase" || fleet.keyboardVisible !== "phase") failures.push("实验键盘 tab 导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || papers.total !== 486 || !papers.hasFilter || papers.visible !== 46) failures.push("论文库推理服务标签或总数异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
|
||||
@@ -0,0 +1,257 @@
|
||||
import { writeFileSync } from "node:fs";
|
||||
|
||||
const cdpPort = process.env.CDP_PORT ?? "9231";
|
||||
const baseUrl = process.env.SITE_URL ?? "http://127.0.0.1:4330";
|
||||
const pages = await fetch(`http://127.0.0.1:${cdpPort}/json/list`).then((response) => response.json());
|
||||
const page = pages.find((entry) => entry.type === "page");
|
||||
if (!page) throw new Error(`CDP ${cdpPort} 没有可用页面`);
|
||||
|
||||
const socket = new WebSocket(page.webSocketDebuggerUrl);
|
||||
await new Promise((resolve, reject) => {
|
||||
socket.addEventListener("open", resolve, { once: true });
|
||||
socket.addEventListener("error", reject, { once: true });
|
||||
});
|
||||
|
||||
let nextId = 0;
|
||||
const pending = new Map();
|
||||
const exceptions = [];
|
||||
socket.addEventListener("message", (event) => {
|
||||
const message = JSON.parse(event.data);
|
||||
if (message.id && pending.has(message.id)) {
|
||||
const { resolve, reject } = pending.get(message.id);
|
||||
pending.delete(message.id);
|
||||
if (message.error) reject(new Error(message.error.message));
|
||||
else resolve(message.result);
|
||||
}
|
||||
if (message.method === "Runtime.exceptionThrown") {
|
||||
exceptions.push(message.params.exceptionDetails.exception?.description ?? message.params.exceptionDetails.text);
|
||||
}
|
||||
});
|
||||
|
||||
const command = (method, params = {}) => new Promise((resolve, reject) => {
|
||||
const id = ++nextId;
|
||||
pending.set(id, { resolve, reject });
|
||||
socket.send(JSON.stringify({ id, method, params }));
|
||||
});
|
||||
const pause = (milliseconds) => new Promise((resolve) => setTimeout(resolve, milliseconds));
|
||||
const evaluate = async (expression) => {
|
||||
const result = await command("Runtime.evaluate", { expression, returnByValue: true, awaitPromise: true });
|
||||
if (result.exceptionDetails) throw new Error(result.exceptionDetails.exception?.description ?? result.exceptionDetails.text);
|
||||
return result.result.value;
|
||||
};
|
||||
const navigate = async (path) => {
|
||||
await command("Page.navigate", { url: `${baseUrl}${path}` });
|
||||
for (let attempt = 0; attempt < 100; attempt += 1) {
|
||||
await pause(100);
|
||||
if (await evaluate("document.readyState === 'complete'")) return;
|
||||
}
|
||||
throw new Error(`${path} 加载超时`);
|
||||
};
|
||||
const screenshot = async (path) => {
|
||||
const result = await command("Page.captureScreenshot", { format: "png", captureBeyondViewport: false });
|
||||
writeFileSync(path, Buffer.from(result.data, "base64"));
|
||||
};
|
||||
|
||||
await command("Page.enable");
|
||||
await command("Runtime.enable");
|
||||
await command("Emulation.setDeviceMetricsOverride", {
|
||||
width: 1440,
|
||||
height: 1100,
|
||||
deviceScaleFactor: 1,
|
||||
mobile: false,
|
||||
});
|
||||
await navigate("/k3/");
|
||||
|
||||
const desktop = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-forward-lab]");
|
||||
root.scrollIntoView({ block: "start", behavior: "instant" });
|
||||
window.scrollBy(0, -78);
|
||||
const text = (selector) => root.querySelector(selector)?.textContent.replace(/\\s+/g, " ").trim();
|
||||
const panel = () => root.querySelector("[data-forward-panel]:not([hidden])")?.dataset.forwardPanel;
|
||||
const setSelect = (selector, value) => {
|
||||
const node = root.querySelector(selector);
|
||||
node.value = value;
|
||||
node.dispatchEvent(new Event("change", { bubbles: true }));
|
||||
};
|
||||
|
||||
const initial = {
|
||||
panel: panel(),
|
||||
tabs: root.querySelectorAll("[data-forward-tab]").length,
|
||||
panels: root.querySelectorAll("[data-forward-panel]").length,
|
||||
ledger: root.querySelectorAll(".forward-ledger article").length,
|
||||
targetGroups: root.querySelectorAll(".group-map article.target").length,
|
||||
spikeLayers: root.querySelectorAll(".group-map i.spike").length,
|
||||
variants: root.querySelectorAll(".variant-grid article").length,
|
||||
boundary: root.textContent.includes("K3 的训练尖峰已被定位") &&
|
||||
root.textContent.includes("Figure 5(c)") &&
|
||||
root.textContent.includes("训练期架构消融"),
|
||||
};
|
||||
|
||||
root.querySelector('[data-forward-tab="trajectory"]').click();
|
||||
const trajectoryInitial = {
|
||||
panel: panel(),
|
||||
state: text("[data-forward-trajectory-state]"),
|
||||
points: root.querySelectorAll("[data-forward-trajectory-series] circle").length,
|
||||
lines: root.querySelectorAll("[data-forward-trajectory-series] polyline").length,
|
||||
readouts: [...root.querySelectorAll("[data-forward-trajectory-readout] article")]
|
||||
.map((node) => node.textContent.replace(/\\s+/g, " ").trim()),
|
||||
};
|
||||
setSelect("[data-forward-trajectory-seed]", "2026073002");
|
||||
root.querySelector('[data-forward-trajectory-metric="peak_normalized"]').click();
|
||||
const trajectoryChanged = {
|
||||
state: text("[data-forward-trajectory-state]"),
|
||||
readouts: [...root.querySelectorAll("[data-forward-trajectory-readout] article")]
|
||||
.map((node) => node.textContent.replace(/\\s+/g, " ").trim()),
|
||||
};
|
||||
|
||||
root.querySelector('[data-forward-tab="gate"]').click();
|
||||
const gate = {
|
||||
panel: panel(),
|
||||
status: text(".status-banner"),
|
||||
rows: root.querySelectorAll(".gate-table tbody tr").length,
|
||||
passedRows: root.querySelectorAll(".gate-table tbody td.good:last-child").length,
|
||||
quality: text(".gate-layout aside"),
|
||||
claims: root.querySelectorAll(".claim-pair article").length,
|
||||
};
|
||||
|
||||
root.querySelector('[data-forward-tab="interaction"]').click();
|
||||
const interactionInitial = {
|
||||
panel: panel(),
|
||||
state: text("[data-forward-interaction-state]"),
|
||||
cells: root.querySelectorAll("[data-forward-interaction-cells] article").length,
|
||||
summary: text("[data-forward-interaction-summary]"),
|
||||
};
|
||||
root.querySelector('[data-forward-interaction-metric="peak_normalized"]').click();
|
||||
const interactionChanged = {
|
||||
state: text("[data-forward-interaction-state]"),
|
||||
summary: text("[data-forward-interaction-summary]"),
|
||||
};
|
||||
|
||||
root.querySelector('[data-forward-tab="spectrum"]').click();
|
||||
const spectrumInitial = {
|
||||
panel: panel(),
|
||||
state: text("[data-forward-spectrum-state]"),
|
||||
points: root.querySelectorAll("[data-forward-spectrum-points] circle").length,
|
||||
contrast: text("[data-forward-spectrum-contrast]"),
|
||||
peak: text("[data-forward-spectrum-peak]"),
|
||||
layer: text("[data-forward-spectrum-layer]"),
|
||||
audits: root.querySelectorAll(".audit-grid article").length,
|
||||
replay: root.textContent.includes("scientific exact"),
|
||||
};
|
||||
setSelect("[data-forward-spectrum-seed]", "2026073002");
|
||||
setSelect("[data-forward-spectrum-variant]", "uniform_groups_6_7_forward");
|
||||
const spectrumChanged = {
|
||||
state: text("[data-forward-spectrum-state]"),
|
||||
points: root.querySelectorAll("[data-forward-spectrum-points] circle").length,
|
||||
contrast: text("[data-forward-spectrum-contrast]"),
|
||||
peak: text("[data-forward-spectrum-peak]"),
|
||||
layer: text("[data-forward-spectrum-layer]"),
|
||||
};
|
||||
|
||||
const first = root.querySelector('[data-forward-tab="contract"]');
|
||||
first.focus();
|
||||
first.dispatchEvent(new KeyboardEvent("keydown", { key: "ArrowRight", bubbles: true }));
|
||||
const keyboard = {
|
||||
selected: root.querySelector('[data-forward-tab][aria-selected="true"]').dataset.forwardTab,
|
||||
panel: panel(),
|
||||
};
|
||||
|
||||
return {
|
||||
initial, trajectoryInitial, trajectoryChanged, gate,
|
||||
interactionInitial, interactionChanged, spectrumInitial, spectrumChanged, keyboard,
|
||||
documentOverflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
|
||||
rootOverflow: root.scrollWidth - root.clientWidth,
|
||||
};
|
||||
})()`);
|
||||
await pause(180);
|
||||
await screenshot("/tmp/llm-atlas-k3-attnres-forward-desktop.png");
|
||||
|
||||
await command("Emulation.setDeviceMetricsOverride", {
|
||||
width: 390,
|
||||
height: 844,
|
||||
deviceScaleFactor: 1,
|
||||
mobile: true,
|
||||
});
|
||||
await navigate("/k3/");
|
||||
const mobile = await evaluate(`(() => {
|
||||
const root = document.querySelector("[data-forward-lab]");
|
||||
root.scrollIntoView({ block: "start", behavior: "instant" });
|
||||
window.scrollBy(0, -64);
|
||||
root.querySelector('[data-forward-tab="gate"]').click();
|
||||
const table = root.querySelector(".gate-table-wrap");
|
||||
const gateTableScrolls = table.scrollWidth > table.clientWidth;
|
||||
root.querySelector('[data-forward-tab="spectrum"]').click();
|
||||
return {
|
||||
tabs: root.querySelectorAll("[data-forward-tab]").length,
|
||||
visiblePanel: root.querySelector("[data-forward-panel]:not([hidden])")?.dataset.forwardPanel,
|
||||
points: root.querySelectorAll("[data-forward-spectrum-points] circle").length,
|
||||
audits: root.querySelectorAll(".audit-grid article").length,
|
||||
gateTableScrolls,
|
||||
documentOverflow: document.documentElement.scrollWidth - document.documentElement.clientWidth,
|
||||
rootOverflow: root.scrollWidth - root.clientWidth,
|
||||
};
|
||||
})()`);
|
||||
await pause(180);
|
||||
await screenshot("/tmp/llm-atlas-k3-attnres-forward-mobile.png");
|
||||
|
||||
const failures = [];
|
||||
if (desktop.initial.panel !== "contract" || desktop.initial.tabs !== 5 || desktop.initial.panels !== 5 ||
|
||||
desktop.initial.ledger !== 6 || desktop.initial.targetGroups !== 2 ||
|
||||
desktop.initial.spikeLayers !== 5 || desktop.initial.variants !== 4) {
|
||||
failures.push("五视图、账本、group 或 selector map 结构异常");
|
||||
}
|
||||
if (!desktop.initial.boundary) failures.push("reduced-model / K3 / Figure 5(c) claim boundary 缺失");
|
||||
if (desktop.trajectoryInitial.panel !== "trajectory" || desktop.trajectoryInitial.points !== 24 ||
|
||||
desktop.trajectoryInitial.lines !== 4 || desktop.trajectoryInitial.readouts.length !== 4 ||
|
||||
!desktop.trajectoryInitial.state.includes("2026073001") ||
|
||||
!desktop.trajectoryInitial.readouts.some((value) => value.includes("+73.5%"))) {
|
||||
failures.push("seed 1 contrast 训练轨迹异常");
|
||||
}
|
||||
if (!desktop.trajectoryChanged.state.includes("2026073002") ||
|
||||
!desktop.trajectoryChanged.state.includes("PEAK / MEAN") ||
|
||||
!desktop.trajectoryChanged.readouts.some((value) => value.includes("+62.0%"))) {
|
||||
failures.push("trajectory seed / metric 切换异常");
|
||||
}
|
||||
