Text Generation
Transformers
Chinese
English
character-level
chinese
gru
mamba
Mixture of Experts
titans
Instructions to use masterofthestars/tianxuan with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use masterofthestars/tianxuan with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="masterofthestars/tianxuan")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("masterofthestars/tianxuan", device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use masterofthestars/tianxuan with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "masterofthestars/tianxuan" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/masterofthestars/tianxuan
- SGLang
How to use masterofthestars/tianxuan with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "masterofthestars/tianxuan" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "masterofthestars/tianxuan", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use masterofthestars/tianxuan with Docker Model Runner:
docker model run hf.co/masterofthestars/tianxuan
File size: 55,795 Bytes
db7bc06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 | import * as fs from "fs";
import * as path from "path";
import { CharGRU, CharGRUConfig, GRUState } from "./gru";
export type { GRUState };
import { CharMultiNeuro, CharMultiNeuroConfig } from "./multineuro";
import { CharTokenizer } from "./tokenizer";
import { CharTransformer, CharTransformerConfig } from "./transformer";
import { mulberry32 } from "./rng";
import { visionFront, speechFront } from "./frontends";
export { visionFront, speechFront };
import { createStation } from "../station";
import { L0Consistency } from "../station/l0probe";
import { Task, TaskResult } from "../station/types";
import { VectorStore } from "./vectorstore";
import { CNVectorStore } from "./cn_vectorstore";
import { MathVectorStore } from "./math_vectorstore";
import { WorldModel } from "./world_model";
import { TitansMemory } from "./titans";
export type { TitansConfig } from "./titans";
import { ConfidenceFloors } from "./floors";
import { chunkText } from "./data";
import { ContextWindow } from "./tool_system";
import { reflect, loadBigrams } from "./reflection";
export interface CheckpointMeta {
config: CharGRUConfig;
paramCount: number;
corpusTokens: number;
epochs: number;
vocab: string[];
trainedAt: string;
/** 是否由 CharMultiNeuro 训练 (含 MoE/专家头) */
multiNeuro?: boolean;
/** 是否由 CharTransformer 训练 */
transformer?: boolean;
moeTopK?: number;
moeGateHidden?: number;
moeNExperts?: number;
moeLoadBalanceWeight?: number;
onlineTitans?: boolean;
/** Transformer 配置 (可选) */
nLayer?: number;
nHead?: number;
/** Titans 记忆日志 (用于恢复记忆) */
titansWriteLog?: Array<{ slot: number; h1: number[] }>;
}
function sampleToken(probs: Float64Array, topK: number, temperature: number, rand: () => number): number {
// [P0-F 修复 2026-10-08 第一性审核] NaN 概率 (非有限 logits 经 logitsToProbs 返回) → 返回 -1,
// 调用方 charOf(-1) 无有效输出 (可证伪: 不伪造满置信的假输出)。
for (let i = 0; i < probs.length; i++) if (Number.isNaN(probs[i])) return -1;
const v = probs.length;
const scaled = new Float64Array(v);
let sum = 0;
for (let k = 0; k < v; k++) {
scaled[k] = Math.exp(Math.log(probs[k] + 1e-12) / Math.max(temperature, 1e-3));
sum += scaled[k];
}
const k = Math.min(topK, v);
const idxs = Array.from({ length: v }, (_, i) => i).sort((a, b) => scaled[b] - scaled[a]).slice(0, k);
let mass = 0;
for (const i of idxs) mass += scaled[i];
let r = rand() * mass;
for (const i of idxs) {
r -= scaled[i];
if (r <= 0) return i;
}
return idxs[idxs.length - 1];
}
export interface GenerateOptions {
prompt: string;
maxLen: number;
budget: number;
temperature: number;
topK: number;
seed?: number;
}
export interface CharRoute {
char: string;
depth: number;
confidence: number;
units: string[];
memoryHit: boolean;
/** 记忆来源:向量库 / 重复上下文 */
memorySource?: "vector" | "repeat" | "titans";
}