if (desktop.gate.panel !== "gate" || desktop.gate.rows !== 6 || desktop.gate.passedRows !== 6 ||
|
||||
desktop.gate.claims !== 2 || !desktop.gate.status.includes("ATTENUATION ESTABLISHED") ||
|
||||
!desktop.gate.status.includes("6 / 6") || !desktop.gate.status.includes("4 / 4") ||
|
||||
!desktop.gate.quality.includes("+0.0066")) {
|
||||
failures.push("冻结主门或 BPC quality readout 异常");
|
||||
}
|
||||
if (desktop.interactionInitial.panel !== "interaction" || desktop.interactionInitial.cells !== 3 ||
|
||||
!desktop.interactionInitial.state.includes("8,000") ||
|
||||
!desktop.interactionInitial.summary.includes("-0.367") ||
|
||||
!desktop.interactionChanged.state.includes("PEAK / MEAN") ||
|
||||
!desktop.interactionChanged.summary.includes("-0.170")) {
|
||||
failures.push("I67 描述性 residual 切换异常");
|
||||
}
|
||||
if (desktop.spectrumInitial.panel !== "spectrum" || desktop.spectrumInitial.points !== 32 ||
|
||||
desktop.spectrumInitial.contrast !== "3.046×" || desktop.spectrumInitial.peak !== "3.219×" ||
|
||||
desktop.spectrumInitial.layer !== "L21" || desktop.spectrumInitial.audits !== 5 ||
|
||||
!desktop.spectrumInitial.replay) {
|
||||
failures.push("reference 32 层谱或 replay 审计异常");
|
||||
}
|
||||
if (desktop.spectrumChanged.points !== 32 || !desktop.spectrumChanged.state.includes("2026073002") ||
|
||||
!desktop.spectrumChanged.state.includes("GROUPS 6+7") ||
|
||||
desktop.spectrumChanged.contrast !== "0.758×" || desktop.spectrumChanged.peak !== "1.375×" ||
|
||||
desktop.spectrumChanged.layer !== "L9") {
|
||||
failures.push("joint variant spectrum 切换异常");
|
||||
}
|
||||
if (desktop.keyboard.selected !== "trajectory" || desktop.keyboard.panel !== "trajectory") {
|
||||
failures.push("键盘 tab 切换异常");
|
||||
}
|
||||
if (desktop.documentOverflow > 1 || mobile.documentOverflow > 1) failures.push("页面出现文档级横向溢出");
|
||||
if (desktop.rootOverflow > 1 || mobile.rootOverflow > 1) failures.push("Round 08 实验室出现横向溢出");
|
||||
if (mobile.tabs !== 5 || mobile.visiblePanel !== "spectrum" || mobile.points !== 32 ||
|
||||
mobile.audits !== 5 || !mobile.gateTableScrolls) {
|
||||
failures.push("390px 移动端布局或局部表格滚动异常");
|
||||
}
|
||||
if (exceptions.length) failures.push(`运行时异常：${exceptions.join(" | ")}`);
|
||||
|
||||
console.log(JSON.stringify({ desktop, mobile, exceptions }, null, 2));
|
||||
socket.close();
|
||||
if (failures.length) {
|
||||
console.error(`FAIL K3 AttnRes forward-training browser\n- ${failures.join("\n- ")}`);
|
||||
process.exit(1);
|
||||
}
|
||||
console.log("PASS K3 AttnRes forward-training browser interactions");
|
||||
@@ -0,0 +1,106 @@
|
||||
import { createHash } from "node:crypto";
|
||||
import { readdirSync, readFileSync } from "node:fs";
|
||||
|
||||
const hash = (bytes) => createHash("sha256").update(bytes).digest("hex");
|
||||
const read = (path) => {
|
||||
const bytes = readFileSync(new URL(path, import.meta.url));
|
||||
return { bytes, json: JSON.parse(bytes), sha256: hash(bytes) };
|
||||
};
|
||||
const aggregate = read("../src/data/k3-attnres-forward.json");
|
||||
const compact = read("../src/data/k3-attnres-forward-compact.json");
|
||||
const reproduction = read("../experiments/k3/attnres_forward/reproduction.json");
|
||||
const manifest = read("../experiments/k3/attnres_forward/manifest.json");
|
||||
const rawDirectory = new URL("../experiments/k3/attnres_forward/results/raw/", import.meta.url);
|
||||
const failures = [];
|
||||
const expect = (condition, message) => {
|
||||
if (!condition) failures.push(message);
|
||||
};
|
||||
const close = (actual, expected, tolerance = 1e-15) =>
|
||||
Math.abs(actual - expected) <= tolerance;
|
||||
|
||||
expect(aggregate.sha256 === "f8df928adb8a563d851bb3c1abbf40bcada33626d9177821e4341d854980a097", "aggregate physical SHA-256 changed");
|
||||
expect(compact.sha256 === "664f6d6226df7c0c9aba6314d54a1cb8ae90f6922016dfa7c7a823729606f0c1", "compact physical SHA-256 changed");
|
||||
expect(reproduction.sha256 === "b029c333596dd1de957efc2b42f7ebfc1d32f241c687f06021c4d08ffa27b1d0", "reproduction physical SHA-256 changed");
|
||||
expect(manifest.sha256 === "49546ed5baf36bcb30885062b7c671fafe4ccff2e606624cd7dbe23717f9a712", "manifest physical SHA-256 changed");
|
||||
|
||||
expect(aggregate.json.canonical_sha256_without_self === "eecf05c623e473ec5eba8afa54d555d2733a0af2fc5488398f1da095206d50c8", "aggregate canonical SHA-256 changed");
|
||||
expect(compact.json.canonical_sha256_without_self === "c3e672adb6879f99efccf3bd2e0fabaab1a1159225a8ddaf4f5f9ea7a002784b", "compact canonical SHA-256 changed");
|
||||
expect(reproduction.json.canonical_sha256_without_self === "57346df80c0d76bd1d306d5fa213feed74ac2094237c22ebcb49f16ea16437e0", "reproduction canonical SHA-256 changed");
|
||||
|
||||
expect(compact.json.protocol_id === "llm-atlas-k3-attnres-forward-training-v1", "protocol identity mismatch");
|
||||
expect(compact.json.status === "forward_training_attenuation_established_within_reduced_protocol", "frozen status changed");
|
||||
expect(compact.json.primary_step === 8000, "primary step changed");
|
||||
expect(compact.json.spike_layers_1based.join(",") === "21,22,23,24,25", "spike window changed");
|
||||
expect(compact.json.thresholds.material_relative_drop === 0.2, "attenuation threshold changed");
|
||||
expect(compact.json.processed_target_bytes.formal_12_cells === 786432000, "formal target bytes changed");
|
||||
expect(compact.json.processed_target_bytes.primary_replay === 65536000, "replay target bytes changed");
|
||||
expect(compact.json.processed_target_bytes.total === 851968000, "new target bytes changed");
|
||||
expect(compact.json.aggregate_sha256 === aggregate.json.canonical_sha256_without_self, "compact→aggregate canonical link mismatch");
|
||||
expect(reproduction.json.scientific_payload_sha256 === compact.json.replay.scientific_payload_sha256, "reproduction→replay hash link mismatch");
|
||||
expect(reproduction.json.passed && compact.json.replay.passed, "full replay is not exact");
|
||||
expect(reproduction.json.post_result_grok_review.blocking_errors === 0, "post-result audit reports a blocking error");
|
||||
expect(reproduction.json.post_result_grok_review.status_confirmed, "post-result audit did not confirm status");
|
||||
expect(reproduction.json.post_result_grok_review.replay_confirmed, "post-result audit did not confirm replay");
|
||||
expect(compact.json.metadata_warnings.length === 0, "historical pairing metadata warning appeared");
|
||||
|
||||
const rawNames = readdirSync(rawDirectory).filter((name) => name.endsWith(".json")).sort();
|
||||
expect(rawNames.length === 13, "raw matrix is not 12 formal + 1 replay");
|
||||
const expectedRaw = [
|
||||
...aggregate.json.input_files.formal.map((item) => item),
|
||||
{ ...aggregate.json.input_files.replay, variant: "uniform_groups_6_7_forward", seed: 2026073001 },
|
||||
];
|
||||
for (const item of expectedRaw) {
|
||||
const name = item.path.split("/").at(-1);
|
||||
const raw = read(`../experiments/k3/attnres_forward/results/raw/${name}`);
|
||||
expect(raw.sha256 === item.sha256, `${name} physical SHA-256 mismatch`);
|
||||
expect(/^[0-9a-f]{64}$/.test(raw.json.canonical_sha256_without_self), `${name} canonical self-hash missing`);
|
||||
expect(raw.json.forward_intervention.passed, `${name} forward audit failed`);
|
||||
expect(raw.json.forward_intervention.forward_calls === 8054, `${name} forward census changed`);
|
||||
expect(raw.json.target_bytes_seen === 65536000, `${name} target bytes changed`);
|
||||
}
|
||||
|
||||
const primary = compact.json.primary;
|
||||
expect(primary.material_response_passed, "primary material response failed");
|
||||
expect(primary.passed_cells === 6 && primary.required_cells === 6, "primary attenuation is not 6/6");
|
||||
expect(primary.cells.every((cell) => cell.relative_drop >= 0.2 && cell.passed), "a primary attenuation cell fell below 20%");
|
||||
expect(primary.quality.passed && primary.quality.passed_checks === 4, "BPC quality gate is not 4/4");
|
||||
expect(close(primary.quality.mean_delta_bpc, 0.0065784582165467525), "mean BPC delta changed");
|
||||
expect(close(Math.min(...primary.cells.filter((cell) => cell.metric === "spike_contrast").map((cell) => cell.relative_drop)), 0.6211080443849011), "minimum contrast drop changed");
|
||||
expect(close(Math.min(...primary.cells.filter((cell) => cell.metric === "peak_normalized").map((cell) => cell.relative_drop)), 0.3234439150908039), "minimum peak drop changed");
|
||||
expect(Object.values(compact.json.secondary_status).every((value) => value === "secondary_material_response"), "secondary status changed");
|
||||
expect(compact.json.trajectories.length === 90, "trajectory cell count changed");
|
||||
expect(compact.json.final_spectra.length === 15, "final spectrum count changed");
|
||||
expect(compact.json.final_spectra.every((item) => item.normalized.length === 32), "a final spectrum is not 32 layers");
|
||||
expect(compact.json.interaction.cells.length === 36, "interaction cell count changed");
|
||||
expect(compact.json.interaction.summaries.length === 12, "interaction summary count changed");
|
||||
expect(close(
|
||||
compact.json.interaction.summaries.find((item) =>
|
||||
item.step === 8000 && item.metric === "spike_contrast").mean_interaction_residual,
|
||||
-0.3670379672721313,
|
||||
), "final contrast I67 changed");
|
||||
expect(close(
|
||||
compact.json.interaction.summaries.find((item) =>
|
||||
item.step === 8000 && item.metric === "peak_normalized").mean_interaction_residual,
|
||||
-0.17044758609080035,
|
||||
), "final peak I67 changed");
|
||||
|
||||
if (failures.length) {
|
||||
console.error(`FAIL K3 AttnRes forward-training data\n- ${failures.join("\n- ")}`);
|
||||
process.exit(1);
|
||||
}
|
||||
|
||||
console.log(JSON.stringify({
|
||||
protocol: compact.json.protocol_id,
|
||||
status: compact.json.status,
|
||||
rawRuns: rawNames.length,
|
||||
attenuation: `${primary.passed_cells}/${primary.required_cells}`,
|
||||
quality: `${primary.quality.passed_checks}/${primary.quality.required_checks}`,
|
||||
replayExact: compact.json.replay.passed,
|
||||
bytes: compact.json.processed_target_bytes,
|
||||
hashes: {
|
||||
aggregate: aggregate.sha256,
|
||||
compact: compact.sha256,
|
||||
reproduction: reproduction.sha256,
|
||||
},
|
||||
}, null, 2));
|
||||
console.log("PASS K3 AttnRes forward-training frozen data");
|
||||
@@ -88,6 +88,10 @@ const overview = await evaluate(`(() => ({
|
||||
localPathTabs: document.querySelectorAll("[data-local-tab]").length,
|
||||
localPathPanels: document.querySelectorAll("[data-local-panel]").length,
|
||||
localPathVerdict: document.querySelector("#attnres-local-path")?.textContent.includes("localization 未建立"),
|
||||