export interface GenerateResult {
text: string;
routes: CharRoute[];
budget: number;
}
/** 测试时记忆块标记:写入向量库的生成内容前缀,检索时跳过,防止自举循环 */
const MEMORY_TAG = "\uE000";
/**
* 推理引擎:模型 + 字符分词 + 工作站 (分级路由 + Titans L0 记忆)
* L0 记忆缓存直出 (零算力) → L1 浅路径 → L2 深路径 (预算内扩深)
*/
export class Engine {
readonly model: CharGRU;
readonly tokenizer: CharTokenizer;
readonly meta: CheckpointMeta;
private station = createStation();
private habit: () => number;
private memory = new Map<string, string>();
private ctxWin: number[] = [];
private store: VectorStore | null;
private recentChars: string[] = [];
private state: GRUState;
/** 在线神经记忆 (Titans 逻辑):上下文嵌入 → 下一字,生成期持续写入 */
private titansMem: TitansMemory;
/** 多神经协同模型 (含 MoE/专家头,可为 null) */
private ctxWinWindow: ContextWindow;
readonly multiNeuro: CharMultiNeuro | null;
/** 置信度下限追踪器 (宪章第1条: 兜底值由真实观测滚动窗口推导, 不写死常数) */
readonly floors: ConfidenceFloors;
/** L0 直出一致性探针 (宪章第7/8条: 在线记忆效果的量化证据, 非写死) */
readonly l0Probe: L0Consistency;
/** CN 语义向量库 (HNSW + N-gram 预过滤) */
private cnStore: CNVectorStore | null = null;
/** 数学向量库 (公式检索) */
private mathStore: MathVectorStore | null = null;
/** 推理协处理器 */
private _reasoner: any = null;
/** 世界模型 (物理常识 + 因果推理) */
private worldModel: WorldModel | null = null;
/** 语义反思层:n-gram 统计用于修正生成 */
private bigramStats: ReturnType<typeof loadBigrams> = null;
private constructor(model: CharGRU, tokenizer: CharTokenizer, meta: CheckpointMeta, multiNeuro: CharMultiNeuro | null = null, dataRoot = "data", prodRoot = "data", wmDir = "") {
this.model = model;
this.tokenizer = tokenizer;
this.meta = meta;
this.multiNeuro = multiNeuro;
this.habit = mulberry32(Date.now() % 2147483647);
this.floors = new ConfidenceFloors();
this.l0Probe = new L0Consistency();
this.state = model.newState();
// dataRoot = sandbox 数据根 (语料快照); prodRoot = 生产数据根 (bigram/世界模型/向量库, 只读)。
// 均由 load() 依 checkpoint 绝对路径锚定, 不依赖 cwd (宪章第10条: 生产 data/ 只读)。
// 逐资源: 优先 sandbox 根, 缺则回退生产根 → 既守沙箱纪律, 又不漏统计资源。
const pick = (rel: string): string => {
const sbx = path.join(dataRoot, rel);
return fs.existsSync(sbx) ? sbx : path.join(prodRoot, rel);
};
this.store = VectorStore.load(pick("vectorstore"));
if (this.store) console.log(`向量库已加载:${this.store.size} 块`);
this.cnStore = CNVectorStore.load(pick("cn_vectorstore"));
if (this.cnStore) console.log(`CN 语义库已加载:${this.cnStore.size} 块`);
this.mathStore = MathVectorStore.load(pick("math_vectorstore"));
if (this.mathStore) console.log(`数学库已加载:${this.mathStore.size} 条`);
// 尝试从 checkpoint 目录加载世界模型 (ckpt 自带 → 优先; 否则数据根; 宪章第10条 ckpt 只读)
const wmPath = wmDir && fs.existsSync(path.join(wmDir, "world_model", "world_model.json"))
? path.join(wmDir, "world_model")
: pick("world_model");
this.worldModel = WorldModel.load(wmPath);
if (this.worldModel) {
console.log(`世界模型已加载:${this.worldModel.paramCount()} 参数 (来源: ${wmPath})`);
} else {
console.log(`世界模型未找到:${wmPath}`);
}
this.titansMem = new TitansMemory({ dim: meta.config.emb, slots: 256 });
// 加载 bigram 统计(用于语义反思层)
this.bigramStats = loadBigrams(pick("bigrams.json"));
if (this.bigramStats) {
console.log(`Bigram 统计已加载:${this.bigramStats.total} bigrams, ${this.bigramStats.charFreq.size} unique chars`);
} else {
console.log(`Bigram 统计未找到:${path.join(prodRoot, "bigrams.json")} (将使用默认统计)`);
}
// 上下文窗口:滑动窗口管理对话历史
this.ctxWinWindow = new ContextWindow(meta.config.ctx * 10); // 10倍上下文
// 尝试从多个可能位置加载 Titans 记忆
const possiblePaths = [
path.join("sandbox", "ck-titans-test", "titans_memory.json"),
path.join("data", "checkpoints", "titans_memory.json"),
path.join("..", "sandbox", "ck-titans-test", "titans_memory.json"),
path.join("data", "titans_memory.json"),
];
let memoryPath = "";
for (const p of possiblePaths) {
if (fs.existsSync(p)) {
memoryPath = p;
break;
}
}
// 也检查当前 checkpoint 目录
const cwd = process.cwd();
const altPaths = [
path.join(cwd, "titans_memory.json"),
path.join(cwd, "sandbox", "ck-titans-test", "titans_memory.json"),
];
for (const p of altPaths) {
if (fs.existsSync(p)) {
memoryPath = p;
break;
}
}
if (fs.existsSync(memoryPath)) {
try {
const memData = JSON.parse(fs.readFileSync(memoryPath, "utf8"));
// 使用 Object.assign 绕过 readonly 限制
Object.assign(this.titansMem, {
chars: memData.chars,
keys: new Float64Array(memData.keys),
values: new Float64Array(memData.values),
persistent: new Float64Array(memData.persistent),
});
console.log(`已从文件加载 Titans 记忆:${this.titansMem.chars.filter((c: string) => c).length} 个非空槽位`);
} catch (e) {
console.warn(`加载 Titans 记忆失败:${e}`);
}
}
this.registerUnits();
// [R1-B1 修复 2026-10-09 第一性审核] 注入真实 model.step 上界给经济学:
// 全链全跑 = 注册单元数 (深链可达深度, 同深度取一故 = 最大 depth) + 1 (推进 advance)。
// 动态推导, 不再硬编码 4 (旧 B1: mamba/attn 关闭时分母失真)。
{
const textUnits = this.station.registry.computeUnits("text");
const maxDepth = textUnits.reduce((m, u) => Math.max(m, u.depth), 0);
const reachable = textUnits.filter((u) => u.cost > 0).length; // 单元前向 (同深度取一, ≈ maxDepth)
this.station.economics.setMaxRealForwards(Math.max(reachable, maxDepth) + 1);
}
}
/** 检索:把查询文本向量化,返回库内 top-k 片段 */
retrieve(q: string, k = 4): Array<{ text: string; score: number }> {
if (!this.store) return [];
const ids = this.tokenizer.encode(q);
return this.store.search(this.model.embedAvg(ids), k);
}
/** 中文语义检索 (CNVectorStore + HNSW + N-gram 预过滤) */
retrieveCN(q: string, k = 4, category?: string): Array<{ text: string; score: number; category?: string }> {
if (!this.cnStore) return [];
const ids = this.tokenizer.encode(q);
return this.cnStore.search(this.model.embedAvg(ids), k, category);
}
/** 数学公式检索 (语义 + 结构 + 关键词融合) */
retrieveMath(q: string, k = 6, category?: string): Array<{ expr: string; description: string; score: number; matchType: string }> {
if (!this.mathStore) return [];
const ids = this.tokenizer.encode(q);
const results = this.mathStore.search(q, this.model.embedAvg(ids), k, category);
return results.map(r => ({
expr: r.expr, description: r.description, score: r.score, matchType: r.matchType,
}));
}
/** 推理:概念理解 + 符号推导 + 物理方程求解 */
reason(q: string) {
const results: Array<{ type: string; content: string; score?: number }> = [];
// 1. 向量检索
if (this.cnStore) {
const cnHits = this.cnStore.search(this.model.embedAvg(this.tokenizer.encode(q)), 3);
for (const h of cnHits) results.push({ type: "cn", content: h.text, score: h.score });
}
if (this.mathStore) {
const mathHits = this.mathStore.search(q, this.model.embedAvg(this.tokenizer.encode(q)), 3);
for (const h of mathHits) results.push({ type: "math", content: h.expr + ": " + h.description, score: h.score });
}
// 2. 世界模型物理约束检查
if (this.worldModel) {
const traj = this.worldModel.simulate(
new Float64Array(this.model.embedAvg(this.tokenizer.encode(q))),
[0],
);
const phys = this.worldModel.evaluatePhysics(traj[0]?.state);
results.push({
type: "world_model",
content: `物理约束:能量=${phys[0].toFixed(3)} 因果=${phys[1].toFixed(3)} 平滑=${phys[2].toFixed(3)}`,
score: 1 - (phys[0] + phys[1] + phys[2]) / 3,
});
}
return { query: q, results, worldModel: this.worldModel };
}
retrieveCount(): number {
return this.store?.size ?? 0;
}
/**
* 测试时记忆:把生成内容切块写入向量库 (可持久化)。
* 宪章: 生产 data/ 只读 — 持久化目标参数化, 默认内存态不落盘。
* 传 dir 且指向沙箱路径时才写盘; 缺省仅在内存累加。
*/
remember(text: string, persistDir?: string): void {
if (!this.store) return;
let added = 0;
for (const chunk of chunkText(text, 64, 32)) {
if (chunk.length < 8) continue;
const ids = this.tokenizer.encode(chunk);
this.store.add(MEMORY_TAG + chunk, this.model.embedAvg(ids));
added += 1;
}
if (persistDir !== undefined && persistDir !== "") {
this.store.save(persistDir);
if (added > 0) console.log(`测试时记忆:新写入 ${added} 块 → ${persistDir},向量库现有 ${this.store.size} 块`);
} else if (added > 0) {
console.log(`测试时记忆:新写入 ${added} 块 (仅内存, 未持久化),向量库现有 ${this.store.size} 块`);
}
}
/** 采样前降权:近 16 个已出字符中重复≥2 次的禁出 + UNK 槽禁出 (防 12 字级自举循环) */
private penalized(probs: Float64Array): Float64Array {