forwardTabs: document.querySelectorAll("[data-forward-tab]").length,
|
||||
forwardPanels: document.querySelectorAll("[data-forward-panel]").length,
|
||||
forwardVerdict: document.querySelector("#attnres-forward")?.textContent.includes("6 / 6 PASS") &&
|
||||
document.querySelector("#attnres-forward")?.textContent.includes("reduced protocol only"),
|
||||
nativeVisionCorrected: document.body.textContent.includes("MoonViT‑V2 从头训练") &&
|
||||
document.body.textContent.includes("同一个 next-token prediction objective"),
|
||||
staleVisionClaim: document.body.textContent.includes("先固定语言模型训练视觉组件"),
|
||||
@@ -300,8 +304,9 @@ const mobile = await evaluate(`(() => {
|
||||
gradientTabs: document.querySelectorAll("[data-gradient-tab]").length,
|
||||
spikeTabs: document.querySelectorAll("[data-spike-tab]").length,
|
||||
localPathTabs: document.querySelectorAll("[data-local-tab]").length,
|
||||
forwardTabs: document.querySelectorAll("[data-forward-tab]").length,
|
||||
offenders: [...document.querySelectorAll("body *")]
|
||||
.filter((node) => !node.closest(".paper-chain, .spec-table-wrap, .cache-strip, .architecture-explorer, [data-k3-lab], [data-k3-artifact-lab], [data-attnres-lab], [data-gradient-lab], [data-spike-lab], [data-local-path-lab]"))
|
||||
.filter((node) => !node.closest(".paper-chain, .spec-table-wrap, .cache-strip, .architecture-explorer, [data-k3-lab], [data-k3-artifact-lab], [data-attnres-lab], [data-gradient-lab], [data-spike-lab], [data-local-path-lab], [data-forward-lab]"))
|
||||
.filter((node) => node.getBoundingClientRect().right > document.documentElement.clientWidth + 1)
|
||||
.slice(0, 15)
|
||||
.map((node) => ({
|
||||
@@ -332,7 +337,7 @@ console.log(JSON.stringify(report, null, 2));
|
||||
const numeric = (text) => Number.parseFloat(text.replaceAll(",", "").replace("−", "-"));
|
||||
const failures = [];
|
||||
if (!overview.title.includes("因果环节")) failures.push("K3 二轮标题异常");
|
||||
if (overview.sections !== 36 || overview.tocLinks !== 36) failures.push("35 个编号专题加阅读链的目录结构异常");
|
||||
if (overview.sections !== 37 || overview.tocLinks !== 37) failures.push("36 个编号专题加阅读链的目录结构异常");
|
||||
if (overview.ledgers !== 32 || overview.reportMap !== 9) failures.push("32 张问题账或报告地图异常");
|
||||
if (overview.figureAtlas !== 21 || overview.paperLinks !== 100 || overview.paperGroups < 12) failures.push("图表审计或 100 节点阅读链异常");
|
||||
if (overview.labTabs !== 8 || overview.labPanels !== 8) failures.push("八联实验结构异常");
|
||||
@@ -341,6 +346,7 @@ if (overview.attnresTabs !== 5 || overview.attnresPanels !== 5) failures.push("A
|
||||
if (overview.gradientTabs !== 5 || overview.gradientPanels !== 5) failures.push("AttnRes 梯度定义扩展五视图异常");
|
||||
if (overview.spikeTabs !== 5 || overview.spikePanels !== 5) failures.push("AttnRes 尖峰路径五视图异常");
|
||||
if (overview.localPathTabs !== 5 || overview.localPathPanels !== 5 || !overview.localPathVerdict) failures.push("AttnRes 局部路径五视图或冻结判定异常");
|
||||
if (overview.forwardTabs !== 5 || overview.forwardPanels !== 5 || !overview.forwardVerdict) failures.push("AttnRes 训练期前向五视图或冻结判定异常");
|
||||
if (!overview.nativeVisionCorrected || overview.staleVisionClaim) failures.push("原生多模态纠错未生效或旧错误残留");
|
||||
if (overview.documentOverflow > 1 || mobile.documentOverflow > 1) failures.push("桌面或移动端存在文档级横向溢出");
|
||||
if (labs.memoryInitial.panel !== "memory" || numeric(labs.memoryInitial.additiveError) <= numeric(labs.memoryInitial.deltaError)) failures.push("Delta memory 初始递推异常");
|
||||
@@ -365,7 +371,7 @@ if (artifacts.parameterChanged.shape !== "[96,128] F32" || !artifacts.parameterC
|
||||
if (artifacts.reproductionInitial.panel !== "reproduction" || numeric(artifacts.reproductionInitial.speedup) !== 1.85 || numeric(artifacts.reproductionInitial.localMean) < 2.6 || !artifacts.reproductionInitial.exactSuite || numeric(artifacts.reproductionInitial.cv) < 2) failures.push("FlashKDA H20、本机 exact suite 或 router 初始探针异常");
|
||||
if (numeric(artifacts.reproductionChanged.speedup) !== 3.27 || numeric(artifacts.reproductionChanged.flash) !== 0.7064 || numeric(artifacts.reproductionChanged.localMean) >= numeric(artifacts.reproductionInitial.localMean) || !artifacts.reproductionChanged.localMode.includes("FP32 state") || numeric(artifacts.reproductionChanged.cv) <= numeric(artifacts.reproductionInitial.cv) || numeric(artifacts.reproductionChanged.zero) <= numeric(artifacts.reproductionInitial.zero)) failures.push("GB200 benchmark、本机 varlen/state 或 synthetic router counterexample 未更新");
|
||||
if (artifacts.keyboardSelected !== "tensors" || artifacts.keyboardVisible !== "tensors") failures.push("开放工件键盘 tab 导航异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 8 || mobile.artifactTabs !== 4 || mobile.artifactLayers !== 93 || mobile.attnresTabs !== 5 || mobile.gradientTabs !== 5 || mobile.spikeTabs !== 5 || mobile.localPathTabs !== 5) failures.push("移动端导航或实验异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 8 || mobile.artifactTabs !== 4 || mobile.artifactLayers !== 93 || mobile.attnresTabs !== 5 || mobile.gradientTabs !== 5 || mobile.spikeTabs !== 5 || mobile.localPathTabs !== 5 || mobile.forwardTabs !== 5) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
if (exceptions.length) failures.push(`浏览器异常：${exceptions.join(" | ")}`);
|
||||
|
||||
|
||||
@@ -246,7 +246,7 @@ if (ocr.unreported.status !== "OUT OF EVIDENCE" || ocr.unreported.accuracy !== "
|
||||
if (loop.toolsStart.state !== "OPEN" || loop.toolsEnd.state !== "VERIFIED" || loop.toolsEnd.evidence !== "97%" || loop.toolsEnd.tools !== "3") failures.push("vision-in-the-loop 终局异常");
|
||||
if (loop.cotEnd.state !== "FAILED" || !loop.cotEnd.takeaway.includes("不能凭空增加")) failures.push("文字 CoT 与新观察没有分开");
|
||||
if (loop.keyboardSelected !== "connector" || loop.keyboardVisible !== "connector") failures.push("实验键盘 tab 导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || papers.total !== 486 || !papers.hasFilter || papers.multimodalVisible < 59) failures.push("论文库多模态标签或总数异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
|
||||
@@ -277,7 +277,7 @@ if (layout.navLinks !== 20 || mobile.mobileLinks !== 20 || home.navLinks !== 20)
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞：${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") {
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") {
|
||||
failures.push("首页 Transformer 新章入口异常");
|
||||
}
|
||||
if (home.paperCount !== "486") failures.push(`首页论文总数异常：${home.paperCount}`);
|
||||
|
||||
@@ -289,7 +289,7 @@ if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentO
|
||||
}
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞：${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强")) failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后")) failures.push("首页 K3 首发入口异常");
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常：${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
@@ -288,7 +288,7 @@ if (numeric(residual.attnres.states) !== 9 || !residual.attnres.routeExplain.inc
|
||||
if (!residual.clamp.activation.includes("V4") || !residual.clamp.bound.includes("100")) failures.push("DeepSeek-V4 clamp 展示异常");
|
||||
if (!residual.situ.activation.includes("KIMI") || !residual.situ.bound.includes("100")) failures.push("K3 SiTU 上界展示异常");
|
||||
if (residual.keyboardSelected !== "position" || residual.keyboardVisible !== "position") failures.push("实验键盘 tab 导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || home.topicCount !== "17" || papers.total !== 486 || !papers.hasFilter || papers.visible < 30) failures.push("首页 / 论文库表示索引异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (mobile.offenders.length) failures.push(`移动端越界元素：${JSON.stringify(mobile.offenders)}`);
|
||||
|
||||
@@ -273,7 +273,7 @@ if (layout.navLinks !== 20 || mobile.mobileLinks !== 20 || home.navLinks !== 20)
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞：${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") {
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") {
|
||||
failures.push("首页 Transformer 新章入口异常");
|
||||
}
|
||||
if (home.paperCount !== "486") failures.push(`首页论文总数异常：${home.paperCount}`);
|
||||
|
||||
@@ -233,7 +233,7 @@ if (layout.articleSections !== 16 || layout.paperLinks !== 37 || layout.labTabs
|
||||
if (layout.documentOverflow > 0 || mobile.documentOverflow > 0 || home.documentOverflow > 0) failures.push("页面存在横向溢出");
|
||||
if (layout.navGap < 0) failures.push(`桌面导航碰撞：${layout.navGap}px`);
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true") failures.push("移动端菜单不可用");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强")) failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后")) failures.push("首页 K3 首发入口异常");
|
||||
if (exceptions.length) failures.push(`浏览器脚本异常：${exceptions.join("; ")}`);
|
||||
|
||||
socket.close();
|
||||
|
||||
@@ -236,7 +236,7 @@ if (block.family.trim() !== "Hybrid MoE" || !block.kv.includes("3 KDA : 1 Gated
|
||||
if (!block.path.some((step) => step.includes("KDA × 3")) || !block.note.includes("AttnRes")) failures.push("K3 Block 路径异常");
|
||||
if (block.context.trim() !== "128K" || numeric(block.mha) !== 400 || numeric(block.kda) !== 1) failures.push("KV 成本缩放异常");
|
||||
if (block.keyboardSelected !== "block" || block.keyboardVisible !== "block") failures.push("实验 tab 键盘导航异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("16 个局部 mixer 单侧证据很强") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.releaseCards !== 17 || !home.firstRelease.includes("局部 uniform routing 进入完整训练后") || home.firstHref !== "/k3/") failures.push("首页 K3 首发入口异常");
|
||||
if (home.paperCount !== "486" || papers.total !== 486 || papers.transformerVisible < 30) failures.push("论文库或首页论文数量异常");
|
||||
if (!mobile.menuVisible || mobile.menuOpen !== "true" || mobile.tabs !== 4) failures.push("移动端导航或实验异常");
|
||||
if (exceptions.length) failures.push(`浏览器异常：${exceptions.join(" | ")}`);
|
||||
|
||||
@@ -0,0 +1,703 @@
|
||||
---
|
||||
import rawLab from "@/data/k3-attnres-forward-compact.json";
|
||||
|
||||
const lab = rawLab as any;
|
||||
const json = JSON.stringify(lab).replaceAll("<", "\\u003c");
|
||||
const seeds = [...new Set(lab.final_spectra.map((item: any) => item.seed))] as number[];
|
||||
const variants = [
|
||||
"uniform_group_6_forward",
|
||||
"uniform_group_7_forward",
|
||||
"uniform_groups_6_7_forward",
|
||||
"uniform_group_7_mlp_forward",
|
||||
];
|
||||
const variantLabels: Record<string, string> = {
|
||||
learned_reference: "HISTORICAL LEARNED REFERENCE",
|
||||
uniform_group_6_forward: "GROUP 6 · UNIFORM FORWARD",
|
||||
uniform_group_7_forward: "GROUP 7 · UNIFORM FORWARD",
|
||||
uniform_groups_6_7_forward: "GROUPS 6+7 · UNIFORM FORWARD",
|
||||
uniform_group_7_mlp_forward: "GROUP 7 MLP · UNIFORM FORWARD",
|
||||
};
|
||||
const statusLabels: Record<string, string> = {
|
||||