const out = probs.slice();
out[out.length - 1] = 0;
const freq = new Map<string, number>();
for (const ch of this.recentChars) freq.set(ch, (freq.get(ch) ?? 0) + 1);
for (const [ch, count] of freq) {
if (count < 2) continue;
const id = this.tokenizer.idOf(ch);
if (id !== undefined && id < out.length) out[id] = 0;
}
let sum = 0;
for (let i = 0; i < out.length; i++) sum += out[i];
for (let i = 0; i < out.length; i++) out[i] /= sum;
return out;
}
private propose(ch: string): void {
this.recentChars.push(ch);
if (this.recentChars.length > 16) this.recentChars.shift();
}
/** 窥视:用当前状态副本推一步,不推进真实状态 */
private peekLogits(depth: 1 | 2 | 3 | 4): Float64Array {
const lastId = this.ctxWin.length > 0 ? this.ctxWin[this.ctxWin.length - 1] : 0;
const copy: GRUState = {
h1: this.state.h1.slice(),
h2: this.state.h2.slice(),
hist: this.state.hist.map((h) => h.slice()),
ids: [...this.state.ids],
};
return this.model.step(lastId, copy, depth);
}
/** 推进:真实状态吸收新字符 (depth = 路由实际执行到的最大深度, 默认 4) */
private advance(chId: number, depth: 1 | 2 | 3 | 4 = 4): void {
this.model.step(chId, this.state, depth);
// SNN 膜电位锁相推进 (与骨干同频, 每字一次)
if (this.multiNeuro?.spike) {
this.multiNeuro.spikeAdvance(chId);
}
}
/** SNN 膜电位重置 (生成开始时调用) */
private resetSpike(): void {
if (this.multiNeuro?.spike) {
this.multiNeuro.resetSpikeSeq();
}
}
/**
* [P0-B 修复 2026-10-08 第一性审核] 上下文指纹: 稳定 FNV-1a 哈希 (非 seed, 不随随机性变),
* 相同 ctxWin → 相同指纹。mem: 键绑定此指纹, 不同上下文的残留键天然失效。
*/
static ctxFingerprint(ctx: number[], vocabSize: number): string {
let h = 2166136261;
for (let i = 0; i < ctx.length; i++) {
h ^= ctx[i] % vocabSize;
h = Math.imul(h, 16777619);
}
return (h >>> 0).toString(16).padStart(8, "0");
}
private recallAt(ctxIds: number[]): { char: string; source: "vector" | "repeat" | "titans" } | null {
if (this.store) {
const q = this.tokenizer.decode(ctxIds);
const hits = this.store.search(this.model.embedAvg(ctxIds), 8);
for (const hit of hits) {
if (hit.text.startsWith(MEMORY_TAG)) continue;
const at = hit.text.indexOf(q);
if (at >= 0 && at + q.length < hit.text.length) {
const ch = hit.text[at + q.length];
if (this.isDegenerate(ch)) return null;
// [P0-A 补漏] 向量库召回: 历史上该上下文真实延续过 → 0.5 点火标记 (非写死满置信, 宪章第1条);
// executor L0 分支读 z:l0:act 推导置信度, 无写入 → 0, 冷启动不伪造满置信
this.station.blackboard.write("z:l0:act", 0.5);
return { char: ch, source: "vector" };
}
}
}
const cached = this.memory.get(ctxIds.join(","));
if (cached !== undefined && !this.isDegenerate(cached)) {
// 重复上下文缓存: 深链曾真实出过该字 → 0.5 点火标记 (同上)
this.station.blackboard.write("z:l0:act", 0.5);
return { char: cached, source: "repeat" };
}
// Titans 在线神经记忆:上下文嵌入检索 → 槽位存储的下一字符
const q = Float64Array.from(this.model.embedAvg(ctxIds));
const cache = { query: new Float64Array(0), keys: new Float64Array(0), values: new Float64Array(0), weights: new Float64Array(0) };
const mem = this.titansMem.read(q, cache);
if (mem.hitSlots > 0 && cache.weights.length > 0) {
let best = 0;
for (let i = 1; i < cache.weights.length; i++) if (cache.weights[i] > cache.weights[best]) best = i;
if (cache.weights[best] > 0.5) {
// 直接从 Titans 获取字符,不再依赖外部映射
const ch = this.titansMem.chars[best % this.titansMem.cfg.slots];
if (ch !== undefined && ch !== "" && !this.isDegenerate(ch)) {
const l0Act = mem.hitSlots / this.titansMem.cfg.slots; // 点火 = 命中槽占比 (真实数值)
this.station.blackboard.write("z:act:titans", l0Act);
this.station.blackboard.write("z:l0:act", l0Act);
return { char: ch, source: "repeat" };
}
}
}
return null;
}
/** 直接读取 Titan 记忆输出(用于 L0 直出) */
private getTitansOutput(ctxIds: number[]): string | null {
const q = Float64Array.from(this.model.embedAvg(ctxIds));
const cache = { query: new Float64Array(0), keys: new Float64Array(0), values: new Float64Array(0), weights: new Float64Array(0) };
const mem = this.titansMem.read(q, cache);
if (mem.hitSlots > 0 && cache.weights.length > 0) {
// 找到权重最大的多个槽位
const topSlots: number[] = [];
let maxWeight = 0;
for (let i = 0; i < cache.weights.length; i++) {
if (cache.weights[i] > maxWeight) {
maxWeight = cache.weights[i];
topSlots.length = 0;
topSlots.push(i);
} else if (cache.weights[i] > maxWeight * 0.9) { // 接近最大的也考虑
topSlots.push(i);
}
}
if (topSlots.length > 0 && maxWeight > 0.3) {
// 从 top 槽位中选择一个非空字符
for (const slot of topSlots) {
const ch = this.titansMem.chars[slot % this.titansMem.cfg.slots];
if (ch !== undefined && ch !== "" && !this.isDegenerate(ch)) {
this.station.blackboard.write("z:act:titans", mem.hitSlots / this.titansMem.cfg.slots);
return ch;
}
}
}
}
return null;
}
/** L0 一致性探针 (宪章第7/8条): 同上下文深链 top-1 字 (peekLogits 无副作用: 状态副本推进, 不污染当前生成状态) */
deepTop1At(ctxIds?: number[]): { char: string; p: number } {
const logits = this.peekLogits(4);
const probs = this.model.logitsToProbs(logits);
let best = 0;
for (let i = 1; i < probs.length; i++) if (probs[i] > probs[best]) best = i;
return { char: this.tokenizer.charOf(best), p: probs[best] };
}
/** 循环防护:连续相同字符 ≥4 次或最近 8 字符整段重复,判定为退化输出 */
private isDegenerate(ch: string): boolean {
// 只检测连续重复,不检测全局频率
const n = this.recentChars.length;
if (n >= 3) {
const last3 = this.recentChars.slice(-3).join("");
if (last3 === ch.repeat(3)) return true;
}
if (n >= 16) {
const a = this.recentChars.slice(-8).join("");
const b = this.recentChars.slice(-16, -8).join("");
if (a === b) return true;
}
return false;
}
static load(checkpointDir: string): Engine {
const meta = JSON.parse(fs.readFileSync(path.join(checkpointDir, "meta.json"), "utf-8")) as CheckpointMeta;
const params = JSON.parse(fs.readFileSync(path.join(checkpointDir, "model.json"), "utf-8")) as number[];
const tokenizer = new CharTokenizer();
tokenizer.load(meta.vocab);
// [P0-F 修复 2026-10-08 第一性审核] 推理侧非有限参数拒载: 检查点若被污染 (含 NaN/Inf),
// 拒绝载入 — 旧实现 Engine.load 直接 load(params) 不扫描, 污染检查点 (85.8% +Inf)
// 载入后置信度恒 1.0, 健康监控 nan:0 永不降级, 生成 avgFinalConf≈1.0 而真实输出是噪声。
// 与 train.ts resume 防护 (防线 3) 同逻辑: 任一非有限 → 抛错拒载。
{
let bad = 0;
for (let i = 0; i < params.length; i++) if (!Number.isFinite(params[i])) bad++;
if (bad > 0) {
throw new Error(
`[推理防护] ${checkpointDir}/model.json 含 ${bad}/${params.length} 非有限参数 (NaN/Inf), ` +
`检查点已被污染, 拒绝载入。请回退到更早的完好检查点。`
);
}
}
let multiNeuro: CharMultiNeuro | null = null;
let baseModel: CharGRU;
if (meta.transformer) {
// Transformer 检查点路径
const tCfg: CharTransformerConfig = {
vocabSize: tokenizer.vocabSize,
emb: (meta.config as any).emb ?? 64,
nLayer: (meta.config as any).nLayer ?? 2,
nHead: (meta.config as any).nHead ?? 2,
ctx: (meta.config as any).ctx ?? 8,
bptt: (meta.config as any).bptt ?? 32,
};
const transformer = new CharTransformer(tCfg, 1);
if (transformer.paramCount() !== params.length) {
throw new Error(`checkpoint 参数数量与配置不匹配:${transformer.paramCount()} != ${params.length}`);
}
transformer.load(params);
// 包装为 GRU 接口兼容层
baseModel = new CharGRU({
vocabSize: tCfg.vocabSize,
emb: tCfg.emb,
hidden: tCfg.emb,
ctx: tCfg.ctx,
bptt: tCfg.bptt,
}, 1);
// 将 transformer 包装到 baseModel 上
(baseModel as any)._transformer = transformer;
(baseModel as any).step = function(xId: number, state: GRUState, depth?: number): Float64Array {
const t = (this as any)._transformer;
if (!t) return new Float64Array(0);
const tState: any = (this as any)._tState ?? ((this as any)._tState = t.newState());
// 路由深度: 允许 1..nLayer 裁剪 (MoD 预算内扩深/裁剪)
const d = Math.max(1, Math.min(depth ?? t.cfg.nLayer, t.cfg.nLayer));
return t.step(xId, tState, d) as Float64Array;
};
} else if (meta.multiNeuro) {
const cfg: CharMultiNeuroConfig = {
...meta.config,
moeTopK: meta.moeTopK ?? 2,
moeGateHidden: meta.moeGateHidden ?? 32,
moeLoadBalanceWeight: meta.moeLoadBalanceWeight ?? 0.01,
onlineTitans: meta.onlineTitans ?? false,
moeNExperts: meta.moeNExperts ?? 4,
// SNN 脉冲支路 (独立参数, 与 checkpoint 训练时一致)
snn: (meta.config as CharMultiNeuroConfig).snn ?? false,