forward_training_attenuation_established_within_reduced_protocol:
|
||||
"ATTENUATION ESTABLISHED · REDUCED PROTOCOL",
|
||||
attenuation_not_established: "ATTENUATION NOT ESTABLISHED",
|
||||
quality_guard_failed: "QUALITY GUARD FAILED",
|
||||
attenuation_and_quality_failed: "ATTENUATION + QUALITY FAILED",
|
||||
};
|
||||
const statusLabel = statusLabels[lab.status] ?? lab.status;
|
||||
const primary = lab.primary;
|
||||
const quality = primary.quality;
|
||||
const meanDrop = (effect: any, metric: string) => {
|
||||
const cells = effect.cells.filter((cell: any) => cell.metric === metric);
|
||||
return cells.reduce((sum: number, cell: any) => sum + cell.relative_drop, 0) / cells.length;
|
||||
};
|
||||
const percent = (value: number) => `${value >= 0 ? "+" : "−"}${Math.abs(value * 100).toFixed(1)}%`;
|
||||
const shortHash = (value: string) => `${value.slice(0, 10)}…${value.slice(-8)}`;
|
||||
---
|
||||
|
||||
<figure class="forward-lab" data-forward-lab>
|
||||
<figcaption>
|
||||
<span>ROUND 08 / TRAIN-TIME FORWARD</span>
|
||||
<div>
|
||||
<h3>不再只改 diagnostic backward：让局部 uniform routing 真正进入 8,000-step 训练</h3>
|
||||
<p>4 variants × 3 seeds · 1 historical paired reference · 1 full replay · frozen analyzer</p>
|
||||
</div>
|
||||
<em>REDUCED-MODEL ARCHITECTURE ABLATION</em>
|
||||
</figcaption>
|
||||
|
||||
<div class="forward-ledger">
|
||||
<article><span>NEW BYTES</span><b>851.968M</b><p>12 formal + 1 replay</p></article>
|
||||
<article><span>PRIMARY CELLS</span><b>{primary.passed_cells} / {primary.required_cells}</b><p>contrast + peak</p></article>
|
||||
<article><span>QUALITY</span><b>{quality.passed_checks} / {quality.required_checks}</b><p>BPC degradation screen</p></article>
|
||||
<article><span>REPLAY</span><b>{lab.replay.passed ? "EXACT" : "FAILED"}</b><p>scientific payload</p></article>
|
||||
<article><span>SELECTED PARAMS</span><b>UNREACHABLE</b><p>0 hooks · 0 optimizer state</p></article>
|
||||
<article class:list={{ pass: primary.material_response_passed, warn: !primary.material_response_passed }}>
|
||||
<span>STATUS</span><b>{primary.material_response_passed ? "PASS" : "NOT ESTABLISHED"}</b><p>reduced protocol only</p>
|
||||
</article>
|
||||
</div>
|
||||
|
||||
<div class="forward-tabs" role="tablist" aria-label="Round 08 训练期前向干预实验视图">
|
||||
<button type="button" role="tab" data-forward-tab="contract" aria-selected="true">01 / FORWARD CONTRACT</button>
|
||||
<button type="button" role="tab" data-forward-tab="trajectory" aria-selected="false">02 / TRAINING TRAJECTORY</button>
|
||||
<button type="button" role="tab" data-forward-tab="gate" aria-selected="false">03 / FINAL GATE</button>
|
||||
<button type="button" role="tab" data-forward-tab="interaction" aria-selected="false">04 / NON-ADDITIVITY</button>
|
||||
<button type="button" role="tab" data-forward-tab="spectrum" aria-selected="false">05 / SPECTRUM × AUDIT</button>
|
||||
</div>
|
||||
|
||||
<section class="forward-panel" data-forward-panel="contract">
|
||||
<div class="panel-lead">
|
||||
<div><span>I / WHAT ACTUALLY CHANGED</span><h4>这是训练期架构消融，不是“只改变 forward”的纯因果实验</h4></div>
|
||||
<p>选中 mixer 每一次 train / eval / diagnostic 都改成参数无关的均匀读取；新的表示自然改变 backward 与后续更新。</p>
|
||||
</div>
|
||||
<div class="equation-pair">
|
||||
<article>
|
||||
<span>LEARNED DEPTH MIXER</span>
|
||||
<code>keys = RMSNorm(sources)</code>
|
||||
<code>w = softmax(query · keys)</code>
|
||||
<b>output = Σ wᵢ · sourceᵢ</b>
|
||||
</article>
|
||||
<i>→</i>
|
||||
<article class="uniform">
|
||||
<span>SELECTED UNIFORM MIXER</span>
|
||||
<code>logits = zeros(N, B, T)</code>
|
||||
<code>w = softmax(logits) = 1 / N</code>
|
||||
<b>output = Σ sourceᵢ / N</b>
|
||||
</article>
|
||||
</div>
|
||||
<div class="group-map">
|
||||
{[1,2,3,4,5,6,7,8].map((group) => (
|
||||
<article class:list={{ target: group === 6 || group === 7 }}>
|
||||
<span>GROUP {group}</span>
|
||||
<b>L{(group - 1) * 4 + 1}–{group * 4}</b>
|
||||
<div>
|
||||
{[0,1,2,3].map((offset) => {
|
||||
const layer = (group - 1) * 4 + offset + 1;
|
||||
return <i class:list={{ spike: layer >= 21 && layer <= 25 }}>{layer}</i>;
|
||||
})}
|
||||
</div>
|
||||
<small>{group === 6 ? "#40–47" : group === 7 ? "#48–55" : "LEARNED"}</small>
|
||||
</article>
|
||||
))}
|
||||
</div>
|
||||
<div class="variant-grid">
|
||||
{variants.map((variant) => {
|
||||
const effect = lab.effects[variant];
|
||||
return (
|
||||
<article class:list={{ primary: variant === "uniform_groups_6_7_forward" }}>
|
||||
<span>{variant === "uniform_groups_6_7_forward" ? "PRIMARY" : "SECONDARY"}</span>
|
||||
<b>{variantLabels[variant]}</b>
|
||||
<dl>
|
||||
<div><dt>CONTRAST</dt><dd>{percent(meanDrop(effect, "spike_contrast"))}</dd></div>
|
||||
<div><dt>PEAK</dt><dd>{percent(meanDrop(effect, "peak_normalized"))}</dd></div>
|
||||
<div><dt>STATUS</dt><dd>{effect.material_response_passed ? "PASS" : "NOT EST."}</dd></div>
|
||||
</dl>
|
||||
</article>
|
||||
);
|
||||
})}
|
||||
</div>
|
||||
<div class="reachability-flow">
|
||||
<article><span>SELECTED QUERY / KEY NORM</span><b>留在 AdamW param groups</b><p>保持 optimizer 结构与 reference 一致。</p></article>
|
||||
<i>×</i>
|
||||
<article><span>COMPUTATION GRAPH</span><b>没有被 forward 调用</b><p>不是 stop-gradient；参数结构性不可达。</p></article>
|
||||
<i>→</i>
|
||||
<article><span>FINAL AUDIT</span><b>0 hook · 0 state · byte-exact init</b><p>不能写成“训练了但没有移动”。</p></article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="forward-panel" data-forward-panel="trajectory" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>II / SIX FROZEN CHECKPOINTS</span><h4>不是只看终点：局部前向干预从什么时候开始分化？</h4></div>
|
||||
<p>纵轴是相对同 seed historical learned reference 的下降；正值表示 attenuation，负值表示 amplification。</p>
|
||||
</div>
|
||||
<div class="forward-controls">
|
||||
<label>SEED
|
||||
<select data-forward-trajectory-seed>
|
||||
{seeds.map((seed) => <option value={String(seed)}>{seed}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<div>
|
||||
<button type="button" data-forward-trajectory-metric="spike_contrast" aria-pressed="true">SPIKE CONTRAST</button>
|
||||
<button type="button" data-forward-trajectory-metric="peak_normalized" aria-pressed="false">PEAK / MEAN</button>
|
||||
</div>
|
||||
<span data-forward-trajectory-state></span>
|
||||
</div>
|
||||
<div class="trajectory-chart">
|
||||
<header><b>RELATIVE DROP VS PAIRED REFERENCE</b><span>checkpoints are equally spaced; labels preserve actual steps</span></header>
|
||||
<svg viewBox="0 0 960 390" role="img" aria-label="四个训练变体的尖峰指标轨迹" data-forward-trajectory-chart>
|
||||
<g data-forward-trajectory-grid></g>
|
||||
<g data-forward-trajectory-series></g>
|
||||
</svg>
|
||||
<div class="series-key">
|
||||
<span><i class="g6"></i>GROUP 6</span>
|
||||
<span><i class="g7"></i>GROUP 7</span>
|
||||
<span><i class="joint"></i>GROUPS 6+7</span>
|
||||
<span><i class="mlp"></i>GROUP 7 MLP</span>
|
||||
</div>
|
||||
</div>
|
||||
<div class="trajectory-readout" data-forward-trajectory-readout></div>
|
||||
<div class="boundary-note">
|
||||
<b>轨迹不参与主闸门</b>
|
||||
<p>主 status 只读取 step 8,000；中间 checkpoint 用来观察适应过程，不能挑一个最好看的时点替代终点。</p>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="forward-panel" data-forward-panel="gate" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>III / PREREGISTERED FINAL VERDICT</span><h4>20% attenuation 与 BPC degradation screen 必须同时通过</h4></div>
|
||||
<p>任何结构、hash、selector、finite 或 replay 错误都会让 analyzer 直接失败，不进入下表的科学状态。</p>
|
||||
</div>
|
||||
<div class:list={{ "status-banner": true, pass: primary.material_response_passed, warn: !primary.material_response_passed }}>
|
||||
<span>FROZEN ANALYZER STATUS</span>
|
||||
<b>{statusLabel}</b>
|
||||
<p>ATTENUATION {primary.passed_cells} / {primary.required_cells} · QUALITY {quality.passed_checks} / {quality.required_checks}</p>
|
||||
</div>
|
||||
<div class="gate-layout">
|
||||
<div class="gate-table-wrap">
|
||||
<table class="gate-table">
|
||||
<thead><tr><th>SEED</th><th>METRIC</th><th>REFERENCE</th><th>VARIANT</th><th>DROP</th><th>≥20%</th></tr></thead>
|
||||
<tbody>
|
||||
{primary.cells.map((cell: any) => (
|
||||
<tr>
|
||||
<th>{cell.seed}</th>
|
||||
<td>{cell.metric === "spike_contrast" ? "CONTRAST" : "PEAK"}</td>
|
||||
<td>{cell.reference.toFixed(3)}×</td>
|
||||
<td>{cell.variant.toFixed(3)}×</td>
|
||||
<td class:list={{ good: cell.passed, bad: !cell.passed }}>{percent(cell.relative_drop)}</td>
|
||||
<td class:list={{ good: cell.passed, bad: !cell.passed }}>{cell.passed ? "PASS" : "FAIL"}</td>
|
||||
</tr>
|
||||
))}
|
||||
</tbody>
|
||||
</table>
|
||||
</div>
|
||||
<aside>
|
||||
<span>QUALITY SCREEN</span>
|
||||
<b>final validation BPC</b>
|
||||
{Object.entries(quality.per_seed).map(([seed, item]: [string, any]) => (
|
||||
<div><em>{seed}</em><strong class:list={{ good: item.passed, bad: !item.passed }}>{item.delta_bpc >= 0 ? "+" : ""}{item.delta_bpc.toFixed(4)}</strong><small>≤ +.050</small></div>
|
||||
))}
|
||||
<div class="mean"><em>3-SEED MEAN</em><strong class:list={{ good: quality.mean_passed, bad: !quality.mean_passed }}>{quality.mean_delta_bpc >= 0 ? "+" : ""}{quality.mean_delta_bpc.toFixed(4)}</strong><small>≤ +.030</small></div>
|
||||
</aside>
|
||||
</div>
|
||||
<div class="numerator-denominator">
|
||||
<article><span>SPIKE WINDOW</span><b>S = layers 21–25</b><p>报告 mean(g[S])，但不单独作为 status。</p></article>
|
||||
<i>÷</i>
|
||||
<article><span>REFERENCE WINDOW</span><b>R = other 27 layers</b><p>layers 26–28 也被 joint intervention 改写。</p></article>
|
||||
<i>=</i>
|
||||
<article class="warning"><span>CONTRAST</span><b>不是“尖峰层关闭”</b><p>下降可能来自 S 降、R 升，或两者同时发生。</p></article>
|
||||
</div>
|
||||
<div class="claim-pair">
|
||||
<article class="yes"><span>可以说</span><b>{statusLabel}</b><p>只限这个固定缩小模型、数据、预算、seed 与 operational metric。</p></article>
|
||||
<article class="no"><span>不能说</span><b>K3 的训练尖峰已被定位</b><p>没有真实 2.8T checkpoint forward，也没有复现未公开 Figure 5(c) telemetry。</p></article>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="forward-panel" data-forward-panel="interaction" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>IV / DESCRIPTIVE LOG RESIDUAL</span><h4>joint effect 等不等于两个 single-run effects 相加？</h4></div>
|
||||
<p>三条 effect 来自三套独立训练；这里画的是跨 run 的 bookkeeping residual，不是因果 interaction 或 Shapley contribution。</p>
|
||||
</div>
|
||||
<div class="forward-controls">
|
||||
<label>STEP
|
||||
<select data-forward-interaction-step>
|
||||
{[0,100,500,2000,4000,8000].map((step) => <option value={String(step)}>{step.toLocaleString()}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<div>
|
||||
<button type="button" data-forward-interaction-metric="spike_contrast" aria-pressed="true">SPIKE CONTRAST</button>
|
||||
<button type="button" data-forward-interaction-metric="peak_normalized" aria-pressed="false">PEAK / MEAN</button>
|
||||
</div>
|
||||
<span data-forward-interaction-state></span>
|
||||
</div>
|
||||
<div class="interaction-equation">
|
||||
<span>E₆₇</span><i>−</i><span>E₆</span><i>−</i><span>E₇</span><b>= I₆₇</b>
|
||||
<p>E = ln(Xref / Xvariant)；I > 0 表示 joint log attenuation 大于两个 single-run effects 的和。</p>
|
||||
</div>
|
||||
<div class="interaction-cells" data-forward-interaction-cells></div>
|
||||
<div class="interaction-summary" data-forward-interaction-summary></div>
|
||||
<div class="boundary-note">
|
||||
<b>没有预注册通过线</b>
|
||||
<p>I₆₇ 只报告原值、三 seed mean 与 range；正负都不能翻译成 group 6 / 7 的真实贡献。</p>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="forward-panel" data-forward-panel="spectrum" hidden>
|
||||
<div class="panel-lead">
|
||||
<div><span>V / 32-LAYER SHAPE × REPRODUCTION</span><h4>终点平均数之外：峰值移动、S/R 分拆与完整 replay</h4></div>
|
||||
<p>每条谱按自身 32-layer mean 归一化；选择 reference 或任一训练变体，不改变 analyzer 的正式 status。</p>
|
||||
</div>
|
||||
<div class="forward-controls spectrum-controls">
|
||||
<label>SEED
|
||||
<select data-forward-spectrum-seed>
|
||||
{seeds.map((seed) => <option value={String(seed)}>{seed}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<label>VARIANT
|
||||
<select data-forward-spectrum-variant>
|
||||
{["learned_reference", ...variants].map((variant) => <option value={variant}>{variantLabels[variant]}</option>)}
|
||||
</select>
|
||||
</label>
|
||||
<span data-forward-spectrum-state></span>
|
||||
</div>
|
||||
<div class="spectrum-layout">
|
||||
<div class="spectrum-chart">
|
||||
<header><b>FINAL NORMALIZED ACTIVATION-GRADIENT RMS</b><span>layer mean = 1</span></header>
|
||||
<svg viewBox="0 0 940 350" role="img" aria-label="训练期前向变体的 32 层梯度谱" data-forward-spectrum-chart>
|
||||
<rect class="spike-zone" x="0" y="26" width="0" height="280" data-forward-spectrum-zone></rect>
|
||||
<g data-forward-spectrum-grid></g>
|
||||
<polyline points="" data-forward-spectrum-line></polyline>
|
||||
<g data-forward-spectrum-points></g>
|
||||
</svg>
|
||||
</div>
|
||||
<aside>
|
||||
<span>SELECTED READOUT</span>
|
||||
<b data-forward-spectrum-label></b>
|
||||
<dl>
|
||||
<div><dt>SPIKE MEAN</dt><dd data-forward-spectrum-smean></dd></div>
|
||||
<div><dt>R MEAN</dt><dd data-forward-spectrum-rmean></dd></div>
|
||||
<div><dt>CONTRAST</dt><dd data-forward-spectrum-contrast></dd></div>
|
||||
<div><dt>PEAK / MEAN</dt><dd data-forward-spectrum-peak></dd></div>
|
||||
<div><dt>PEAK LAYER</dt><dd data-forward-spectrum-layer></dd></div>
|
||||
</dl>
|
||||
</aside>
|
||||
</div>
|
||||
<div class="audit-grid">
|
||||
<article><span>STEP-0 IDENTITY</span><b>5 / 5 exact</b><p>logits、CE、validation、32-layer diagnostic 与 capture summary。</p></article>
|
||||
<article><span>EMPTY SELECTOR</span><b>parent exact</b><p>20-step model、optimizer、history、evaluations 与 hashes。</p></article>
|
||||
<article><span>PRIMARY REPLAY</span><b>{lab.replay.passed ? "scientific exact" : "FAILED"}</b><p><code>{shortHash(lab.replay.scientific_payload_sha256)}</code></p></article>
|
||||
<article><span>PROCESSED</span><b>851,968,000 bytes</b><p>历史 references 的 196,608,000 bytes 单列、不重复计入。</p></article>
|
||||
<article class="boundary"><span>REAL K3</span><b>not executed</b><p><code>A_log [128]↔[96]</code> 尚无官方裁决。</p></article>
|
||||
</div>
|
||||
<div class="hash-strip">
|
||||
<span>PROTOCOL <code>{lab.protocol_id}</code></span>
|
||||
<span>AGGREGATE <code>{shortHash(lab.aggregate_sha256)}</code></span>
|
||||
<span>COMPACT <code>{shortHash(lab.canonical_sha256_without_self)}</code></span>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<script is:inline type="application/json" data-forward-payload set:html={json}></script>
|
||||
</figure>
|
||||
|
||||
<script>
|
||||
const initializeForwardLab = (root: HTMLElement) => {
|
||||
if (root.dataset.ready === "true") return;
|
||||
root.dataset.ready = "true";
|
||||
const payload = root.querySelector<HTMLScriptElement>("[data-forward-payload]");
|
||||
if (!payload) return;
|
||||
const data = JSON.parse(payload.textContent || "{}");
|
||||
const variants = [
|
||||
"uniform_group_6_forward",
|
||||
"uniform_group_7_forward",
|
||||
"uniform_groups_6_7_forward",
|
||||
"uniform_group_7_mlp_forward",
|
||||
];
|
||||
const labels: Record<string, string> = {
|
||||
learned_reference: "HISTORICAL LEARNED REFERENCE",
|
||||
uniform_group_6_forward: "GROUP 6",
|
||||
uniform_group_7_forward: "GROUP 7",
|
||||
uniform_groups_6_7_forward: "GROUPS 6+7",
|
||||
uniform_group_7_mlp_forward: "GROUP 7 MLP",
|
||||
};
|
||||
const colors: Record<string, string> = {
|
||||
uniform_group_6_forward: "#c98a58",
|
||||
uniform_group_7_forward: "#8794aa",
|
||||
uniform_groups_6_7_forward: "#8aae8d",
|
||||
uniform_group_7_mlp_forward: "#b487a7",
|
||||
};
|
||||
const tabs = [...root.querySelectorAll<HTMLButtonElement>("[data-forward-tab]")];
|
||||
const panels = [...root.querySelectorAll<HTMLElement>("[data-forward-panel]")];
|
||||
const activate = (name: string, focus = false) => {
|
||||
tabs.forEach((tab) => {
|
||||
const selected = tab.dataset.forwardTab === name;
|
||||
tab.setAttribute("aria-selected", String(selected));
|
||||
if (selected && focus) tab.focus();
|
||||
});
|
||||
panels.forEach((panel) => panel.hidden = panel.dataset.forwardPanel !== name);
|
||||
};
|
||||
tabs.forEach((tab, index) => {
|
||||
tab.addEventListener("click", () => activate(tab.dataset.forwardTab || "contract"));
|
||||
tab.addEventListener("keydown", (event) => {
|
||||
if (!["ArrowLeft", "ArrowRight", "Home", "End"].includes(event.key)) return;
|
||||
event.preventDefault();
|
||||
let next = index;
|
||||
if (event.key === "ArrowRight") next = (index + 1) % tabs.length;
|
||||
if (event.key === "ArrowLeft") next = (index - 1 + tabs.length) % tabs.length;
|
||||
if (event.key === "Home") next = 0;
|
||||
if (event.key === "End") next = tabs.length - 1;
|
||||
activate(tabs[next].dataset.forwardTab || "contract", true);
|
||||
});
|
||||
});
|
||||
|
||||
const ns = "http://www.w3.org/2000/svg";
|
||||
const make = (name: string, attributes: Record<string, string>) => {
|
||||
const node = document.createElementNS(ns, name);
|
||||
Object.entries(attributes).forEach(([key, value]) => node.setAttribute(key, value));
|
||||
return node;
|
||||
};
|
||||
const setText = (selector: string, value: string) => {
|
||||
const node = root.querySelector<HTMLElement>(selector);
|
||||
if (node) node.textContent = value;
|
||||
};
|
||||
|
||||
let trajectoryMetric = "spike_contrast";
|
||||
const trajectorySeed = root.querySelector<HTMLSelectElement>("[data-forward-trajectory-seed]");
|
||||
const renderTrajectory = () => {
|
||||
if (!trajectorySeed) return;
|
||||
const seed = Number(trajectorySeed.value);
|
||||
const rows = data.trajectories.filter((row: any) => row.seed === seed && variants.includes(row.variant));
|
||||
const steps = [...new Set(rows.map((row: any) => row.step))] as number[];
|
||||
const values = rows.map((row: any) => row.relative_drop[trajectoryMetric]);
|
||||
const minimum = Math.min(-.05, ...values);
|
||||
const maximum = Math.max(.25, ...values);
|
||||
const left = 75, right = 925, top = 35, bottom = 330;
|
||||
const x = (index: number) => left + index * (right - left) / (steps.length - 1);
|
||||
const y = (value: number) => bottom - (value - minimum) / (maximum - minimum) * (bottom - top);
|
||||
const grid = root.querySelector<SVGGElement>("[data-forward-trajectory-grid]");
|
||||
const series = root.querySelector<SVGGElement>("[data-forward-trajectory-series]");
|
||||
if (!grid || !series) return;
|
||||
grid.innerHTML = ""; series.innerHTML = "";
|
||||
[minimum, 0, .2, maximum].filter((value, index, all) => all.indexOf(value) === index).forEach((value) => {
|
||||
grid.append(make("line", { x1: String(left), x2: String(right), y1: String(y(value)), y2: String(y(value)), class: value === 0 ? "zero" : "" }));
|
||||
const label = make("text", { x: "12", y: String(y(value) + 4) });
|
||||
label.textContent = `${value >= 0 ? "+" : ""}${(value * 100).toFixed(0)}%`;
|
||||
grid.append(label);
|
||||
});
|
||||
steps.forEach((step, index) => {
|
||||
const label = make("text", { x: String(x(index)), y: "360", "text-anchor": "middle" });
|
||||
label.textContent = step.toLocaleString();
|
||||
grid.append(label);
|
||||
});
|
||||
variants.forEach((variant) => {
|
||||
const selected = rows.filter((row: any) => row.variant === variant).sort((a: any, b: any) => a.step - b.step);
|
||||
const points = selected.map((row: any, index: number) => `${x(index)},${y(row.relative_drop[trajectoryMetric])}`).join(" ");
|
||||
series.append(make("polyline", { points, fill: "none", stroke: colors[variant], "stroke-width": variant.includes("groups_6_7") ? "3" : "2" }));
|
||||
selected.forEach((row: any, index: number) => {
|
||||
const circle = make("circle", { cx: String(x(index)), cy: String(y(row.relative_drop[trajectoryMetric])), r: "4", fill: colors[variant], "data-variant": variant, "data-step": String(row.step) });
|
||||
series.append(circle);
|
||||
});
|
||||
});
|
||||
const metricLabel = trajectoryMetric === "spike_contrast" ? "SPIKE CONTRAST" : "PEAK / MEAN";
|
||||
setText("[data-forward-trajectory-state]", `${seed} · ${metricLabel}`);
|
||||
const readout = root.querySelector<HTMLElement>("[data-forward-trajectory-readout]");
|
||||
if (readout) {
|
||||
readout.innerHTML = variants.map((variant) => {
|
||||
const final = rows.find((row: any) => row.variant === variant && row.step === 8000);
|
||||
const value = final.relative_drop[trajectoryMetric];
|
||||
return `<article><span>${labels[variant]}</span><b>${value >= 0 ? "+" : "−"}${Math.abs(value * 100).toFixed(1)}%</b><p>step 8,000 vs same-seed reference</p></article>`;
|
||||
}).join("");
|
||||
}
|
||||
};
|
||||
trajectorySeed?.addEventListener("change", renderTrajectory);
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-forward-trajectory-metric]").forEach((button) => {
|
||||
button.addEventListener("click", () => {
|
||||
trajectoryMetric = button.dataset.forwardTrajectoryMetric || "spike_contrast";
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-forward-trajectory-metric]").forEach((candidate) =>
|
||||
candidate.setAttribute("aria-pressed", String(candidate === button)));
|
||||
renderTrajectory();
|
||||
});
|
||||
});
|
||||
renderTrajectory();
|
||||
|
||||
let interactionMetric = "spike_contrast";
|
||||
const interactionStep = root.querySelector<HTMLSelectElement>("[data-forward-interaction-step]");
|
||||
if (interactionStep) interactionStep.value = "8000";
|
||||
const renderInteraction = () => {
|
||||
if (!interactionStep) return;
|
||||
const step = Number(interactionStep.value);
|
||||