snnLeak: (meta.config as CharMultiNeuroConfig).snnLeak,
snnThresh: (meta.config as CharMultiNeuroConfig).snnThresh,
};
multiNeuro = new CharMultiNeuro(cfg, 1);
if (multiNeuro.paramCount() !== params.length) {
throw new Error(`checkpoint 参数数量与配置不匹配:${multiNeuro.paramCount()} != ${params.length}`);
}
multiNeuro.load(params);
baseModel = multiNeuro.model;
// 恢复 Titans 记忆 (如果有的话)
if (meta.onlineTitans && meta.titansWriteLog && meta.titansWriteLog.length > 0) {
const log = multiNeuro.getTitansWriteLog();
for (const rec of meta.titansWriteLog as any[]) {
const h1 = new Float64Array(rec.h1);
log.push({ slot: rec.slot, h1: h1 });
}
console.log(`恢复 Titans 记忆:${meta.titansWriteLog.length} 条记录`);
// 直接从日志重建 Titans 记忆槽位
const titansMem = new TitansMemory({ dim: meta.config.emb, slots: 256 });
const dim = meta.config.emb;
for (const rec of meta.titansWriteLog as any[]) {
const h1 = new Float64Array(rec.h1);
const slot = rec.slot % titansMem.cfg.slots;
// 从词表中选择合适的字符(高频字)
const vocab = meta.vocab;
titansMem.chars[slot] = vocab[rec.slot % vocab.length] || "。";
for (let i = 0; i < dim; i++) {
titansMem.keys[slot * dim + i] = h1[i];
titansMem.values[slot * dim + i] = h1[i];
}
}
console.log(`重建 Titans 记忆槽位:${titansMem.chars.filter((c: string) => c).length} 个非空槽位`);
}
} else {
baseModel = new CharGRU(meta.config, 1);
if (baseModel.paramCount() !== params.length) {
throw new Error(`checkpoint 参数数量与配置不匹配:${baseModel.paramCount()} != ${params.length}`);
}
baseModel.load(params);
}
// 统一在这里创建 Engine,避免 multiNeuro=null 导致多神经协同功能失效
// 宪章第10条: 生产 data/ 只读 — 数据根从 checkpoint 目录推导 (sandbox 检查点 → <root>/sandbox/data, 生产检查点 → <root>/data)
// [FIX 2026-10-06] 旧逻辑用相对 "sandbox/data", 脚本 cd sandbox 后跑 generate 时拼成
// sandbox/sandbox/data → bigram/世界模型静默缺失 (gen-smoke 置信 0.00 即因此)。
// 现锚定到 ck 绝对路径里 sandbox/data 段的位置, 子目录检查点 (checkpoints-mamba/…) 也解析对。
const ckAbs = path.resolve(checkpointDir);
const parts = ckAbs.split(path.sep);
const iSbx = parts.lastIndexOf("sandbox");
const iData = parts.lastIndexOf("data");
// 项目根 = sandbox 段 (或生产 ck 的 data 段) 之前的那一段 → 全绝对, 不依赖 cwd。
const projRoot = (iSbx >= 0 ? parts.slice(0, iSbx) : iData >= 0 ? parts.slice(0, iData) : []).join(path.sep)
|| process.cwd();
// sandbox 数据根: 语料快照; 生产数据根: bigram/世界模型/向量库 (只读)
const dataRoot = iSbx >= 0 ? path.join(projRoot, "sandbox", "data") : path.join(projRoot, "data");
const prodRoot = path.join(projRoot, "data");
const eng = new Engine(baseModel, tokenizer, meta, multiNeuro, dataRoot, prodRoot, checkpointDir);
return eng;
}
private registerUnits(): void {
this.registerUnitsTo(this.station.registry, this.station.economics);
}
/**
* [P0-2 修复 2026-10-13 第一性审核] 把引擎侧计算单元注册到任意工作站注册表,
* 并同步注入经济学的真实 model.step 上界。
* 背景: SessionEngine 每会话自建 createStation(), 旧实现其 station 注册表只有
* SEED 种子 (text 链零单元) → executor 恒走「零单元」分支, MoD 路由/审计/经济性
* 在服务器路径上是死代码 (真实输出字来自 peekLogits(4) 全深兜底, 审计记 depth=0)。
* 现: SessionEngine 构造时对 baseEngine 调本方法, 把与共享 Engine 完全相同的
* 计算单元 (titans/gru/mamba/attn/cnn/snn/world/vision/speech/output) 注入
* 会话工作站 → 服务器路径多神经协同真跑 (宪章第1/2条)。
* 单元 forward 闭包引用的 engine/model/titansMem/floors 均为 baseEngine 共享只读
* 资源; 黑板键带 task.id (每会话独立编号+上下文指纹), 会话间无串读。
*/
registerUnitsTo(reg: import("../station/registry").Registry, economics?: import("../station/economics").OnDemandEconomics): void {
const engine = this;
// L0 — Titans 记忆 (零算力直出,depth 0)
// [P1-H 修复 2026-10-08 第一性审核] structural 白名单化: 豁免须 evidence.status!=="fail"
// + note 含 checkpoint 证据物。titans SEED 证据 status="fail" (一致性 33.8% < 0.5 门槛)
// → 不得豁免, 保持 enabled=false (与 SEED registry 一致)。
if (this.meta.onlineTitans) {
reg.register({
kind: "titans", name: "Titans 记忆", role: "L0 零算力记忆直出", stage: "phase3", enabled: false, status: "idle",
structural: true,
justification: "在线神经记忆:上下文嵌入检索 → 槽位字符输出 (已训 checkpoint 结构层); SEED 证据 L0 一致性 33.8% < 0.5 门槛 (fail) → 不得豁免 (P1-H 白名单化), 保持货架待大语料重测",