const cells = data.interaction.cells.filter((cell: any) => cell.step === step && cell.metric === interactionMetric);
|
||||
const container = root.querySelector<HTMLElement>("[data-forward-interaction-cells]");
|
||||
if (container) {
|
||||
container.innerHTML = cells.map((cell: any) => {
|
||||
const residual = cell.interaction_residual;
|
||||
return `<article data-sign="${residual >= 0 ? "positive" : "negative"}"><span>SEED ${cell.seed}</span><div><em>E6</em><b>${cell.log_effects.group6.toFixed(3)}</b></div><div><em>E7</em><b>${cell.log_effects.group7.toFixed(3)}</b></div><div><em>E67</em><b>${cell.log_effects.groups6_7.toFixed(3)}</b></div><strong>I67 ${residual >= 0 ? "+" : ""}${residual.toFixed(3)}</strong></article>`;
|
||||
}).join("");
|
||||
}
|
||||
const summary = data.interaction.summaries.find((item: any) => item.step === step && item.metric === interactionMetric);
|
||||
const summaryNode = root.querySelector<HTMLElement>("[data-forward-interaction-summary]");
|
||||
if (summaryNode && summary) {
|
||||
summaryNode.innerHTML = `<span>3-SEED DESCRIPTIVE SUMMARY</span><b>mean I67 ${summary.mean_interaction_residual >= 0 ? "+" : ""}${summary.mean_interaction_residual.toFixed(3)}</b><p>range ${summary.minimum.toFixed(3)} → ${summary.maximum.toFixed(3)}</p>`;
|
||||
}
|
||||
setText("[data-forward-interaction-state]", `${step.toLocaleString()} · ${interactionMetric === "spike_contrast" ? "SPIKE CONTRAST" : "PEAK / MEAN"}`);
|
||||
};
|
||||
interactionStep?.addEventListener("change", renderInteraction);
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-forward-interaction-metric]").forEach((button) => {
|
||||
button.addEventListener("click", () => {
|
||||
interactionMetric = button.dataset.forwardInteractionMetric || "spike_contrast";
|
||||
root.querySelectorAll<HTMLButtonElement>("[data-forward-interaction-metric]").forEach((candidate) =>
|
||||
candidate.setAttribute("aria-pressed", String(candidate === button)));
|
||||
renderInteraction();
|
||||
});
|
||||
});
|
||||
renderInteraction();
|
||||
|
||||
const spectrumSeed = root.querySelector<HTMLSelectElement>("[data-forward-spectrum-seed]");
|
||||
const spectrumVariant = root.querySelector<HTMLSelectElement>("[data-forward-spectrum-variant]");
|
||||
const renderSpectrum = () => {
|
||||
if (!spectrumSeed || !spectrumVariant) return;
|
||||
const seed = Number(spectrumSeed.value);
|
||||
const variant = spectrumVariant.value;
|
||||
const record = data.final_spectra.find((item: any) => item.seed === seed && item.variant === variant);
|
||||
if (!record) return;
|
||||
const values = record.normalized;
|
||||
const left = 48, right = 915, top = 28, bottom = 305;
|
||||
const maximum = Math.max(2, ...values) * 1.08;
|
||||
const x = (index: number) => left + index * (right - left) / 31;
|
||||
const y = (value: number) => bottom - value / maximum * (bottom - top);
|
||||
const grid = root.querySelector<SVGGElement>("[data-forward-spectrum-grid]");
|
||||
const points = root.querySelector<SVGGElement>("[data-forward-spectrum-points]");
|
||||
const line = root.querySelector<SVGPolylineElement>("[data-forward-spectrum-line]");
|
||||
const zone = root.querySelector<SVGRectElement>("[data-forward-spectrum-zone]");
|
||||
if (!grid || !points || !line || !zone) return;
|
||||
grid.innerHTML = ""; points.innerHTML = "";
|
||||
[0, 1, Math.ceil(maximum)].forEach((value) => {
|
||||
grid.append(make("line", { x1: String(left), x2: String(right), y1: String(y(value)), y2: String(y(value)) }));
|
||||
const label = make("text", { x: "8", y: String(y(value) + 4) });
|
||||
label.textContent = `${value}×`; grid.append(label);
|
||||
});
|
||||
[1,5,9,13,17,21,25,29,32].forEach((layer) => {
|
||||
const label = make("text", { x: String(x(layer - 1)), y: "330", "text-anchor": "middle" });
|
||||
label.textContent = String(layer); grid.append(label);
|
||||
});
|
||||
zone.setAttribute("x", String(x(20) - 8));
|
||||
zone.setAttribute("width", String(x(24) - x(20) + 16));
|
||||
line.setAttribute("points", values.map((value: number, index: number) => `${x(index)},${y(value)}`).join(" "));
|
||||
values.forEach((value: number, index: number) => {
|
||||
points.append(make("circle", { cx: String(x(index)), cy: String(y(value)), r: index + 1 === record.peak_layer_1based ? "4.5" : "2.6", "data-layer": String(index + 1) }));
|
||||
});
|
||||
setText("[data-forward-spectrum-state]", `${seed} · ${labels[variant]}`);
|
||||
setText("[data-forward-spectrum-label]", labels[variant]);
|
||||
setText("[data-forward-spectrum-smean]", record.spike_mean.toExponential(3));
|
||||
setText("[data-forward-spectrum-rmean]", record.reference_mean.toExponential(3));
|
||||
setText("[data-forward-spectrum-contrast]", `${record.spike_contrast.toFixed(3)}×`);
|
||||
setText("[data-forward-spectrum-peak]", `${record.peak_normalized.toFixed(3)}×`);
|
||||
setText("[data-forward-spectrum-layer]", `L${record.peak_layer_1based}`);
|
||||
};
|
||||
spectrumSeed?.addEventListener("change", renderSpectrum);
|
||||
spectrumVariant?.addEventListener("change", renderSpectrum);
|
||||
renderSpectrum();
|
||||
};
|
||||
|
||||
document.querySelectorAll<HTMLElement>("[data-forward-lab]").forEach(initializeForwardLab);
|
||||
document.addEventListener("astro:page-load", () => {
|
||||
document.querySelectorAll<HTMLElement>("[data-forward-lab]").forEach(initializeForwardLab);
|
||||
});
|
||||
</script>
|
||||
|
||||
<style>
|
||||
.forward-lab {
|
||||
--ink: #d9d7ce;
|
||||
--muted: #8e918d;
|
||||
--line: rgba(217,215,206,.14);
|
||||
--panel: rgba(14,17,16,.78);
|
||||
--green: #8aae8d;
|
||||
--copper: #c98a58;
|
||||
--blue: #8794aa;
|
||||
--pink: #b487a7;
|
||||
--red: #c77768;
|
||||
margin: 32px 0;
|
||||
border: 1px solid var(--line);
|
||||
background: #0d0f0e;
|
||||
color: var(--ink);
|
||||
}
|
||||
.forward-lab h3, .forward-lab h4 { color: var(--ink); }
|
||||
.forward-lab figcaption { display: grid; grid-template-columns: 150px 1fr auto; gap: 24px; align-items: start; padding: 25px; border-bottom: 1px solid var(--line); }
|
||||
.forward-lab figcaption > span, .forward-lab figcaption > em { color: var(--copper); font: normal .52rem var(--mono); letter-spacing: .12em; }
|
||||
.forward-lab figcaption > em { color: var(--muted); text-align: right; }
|
||||
.forward-lab figcaption h3 { margin: 0 0 8px; font-size: 1.02rem; line-height: 1.35; }
|
||||
.forward-lab figcaption p { margin: 0; color: var(--muted); font: .56rem/1.5 var(--mono); }
|
||||
.forward-ledger { display: grid; grid-template-columns: repeat(6, 1fr); border-bottom: 1px solid var(--line); }
|
||||
.forward-ledger article { min-height: 96px; padding: 15px; border-right: 1px solid var(--line); }
|
||||
.forward-ledger article:last-child { border-right: 0; }
|
||||
.forward-ledger span, .variant-grid span, .audit-grid span { color: var(--muted); font: .48rem var(--mono); }
|
||||
.forward-ledger b { display: block; margin: 11px 0 6px; font: 800 .66rem var(--mono); }
|
||||
.forward-ledger p { margin: 0; color: var(--muted); font-size: .51rem; }
|
||||
.forward-ledger .pass b { color: var(--green); }
|
||||
.forward-ledger .warn b { color: var(--red); }
|
||||
.forward-tabs { display: grid; grid-template-columns: repeat(5, 1fr); border-bottom: 1px solid var(--line); }
|
||||
.forward-tabs button { min-height: 48px; border: 0; border-right: 1px solid var(--line); background: transparent; color: var(--muted); font: .51rem var(--mono); cursor: pointer; }
|
||||
.forward-tabs button:last-child { border-right: 0; }
|
||||
.forward-tabs button[aria-selected="true"] { background: rgba(201,138,88,.09); color: var(--copper); box-shadow: inset 0 -2px var(--copper); }
|
||||
.forward-tabs button:focus-visible { outline: 2px solid var(--green); outline-offset: -3px; }
|
||||
.forward-panel { padding: 26px; }
|
||||
.panel-lead { display: grid; grid-template-columns: 1fr minmax(260px, 42%); gap: 32px; margin-bottom: 22px; }
|
||||
.panel-lead span { color: var(--copper); font: .5rem var(--mono); }
|
||||
.panel-lead h4 { margin: 8px 0 0; font-size: .9rem; }
|
||||
.panel-lead p { margin: 0; color: var(--muted); font-size: .64rem; line-height: 1.65; }
|
||||
.equation-pair { display: grid; grid-template-columns: 1fr auto 1fr; gap: 14px; align-items: center; }
|
||||
.equation-pair article { padding: 18px; border: 1px solid var(--line); background: var(--panel); }
|
||||
.equation-pair article.uniform { border-color: rgba(138,174,141,.45); }
|
||||
.equation-pair span { color: var(--muted); font: .5rem var(--mono); }
|
||||
.equation-pair code, .equation-pair b { display: block; margin-top: 10px; font: .61rem var(--mono); }
|
||||
.equation-pair code { color: var(--muted); }
|
||||
.equation-pair b { color: var(--green); }
|
||||
.equation-pair > i, .reachability-flow > i, .numerator-denominator > i { color: var(--copper); font-style: normal; }
|
||||
.group-map { display: grid; grid-template-columns: repeat(8, 1fr); margin-top: 18px; border: 1px solid var(--line); }
|
||||
.group-map article { min-height: 118px; padding: 12px; border-right: 1px solid var(--line); }
|
||||
.group-map article:last-child { border-right: 0; }
|
||||
.group-map article.target { background: rgba(201,138,88,.08); box-shadow: inset 0 3px var(--copper); }
|
||||
.group-map span, .group-map small { color: var(--muted); font: .45rem var(--mono); }
|
||||
.group-map b { display: block; margin: 9px 0; font: .64rem var(--mono); }
|
||||
.group-map div { display: grid; grid-template-columns: repeat(4, 1fr); gap: 3px; }
|
||||
.group-map i { display: grid; place-items: center; aspect-ratio: 1; border: 1px solid var(--line); color: var(--muted); font: normal .44rem var(--mono); }
|
||||
.group-map i.spike { border-color: rgba(138,174,141,.6); color: var(--green); }
|
||||
.group-map small { display: block; margin-top: 8px; }
|
||||
.variant-grid { display: grid; grid-template-columns: repeat(4, 1fr); gap: 12px; margin-top: 18px; }
|
||||
.variant-grid article { padding: 15px; border: 1px solid var(--line); }
|
||||
.variant-grid article.primary { border-color: rgba(138,174,141,.5); }
|
||||
.variant-grid b { display: block; min-height: 34px; margin: 8px 0 12px; font: .6rem/1.4 var(--mono); }
|
||||
.variant-grid dl { margin: 0; }
|
||||
.variant-grid dl div { display: flex; justify-content: space-between; padding: 7px 0; border-top: 1px solid var(--line); }
|
||||
.variant-grid dt, .variant-grid dd { margin: 0; font: .48rem var(--mono); }
|
||||
.variant-grid dt { color: var(--muted); }
|
||||
.variant-grid dd { color: var(--green); }
|
||||
.reachability-flow, .numerator-denominator { display: grid; grid-template-columns: 1fr auto 1fr auto 1fr; gap: 12px; align-items: center; margin-top: 18px; }
|
||||