evidence: { metric: "L0 直出一致性 (探针 l0-probe-check.js)", baseline: 0.5, result: 0.338, status: "fail", note: "2026-10-05 L0 一致性探针: 命中率 77.1% 但一致性 33.8% < 0.5 — 省算力真实, 正确性收益未过门槛; P1-H 白名单化后 fail 证据不得 structural 豁免 (宪章第7条), 待健康检查点重测" },
compute: {
unitId: "titans:depth0", depth: 0, cost: 0, chains: ["text"],
forward: async (task, state) => {
// 直接从 TitansMemory 检索最佳槽位
const ctx = task.input as number[];
const q = Float64Array.from(engine.model.embedAvg(ctx));
const cache = { query: new Float64Array(0), keys: new Float64Array(0), values: new Float64Array(0), weights: new Float64Array(0) };
const mem = engine.titansMem.read(q, cache);
if (mem.hitSlots > 0 && cache.weights.length > 0) {
let best = 0;
for (let i = 1; i < cache.weights.length; i++) if (cache.weights[i] > cache.weights[best]) best = i;
if (cache.weights[best] > 0.3) {
const ch = engine.titansMem.getChar(best % engine.titansMem.cfg.slots);
if (ch !== undefined && ch !== "" && !engine.isDegenerate(ch)) {
// 同时写入 mem 和 o,确保 L0 路径能检测到
state.data[`mem:${task.id}`] = ch;
state.data[`o:${task.id}`] = ch;
state.data["z:act:titans"] = mem.hitSlots / engine.titansMem.cfg.slots;
const conf = Math.min(1, mem.hitSlots / 5); // 归一化置信度
engine.floors.record("titans:depth0", conf);
return conf;
}
}
}
// 无命中兜底: 由本单元真实观测的最低置信度推导 (冷启动=0, 非写死常数, 宪章第1条)
return engine.floors.fallback("titans:depth0");
},
},
});
}
// 感知神经:第 1 层 GRU(emb → 状态流),神经链入口
reg.register({
kind: "gru", name: "GRU L1", role: "感知:字符嵌入 → 状态流", stage: "phase1", enabled: true, status: "idle",
roleTag: "perception",
structural: true,
justification: "基线状态流;无此层无任何输出 (已训骨干结构层, 推理分布一致性豁免证据门槛)",
compute: {
unitId: "gru:shallow", depth: 1, cost: 1, chains: ["text"],
forward: async (task, state) => {
const logits = engine.peekLogits(1);
state.data[`z:logits:${task.id}`] = logits;
state.data["z:act:gru:shallow"] = engine.model.confidence(logits);
return engine.model.confidence(logits);
},
},
});
if (engine.model.cfg.mamba) {
// 中央神经 (快动力学):真选择性 SSM(depth 2),点火 = Δ 均值
// [P1-H 修复 2026-10-08 第一性审核] structural 白名单化: mamba SEED 证据 status="fail"
// (3 种子均值 Δ=+0.4pt < +1.0 门槛) → 不得豁免, 保持 enabled=false (宪章第7条)。
reg.register({
kind: "mamba", name: "Mamba", role: "快动力学:选择性 SSM 状态流", stage: "phase1", enabled: false, status: "idle",
roleTag: "central",
structural: true,
justification: "对照基线:线性状态建模 vs 纯骨干(同配置) (已训 checkpoint 中央结构层); P1-H 白名单化后 fail 证据不得豁免, 待大真实语料重测过门槛",
evidence: { metric: "事件语料 top-1(30k×4ep, 3 种子均值单变量消融)", baseline: 10.4, result: 10.8, status: "fail", note: "2026-10-05 ablation-mamba-multiseed.sh: 3 种子均值 Δ=+0.4pt < +1.0 门槛 → 未过 (单种子7 的 +1.7 为运气, seed3/11 Δ=0.0)。容量混淆已控 (on 190k < base 233k); 小语料噪声淹没效应; P1-H 白名单化: fail 证据不得 structural 豁免 (宪章第7条), 待大真实语料重测" },
compute: {
unitId: "mamba:ssm", depth: 2, cost: 2, chains: ["text"],
forward: async (task, state) => {
const logits = engine.peekLogits(2);
state.data[`z:logits:${task.id}`] = logits;
state.data["z:act:mamba:ssm"] = Math.min(1, engine.model.ssm!.lastDeltaMean / 1.5);
return engine.model.confidence(logits);
},
},
});
} else {
// 中央神经 (快动力学):GRU 第 2 层
reg.register({
kind: "gru", name: "GRU L2", role: "快动力学:第二层 GRU 状态流", stage: "phase1", enabled: true, status: "idle",
roleTag: "central",
structural: true,
justification: "共享底座深路径 (已训骨干结构层, 推理分布一致性豁免证据门槛)",
compute: {
unitId: "gru:deep", depth: 2, cost: 2, chains: ["text"],
forward: async (task, state) => {
const logits = engine.peekLogits(2);
state.data[`z:logits:${task.id}`] = logits;
state.data["z:act:gru:deep"] = engine.model.confidence(logits);
return engine.model.confidence(logits);
},
},
});
}
if (engine.model.cfg.attn) {
// 中央神经 (慢语义):学习型自注意力 (depth 3),点火 = gate
reg.register({
kind: "sparse-attention", name: "稀疏 Attention", role: "慢语义:学习型自注意力", stage: "phase2", enabled: true, status: "idle",
roleTag: "central",
structural: true,
justification: "跨位置信息混合,独立于循环路径;真实计算单元 (已训 checkpoint 中央结构层, 推理分布一致性豁免证据门槛)",