.reachability-flow article, .numerator-denominator article { min-height: 108px; padding: 15px; border: 1px solid var(--line); }
|
||||
.reachability-flow span, .numerator-denominator span { color: var(--muted); font: .47rem var(--mono); }
|
||||
.reachability-flow b, .numerator-denominator b { display: block; margin: 9px 0; font: .64rem var(--mono); }
|
||||
.reachability-flow p, .numerator-denominator p { margin: 0; color: var(--muted); font-size: .56rem; line-height: 1.5; }
|
||||
.numerator-denominator .warning { border-color: rgba(199,119,104,.45); }
|
||||
.forward-controls { display: flex; align-items: stretch; margin-bottom: 16px; border: 1px solid var(--line); }
|
||||
.forward-controls label { display: flex; align-items: center; gap: 9px; padding: 10px 13px; border-right: 1px solid var(--line); color: var(--muted); font: .48rem var(--mono); }
|
||||
.forward-controls select { max-width: 260px; border: 1px solid var(--line); background: #0d0f0e; color: var(--ink); font: .53rem var(--mono); }
|
||||
.forward-controls > div { display: flex; }
|
||||
.forward-controls button { border: 0; border-right: 1px solid var(--line); background: transparent; color: var(--muted); font: .48rem var(--mono); }
|
||||
.forward-controls button[aria-pressed="true"] { color: var(--green); background: rgba(138,174,141,.08); }
|
||||
.forward-controls > span { display: grid; place-items: center; margin-left: auto; padding: 0 14px; color: var(--green); font: .49rem var(--mono); }
|
||||
.trajectory-chart, .spectrum-chart { border: 1px solid var(--line); background: var(--panel); }
|
||||
.trajectory-chart header, .spectrum-chart header { display: flex; justify-content: space-between; padding: 11px 14px; border-bottom: 1px solid var(--line); }
|
||||
.trajectory-chart header b, .trajectory-chart header span, .spectrum-chart header b, .spectrum-chart header span { font: .49rem var(--mono); }
|
||||
.trajectory-chart header span, .spectrum-chart header span { color: var(--muted); }
|
||||
.trajectory-chart svg, .spectrum-chart svg { display: block; width: 100%; height: auto; }
|
||||
[data-forward-trajectory-grid] :global(line), [data-forward-spectrum-grid] :global(line) { stroke: var(--line); stroke-width: 1; }
|
||||
[data-forward-trajectory-grid] :global(line.zero) { stroke: rgba(217,215,206,.5); stroke-dasharray: 4 4; }
|
||||
[data-forward-trajectory-grid] :global(text), [data-forward-spectrum-grid] :global(text) { fill: var(--muted); font: 10px var(--mono); }
|
||||
.series-key { display: flex; flex-wrap: wrap; gap: 18px; padding: 10px 14px; border-top: 1px solid var(--line); color: var(--muted); font: .48rem var(--mono); }
|
||||
.series-key span { display: flex; align-items: center; gap: 7px; }
|
||||
.series-key i { width: 16px; height: 2px; }
|
||||
.series-key .g6 { background: var(--copper); }
|
||||
.series-key .g7 { background: var(--blue); }
|
||||
.series-key .joint { background: var(--green); }
|
||||
.series-key .mlp { background: var(--pink); }
|
||||
.trajectory-readout { display: grid; grid-template-columns: repeat(4, 1fr); gap: 12px; margin-top: 14px; }
|
||||
.trajectory-readout :global(article) { padding: 14px; border: 1px solid var(--line); }
|
||||
.trajectory-readout :global(span) { color: var(--muted); font: .47rem var(--mono); }
|
||||
.trajectory-readout :global(b) { display: block; margin: 8px 0; color: var(--green); font: .7rem var(--mono); }
|
||||
.trajectory-readout :global(p) { margin: 0; color: var(--muted); font-size: .53rem; }
|
||||
.boundary-note { margin-top: 16px; padding: 15px; border-left: 3px solid var(--copper); background: rgba(201,138,88,.07); }
|
||||
.boundary-note b { font-size: .64rem; }
|
||||
.boundary-note p { margin: 6px 0 0; color: var(--muted); font-size: .58rem; line-height: 1.55; }
|
||||
.status-banner { display: grid; grid-template-columns: 180px 1fr auto; gap: 18px; align-items: center; padding: 16px; border: 1px solid var(--line); }
|
||||
.status-banner.pass { border-color: rgba(138,174,141,.5); }
|
||||
.status-banner.warn { border-color: rgba(199,119,104,.5); }
|
||||
.status-banner span, .status-banner p { color: var(--muted); font: .49rem var(--mono); }
|
||||
.status-banner b { color: var(--green); font: .72rem var(--mono); }
|
||||
.status-banner.warn b { color: var(--red); }
|
||||
.gate-layout { display: grid; grid-template-columns: 1fr 230px; gap: 15px; margin-top: 16px; }
|
||||
.gate-table-wrap { overflow-x: auto; border: 1px solid var(--line); }
|
||||
.gate-table { width: 100%; border-collapse: collapse; min-width: 680px; font: .53rem var(--mono); }
|
||||
.gate-table th, .gate-table td { padding: 12px 10px; border-right: 1px solid var(--line); border-bottom: 1px solid var(--line); text-align: right; }
|
||||
.gate-table thead { color: var(--muted); }
|
||||
.good { color: var(--green) !important; }
|
||||
.bad { color: var(--red) !important; }
|
||||
.gate-layout aside { padding: 15px; border: 1px solid var(--line); }
|
||||
.gate-layout aside > span { color: var(--muted); font: .48rem var(--mono); }
|
||||
.gate-layout aside > b { display: block; margin: 8px 0 14px; font: .63rem var(--mono); }
|
||||
.gate-layout aside > div { display: grid; grid-template-columns: 1fr auto; gap: 4px 8px; padding: 9px 0; border-top: 1px solid var(--line); }
|
||||
.gate-layout aside em, .gate-layout aside strong, .gate-layout aside small { font: normal .49rem var(--mono); }
|
||||
.gate-layout aside small { grid-column: 1 / -1; color: var(--muted); }
|
||||
.gate-layout aside .mean { margin-top: 5px; }
|
||||
.claim-pair { display: grid; grid-template-columns: 1fr 1fr; gap: 14px; margin-top: 16px; }
|
||||
.claim-pair article { padding: 16px; border: 1px solid var(--line); }
|
||||
.claim-pair span { color: var(--muted); font: .49rem var(--mono); }
|
||||
.claim-pair b { display: block; margin: 9px 0; font: .65rem var(--mono); }
|
||||
.claim-pair p { margin: 0; color: var(--muted); font-size: .57rem; line-height: 1.5; }
|
||||
.claim-pair .yes { border-color: rgba(138,174,141,.45); }
|
||||
.claim-pair .yes b { color: var(--green); }
|
||||
.claim-pair .no { border-color: rgba(199,119,104,.45); }
|
||||
.claim-pair .no b { color: var(--red); }
|
||||
.interaction-equation { display: flex; flex-wrap: wrap; align-items: center; gap: 12px; padding: 17px; border: 1px solid var(--line); }
|
||||
.interaction-equation span, .interaction-equation b { padding: 8px 12px; background: var(--panel); font: .68rem var(--mono); }
|
||||
.interaction-equation i { color: var(--copper); font-style: normal; }
|
||||
.interaction-equation b { color: var(--green); }
|
||||
.interaction-equation p { flex-basis: 100%; margin: 0; color: var(--muted); font-size: .56rem; }
|
||||
.interaction-cells { display: grid; grid-template-columns: repeat(3, 1fr); gap: 14px; margin-top: 16px; }
|
||||
.interaction-cells :global(article) { padding: 16px; border: 1px solid var(--line); }
|
||||
.interaction-cells :global(article[data-sign="positive"]) { border-color: rgba(138,174,141,.4); }
|
||||
.interaction-cells :global(article[data-sign="negative"]) { border-color: rgba(199,119,104,.4); }
|
||||
.interaction-cells :global(span) { color: var(--muted); font: .48rem var(--mono); }
|
||||
.interaction-cells :global(div) { display: flex; justify-content: space-between; margin-top: 9px; padding-top: 8px; border-top: 1px solid var(--line); }
|
||||
.interaction-cells :global(em), .interaction-cells :global(b) { font: normal .52rem var(--mono); }
|
||||
.interaction-cells :global(em) { color: var(--muted); }
|
||||
.interaction-cells :global(strong) { display: block; margin-top: 12px; color: var(--green); font: .65rem var(--mono); }
|
||||
.interaction-summary { margin-top: 14px; padding: 16px; border: 1px solid var(--line); }
|
||||
.interaction-summary :global(span) { color: var(--muted); font: .48rem var(--mono); }
|
||||
.interaction-summary :global(b) { display: block; margin: 8px 0; font: .7rem var(--mono); }
|
||||
.interaction-summary :global(p) { margin: 0; color: var(--muted); font: .52rem var(--mono); }
|
||||
.spectrum-controls label:nth-child(2) { flex: 1; }
|
||||
.spectrum-controls label:nth-child(2) select { width: 100%; max-width: none; }
|
||||
.spectrum-layout { display: grid; grid-template-columns: 1fr 225px; gap: 15px; }
|
||||
.spike-zone { fill: rgba(201,138,88,.1); }
|
||||
[data-forward-spectrum-line] { fill: none; stroke: var(--green); stroke-width: 2; }
|
||||
[data-forward-spectrum-points] :global(circle) { fill: var(--green); stroke: #0d0f0e; stroke-width: 1; }
|
||||
.spectrum-layout aside { padding: 16px; border: 1px solid var(--line); background: var(--panel); }
|
||||
.spectrum-layout aside > span { color: var(--muted); font: .48rem var(--mono); }
|
||||
.spectrum-layout aside > b { display: block; margin: 9px 0 16px; font: .61rem/1.4 var(--mono); }
|
||||
.spectrum-layout dl { margin: 0; }
|
||||
.spectrum-layout dl div { display: flex; justify-content: space-between; gap: 8px; padding: 10px 0; border-top: 1px solid var(--line); }
|
||||
.spectrum-layout dt, .spectrum-layout dd { margin: 0; font: .49rem var(--mono); }
|
||||
.spectrum-layout dt { color: var(--muted); }
|
||||
.spectrum-layout dd { color: var(--green); }
|
||||
.audit-grid { display: grid; grid-template-columns: repeat(5, 1fr); gap: 12px; margin-top: 16px; }
|
||||
.audit-grid article { padding: 14px; border: 1px solid var(--line); }
|
||||
.audit-grid b { display: block; margin: 9px 0; font: .63rem var(--mono); }
|
||||
.audit-grid p { margin: 0; color: var(--muted); font-size: .53rem; line-height: 1.5; }
|
||||
.audit-grid .boundary { border-color: rgba(199,119,104,.45); }
|
||||
.hash-strip { display: flex; flex-wrap: wrap; gap: 18px; margin-top: 14px; padding: 12px 14px; border: 1px solid var(--line); color: var(--muted); font: .48rem var(--mono); }
|
||||
.hash-strip code { color: var(--green); }
|
||||
|
||||
@media (max-width: 980px) {
|
||||
.forward-ledger { grid-template-columns: repeat(3, 1fr); }
|
||||
.forward-tabs { grid-template-columns: repeat(3, 1fr); }
|
||||
.group-map { grid-template-columns: repeat(4, 1fr); }
|
||||
.variant-grid, .trajectory-readout { grid-template-columns: repeat(2, 1fr); }
|
||||
.gate-layout, .spectrum-layout { grid-template-columns: 1fr; }
|
||||
.audit-grid { grid-template-columns: repeat(3, 1fr); }
|
||||
}
|
||||
@media (max-width: 680px) {
|
||||
.forward-lab figcaption { grid-template-columns: 1fr; padding: 20px; }
|
||||
.forward-lab figcaption > span { order: -1; }
|
||||
.forward-lab figcaption > em { text-align: left; }
|
||||
.forward-ledger { grid-template-columns: repeat(2, 1fr); }
|
||||
.forward-tabs { display: flex; overflow-x: auto; }
|
||||
.forward-tabs button { min-width: 165px; }
|
||||
.forward-panel { padding: 20px 14px; }
|
||||
.panel-lead { grid-template-columns: 1fr; gap: 12px; }
|
||||
.equation-pair, .reachability-flow, .numerator-denominator { grid-template-columns: 1fr; }
|
||||
.equation-pair > i, .reachability-flow > i, .numerator-denominator > i { text-align: center; transform: rotate(90deg); }
|
||||
.group-map { grid-template-columns: repeat(2, 1fr); }
|
||||
.variant-grid, .trajectory-readout, .claim-pair, .interaction-cells, .audit-grid { grid-template-columns: 1fr; }
|
||||
.forward-controls { flex-wrap: wrap; }
|
||||
.forward-controls > span { width: 100%; min-height: 34px; border-top: 1px solid var(--line); }
|
||||
.status-banner { grid-template-columns: 1fr; }
|
||||
.gate-table { min-width: 680px; }
|
||||
}
|
||||
</style>
|
||||
File diff suppressed because it is too large
Load Diff
File diff suppressed because it is too large
Load Diff
@@ -128,20 +128,20 @@ const paths = [
|
||||
<div class="release-grid">
|
||||
<a class="release-card k3-release" href="/k3/">
|
||||
<div>
|
||||
<p class="eyebrow"><span>NEW / K3 ROUND 07</span> LOCAL MIXER PATH · BIDIRECTIONAL GATE</p>
|
||||
<h2>16 个局部 mixer 单侧证据很强，但双向定位仍然没有闭合</h2>
|
||||
<p class="eyebrow"><span>NEW / K3 ROUND 08</span> TRAIN-TIME FORWARD · PREREGISTERED GATE</p>
|
||||
<h2>局部 uniform routing 进入完整训练后，六个尖峰指标格全部衰减</h2>
|
||||
<p>
|
||||
以 14 个冻结 mask 同时检查 sufficiency 与 restoration：groups 6+7 在 learned
|
||||
背景的 6 / 6 格全部超过 50% global log gap；但从 uniform 背景恢复时，peak 三个
|
||||
seed 全部未过线。因此只能报告 one-sided evidence，不能宣布尖峰已定位到这 16 个 mixer。
|
||||
四个前向架构变体各跑三个 8,000-step seed，再完整 replay 主格：groups 6+7 的
|
||||
contrast / peak 六格降幅为 32.3%–77.3%，BPC 质量门 4 / 4 通过。结论只限缩小
|
||||
depth-32 Block 协议，不是真实 K3 checkpoint 或 Figure 5(c) 复现。
|
||||
</p>
|
||||
</div>
|
||||
<dl>
|
||||
<div><dt>MATRIX</dt><dd>14 masks × 3 seeds</dd></div>
|
||||
<div><dt>SUFFICIENCY</dt><dd>6 / 6 pass</dd></div>
|
||||
<div><dt>RESTORATION</dt><dd>3 / 6 fail</dd></div>
|
||||
<div><dt>MATRIX</dt><dd>12 formal + 1 replay</dd></div>
|
||||
<div><dt>ATTENUATION</dt><dd>6 / 6 pass</dd></div>
|
||||
<div><dt>QUALITY</dt><dd>4 / 4 pass</dd></div>
|
||||
</dl>
|
||||
<span class="release-arrow" aria-hidden="true">进入局部路径图、双向门与 32 层原始谱 →</span>
|
||||
<span class="release-arrow" aria-hidden="true">进入训练轨迹、主门、non-additivity 与 32 层谱 →</span>
|
||||
</a>
|
||||
<a class="release-card deepseek-release" href="/deepseek/">
|
||||
<div>
|
||||
|
||||
@@ -2,6 +2,7 @@
|
||||
import BaseLayout from "@/layouts/BaseLayout.astro";
|
||||
import ArchitectureExplorer from "@/components/ArchitectureExplorer.astro";
|
||||
import K3ArtifactLab from "@/components/K3ArtifactLab.astro";
|
||||
import K3AttnResForwardLab from "@/components/K3AttnResForwardLab.astro";
|
||||
import K3AttnResGradientLab from "@/components/K3AttnResGradientLab.astro";
|
||||
import K3AttnResLocalPathLab from "@/components/K3AttnResLocalPathLab.astro";
|
||||
import K3AttnResSpikeLab from "@/components/K3AttnResSpikeLab.astro";
|
||||
@@ -44,7 +45,8 @@ const toc = [
|
||||
["31", "attnres-gradient", "梯度定义与深度扩展"],
|
||||
["32", "attnres-spike", "尖峰轨迹与反向路径"],
|
||||
["33", "attnres-local-path", "局部 mixer 双向干预"],
|
||||
["34", "audit", "21 张图表审计"],
|
||||
["34", "attnres-forward", "训练期前向干预"],
|
||||
["35", "audit", "21 张图表审计"],
|
||||
["↳", "papers", "100 节点阅读链"],
|
||||
];
|
||||
|
||||
@@ -113,13 +115,13 @@ const paperGroups = [
|
||||
|
||||
<BaseLayout
|
||||
title="Kimi K3 技术报告完整深读：架构、训练、RL、系统与评测"
|
||||
description="用三十二张问题账、二十一张图表审计、八个机制实验、四个开放工件视图、四轮二十个 AttnRes 独立实验视图与一百个一手阅读节点，逐节读懂 Kimi K3。"
|
||||
description="用三十二张问题账、二十一张图表审计、八个机制实验、四个开放工件视图、五轮二十五个 AttnRes 独立实验视图与一百个一手阅读节点，逐节读懂 Kimi K3。"
|
||||
section="k3"
|
||||
>
|
||||
<header class="page-hero k3-hero">
|
||||
<div class="page-hero-inner">
|
||||
<div>
|
||||
<p class="eyebrow"><span>ANCHOR REPORT / ROUND 07</span> KIMI K3 · REPORT → ARTIFACTS → INDEPENDENT PROBE</p>
|
||||
<p class="eyebrow"><span>ANCHOR REPORT / ROUND 08</span> KIMI K3 · REPORT → ARTIFACTS → INDEPENDENT PROBE</p>
|
||||
<h1>不把报告压成摘要<br />把每个因果环节<br />重新展开</h1>
|
||||
<p class="lead">
|
||||
K3 同时扩展序列、深度、宽度、视觉与 Agent 轨迹。真正值得读的不是 2.8T 这个最大数字，
|
||||
@@ -129,11 +131,11 @@ const paperGroups = [
|
||||
<dl class="page-facts">
|
||||
<div><dt>QUESTIONS</dt><dd>32 张问题账</dd></div>
|
||||
<div><dt>REPORT</dt><dd>16 Figures · 5 Tables</dd></div>
|
||||
<div><dt>LABS</dt><dd>8 + 4 + 5 + 5 + 5 个交互视图</dd></div>
|
||||
<div><dt>LABS</dt><dd>8 + 4 + 5 + 5 + 5 + 5 个交互视图</dd></div>
|
||||
<div><dt>READING</dt><dd>100 个一手 / 官方节点</dd></div>
|
||||
<div><dt>MODEL</dt><dd>2.78T total / 104.2B active</dd></div>
|
||||
<div><dt>ARTIFACTS</dt><dd>96 shards · 497,220 tensors</dd></div>
|
||||
<div><dt>STATUS</dt><dd>K3 七轮 · 局部路径审计</dd></div>
|
||||
<div><dt>STATUS</dt><dd>K3 八轮 · 训练期消融审计</dd></div>
|
||||
</dl>
|
||||
</div>
|
||||
</header>
|
||||
@@ -977,8 +979,33 @@ const paperGroups = [
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="article-section" id="attnres-forward">
|
||||
<p class="eyebrow"><span>34</span> TRAIN-TIME FORWARD INTERVENTION</p>
|
||||
<h2>诊断期的局部敏感性进入完整训练后，尖峰指标还会衰减吗？</h2>
|
||||
<p class="lede">
|
||||
第八轮不再只改 diagnostic backward，而是让 group 6 / 7 的 parameter-free
|
||||
uniform mixer 进入每一次 train、eval 与 diagnostic forward。四个变体各跑三个
|
||||
8,000-step seed，并把同 seed 的 Round 05 learned run 锁为历史配对 reference；
|
||||
主变体 groups 6+7 的 contrast 与 peak 六格降幅全部超过 20%，同时 BPC 质量门
|
||||
4 / 4 通过，指定 seed 的完整重训 scientific payload exact。
|
||||
</p>
|
||||
<div class="artifact-callout">
|
||||
<article><span>F / FROZEN</span><b>12 formal + 1 replay</b><p>851,968,000 个新 target bytes；每格 65,536,000。</p></article>
|
||||
<article><span>X / ATTENUATION</span><b>6 / 6 PASS</b><p>contrast drop 62.1%–77.3%；peak drop 32.3%–62.0%。</p></article>
|
||||
<article><span>X / QUALITY</span><b>4 / 4 PASS</b><p>三 seed ΔBPC 最大 +0.00960；均值 +0.00658。</p></article>
|
||||
<article class="warning"><span>B / BOUNDARY</span><b>reduced protocol only</b><p>训练期架构消融；不是 K3 checkpoint 或 Figure 5(c) 复现。</p></article>
|
||||
</div>
|
||||
<K3AttnResForwardLab />
|
||||
<div class="hero-actions">
|
||||
<a class="button primary" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_FORWARD_TRAINING_AUDIT.md">阅读完整结果审计</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_FORWARD_TRAINING_PROTOCOL.md">核对预注册协议</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/research/K3_ATTNRES_FORWARD_TRAINING_IMPLEMENTATION_REVIEW.md">查看两阶段实现审阅</a>
|
||||
<a class="button" href="https://git.k1412.top/wuyang/llm-atlas/src/branch/main/experiments/k3/attnres_forward">复跑训练、分析与 replay</a>
|
||||
</div>
|
||||
</section>
|
||||
|
||||
<section class="article-section" id="audit">
|
||||
<p class="eyebrow"><span>34</span> FIGURE & TABLE AUDIT</p>
|
||||
<p class="eyebrow"><span>35</span> FIGURE & TABLE AUDIT</p>
|
||||
<h2>Figure 1–16、Table 1–5：每张图究竟支持什么，不能支持什么</h2>
|
||||
<div class="figure-atlas">
|
||||
{k3FigureAtlas.map(([id, report, title, contract]) => (
|
||||
|
||||
@@ -97,7 +97,7 @@ const workstreams = [
|
||||
<article><span>✓</span><h3>K3 报告已结构化拆解</h3><p>47 页报告目录、151 条参考来源和架构/后训练/系统主线已经提取。</p></article>
|
||||
<article><span>✓</span><h3>17 专题知识图</h3><p>从语言模型基础到评测安全，包含先修依赖和三条贯穿案例。</p></article>
|
||||
<article><span>✓</span><h3>编辑式网站系统</h3><p>响应式导航、章节模板、侧栏、进度、论文链和证据提示组件。</p></article>
|
||||
<article><span>✓</span><h3>一百零九个原创交互视图</h3><p>K3 三轴图、八联报告实验、四联开放工件实验与四轮二十联 AttnRes 独立实验，DeepSeek 四联公式实验、十三联 Base 工件实验、Chat 行为、completion/full-depth、multi-seed、cross-source 与 task-bootstrap CRN 五轮实验，以及语言模型前史、Transformer、表示深度、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>一百一十四个原创交互视图</h3><p>K3 三轴图、八联报告实验、四联开放工件实验与五轮二十五联 AttnRes 独立实验，DeepSeek 四联公式实验、十三联 Base 工件实验、Chat 行为、completion/full-depth、multi-seed、cross-source 与 task-bootstrap CRN 五轮实验，以及语言模型前史、Transformer、表示深度、长上下文、MoE、推理、Agent、多模态、训练系统、推理服务、Scaling、数据工程、数值、Alignment 与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>十七篇首版长文</h3><p>K3、语言模型前史、Transformer、表示/位置/残差、DeepSeek、Scaling、数据工程、长上下文、MoE、后训练、推理、Agent、原生多模态、训练系统、推理服务、数值优化与评测安全专题。</p></article>
|
||||
<article><span>✓</span><h3>语言模型前史深度专题</h3><p>八张独立问题账、33 个正式节点、20 段长文与概率—向量—记忆—对齐四联实验。</p></article>
|
||||
<article><span>✓</span><h3>Transformer 深度专题</h3><p>十张独立问题账、40 个正式节点、21 段正文与 QKV—Mask—多头位置—Block 成本四联实验。</p></article>
|
||||
@@ -109,6 +109,7 @@ const workstreams = [
|
||||
<article><span>✓</span><h3>Kimi K3 五轮梯度定义与深度扩展</h3><p>先确认 Figure 5 没有公开唯一 gradient telemetry 合同，再冻结 16/32 blocks × Baseline/Block × 3 seeds 的 12 个 8,000-step 格。Block 的首尾失衡 6/6 改善但全层 CV 6/6 恶化，两个深度都判为 mixed；指定 32 层格完整重训的模型、优化器与全部冻结字段 exact。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 六轮尖峰轨迹与反向路径</h3><p>严格复用 Round 05 depth-32 Block 的三个正式格：尖峰在 step 500 后形成，六个位置 3/3 seed 可见，四种 reduction 12/12 格稳健。切断 key/softmax 源梯度没有降低尖峰；uniform value-backward 让 contrast 平均下降 70.2%，只判为全局 backward-rule sensitivity。完整 replay 的 16 组冻结字段 exact。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 七轮局部路径双向审计</h3><p>冻结 14 个 same-forward mask，把 groups 6+7 的 16 个 depth mixers 同时放进 sufficiency 与 restoration 两个方向。充分性 6/6 过 50%，恢复性却只有 contrast 3/3 通过、peak 0/3 通过，因此正式状态为 one-sided evidence / localization not established。三个正式格、完整 replay、selector 与 forward identity 全部 exact。</p></article>
|
||||
<article><span>✓</span><h3>Kimi K3 八轮训练期前向干预</h3><p>四个 forward architecture variants × 三 seed × 8,000 steps，加一格完整 replay；groups 6+7 的 contrast / peak 六格降幅全部超过 20%，BPC 质量门 4/4 通过。13 个 raw、自哈希、historical pairing 与 scientific replay 全部过闸；Grok 结果后复算 blocking error 为 0。结论只限固定缩小协议，不是真实 K3 checkpoint 或 Figure 5(c) 复现。</p></article>
|
||||
<article><span>✓</span><h3>FlashKDA RTX 5090 执行闸门</h3><p>隔离 CUDA 13.0 / glibc 2.39 编译 sm_120a wheel；6/6 官方参考逐元素相等，并完成 fixed / varlen、三种 state mode 的 1,800 个 CUDA Event samples。</p></article>
|
||||
<article><span>✓</span><h3>Scaling Laws 深度专题</h3><p>九张账、29 个一手节点、DeepSeek/Kimi 双谱系与曲面—部署—复用—涌现四联实验。</p></article>
|
||||
<article><span>✓</span><h3>数据工程深度专题</h3><p>十二张账、31 个一手节点、DeepSeek/Kimi 双谱系与流水线—去重—混合—改写四联实验。</p></article>
|
||||
@@ -137,7 +138,6 @@ const workstreams = [
|
||||
</div>
|
||||
<div class="queue-table">
|
||||
<div class="head"><b>优先级</b><b>专题</b><b>本轮交付</b><b>完成闸门</b></div>
|
||||
<div><span>P0</span><strong>K3 七轮后续</strong><p>前向训练变体 → 非加性局部交互地图 → 等待 A_log 社区候选的官方裁决后进入真实 checkpoint forward</p><em>局部机制 + 工件边界</em></div>
|
||||
<div><span>P0</span><strong>DeepSeek 八轮后续</strong><p>干预式 mediation → SM90 FlashMLA / FP8 / pipeline traces → R1-like RL 小模型复现</p><em>运行证据 + 独立复现</em></div>
|
||||
<div><span>P0</span><strong>Transformer 二轮</strong><p>多头电路逐图 → Pre/Post-LN 真实 traces → Flash/KV 配置与 kernel 对照</p><em>逐图笔记 + 实测边界</em></div>
|
||||
<div><span>P0</span><strong>表示、位置与残差二轮</strong><p>真实 hidden-state / norm traces → 长上下文位置外推 → mHC / AttnRes 深层稳定性消融</p><em>可复现实验 + 逐图笔记</em></div>
|
||||
|
||||
Reference in New Issue
Block a user