evidence: { metric: "事件语料 top-1(30k×2ep)", baseline: 49.1, result: 47.8, status: "pending", note: "机制正确+gate 选择性;任务级证据未达成; 已训 checkpoint 结构层 (attn 中央层), P1-H 白名单: pending+checkpoint 证据物允许豁免, 独立 Attention 系统仍守门槛" },
compute: {
unitId: "attn:layer3", depth: 3, cost: 3, chains: ["text"],
forward: async (task, state) => {
const logits = engine.peekLogits(3);
state.data[`z:logits:${task.id}`] = logits;
state.data["z:act:attn:layer3"] = engine.model.lastAttnGate;
return engine.model.confidence(logits);
},
},
});
}
if (engine.model.cfg.cnn) {
// 感知神经 (局部):1D 卷积 n-gram(depth 4),点火 = 特征激活均值
reg.register({
kind: "cnn", name: "CNN", role: "局部感知:1D 卷积 n-gram 特征", stage: "phase1", enabled: true, status: "idle",
roleTag: "perception",
justification: "局部字符模式特征,独立预测头与循环/注意力输出融合 (pass 证据直接过门槛, P1-H: 无需 structural 豁免)",
evidence: { metric: "事件语料 top-1(30k×4ep 同配置消融)", baseline: 47.5, result: 49.3, status: "pass", note: "gradcheck-cnn=0.0000; 修复 top-1 缺口 +1.8pt; 就后召回探针 9.7→6.5(小规模代价,大语料复查); P1-H: pass 证据直接过门槛, 无需 structural 豁免" },
compute: {
unitId: "cnn:layer4", depth: 4, cost: 1, chains: ["text"],
forward: async (task, state) => {
const logits = engine.peekLogits(4);
state.data[`z:logits:${task.id}`] = logits;
state.data["z:act:cnn:layer4"] = engine.model.lastCnnAct;
return engine.model.confidence(logits);
},
},
});
}
// SNN 脉冲支路 (快动力学·脉冲门控中央神经, 独立 LIF 膜电位)
if (engine.multiNeuro?.spike) {
reg.register({
kind: "snn", name: "SNN SpikeSSM", role: "快动力学·脉冲门控:LIF 膜电位 + PULSE surrogate", stage: "phase1", enabled: false, status: "shelf",
roleTag: "central",
justification: "独立 LIF 脉冲支路, 与骨干零共享参数, 输出端加法融合 logits (2026-10-04 降 shelf: 原 pass 证据混淆未控, 待重消融)",
evidence: { metric: "事件语料 top-1(消融)", baseline: 0, result: 0, status: "pending", note: "SNN 支路已接入训练/推理; 旧消融 Δ=+27.9pt 因参数不等/单种子混淆降 pending, 重消融通过门槛前保持 shelf (宪章第7条)" },
compute: {
unitId: "snn:spike", depth: 2, cost: 2, chains: ["text"],
forward: async (task, state) => {
const sLogits = engine.multiNeuro!.spikeReadLast();
if (sLogits) {
state.data[`z:spike:${task.id}`] = sLogits;
state.data["z:act:snn:spike"] = engine.multiNeuro!.spikeIgnition();
const conf = engine.model.confidence(sLogits);
engine.floors.record("snn:spike", conf);
return conf;
}
// 无脉冲输出兜底: 由本单元真实观测的最低置信度推导 (冷启动=0, 非写死常数, 宪章第1条)
return engine.floors.fallback("snn:spike");
},
},
});
}
// 世界域中央神经: 世界模型状态转移 + 物理一致性 (已训 checkpoint 结构层,
// 仅 world 任务按需调用 — 每字符 text 生成零额外开销, 宪章「按需调用」)
if (this.worldModel) {
const eng = this;
reg.register({
kind: "world-model", name: "World Model", role: "世界域中央: 状态转移 + 物理一致性 (Dreamer/Genie 逻辑自研)", stage: "deferred", enabled: true, status: "idle",
roleTag: "central",
structural: true,
justification: "已训 checkpoint 结构层 (world_model.json, 编码器/转移/解码器/物理约束); 仅 world 任务按需调用, text 生成零额外开销",
evidence: { metric: "世界一致性置信度 (任务级缺口消融)", baseline: 0, result: 0, status: "pending", note: "已训 checkpoint 结构层 (world_model.json 加载侧 paramCount 校验通过); P1-H 白名单: pending+checkpoint 证据物允许 structural 豁免, 待补「世界模型单元 vs 无」消融回填 (宪章第7条)" },
compute: {
unitId: "world-model:wm", depth: 3, cost: 2, chains: ["world"],
forward: async (task, state) => {
const wm = eng.worldModel!;
const ids = (task.input as number[] | undefined) ?? [];
const obs = ids.length > 0 ? Float64Array.from(eng.model.embedAvg(ids)) : new Float64Array(wm.embDim);
const ws = wm.encode(obs, 0);
const phys = wm.evaluatePhysics(ws);
// 置信度 = 状态相干性 × 能量守恒一致性 (真实数值, 无写死)
const conf = Math.max(0, Math.min(1, ws.coherence * (0.5 + 0.5 * (1 - phys[0]))));
state.data[`z:world:phys:${task.id}`] = Array.from(phys);
state.data[`z:world:state:${task.id}`] = Array.from(ws.s);
state.data[`z:act:world-model:wm`] = conf;
return conf;
},
},
});
}
// 视觉域感知前端: 强度网格 → 真实边缘/结构响应 (自研数值, 非外部模型/写死)
// 宪章第7条: 域消融 v2 (vision-ablation-v2.js) 分离度 0.8929 ≥ 0.2 门槛 → 过,
// enabled; v1 (8×8+噪声) 0.129 未过记录在案, v2 强化协议 (32×32 + 阶梯刺激 + 阴性对照 + 噪声鲁棒) 通过
// [P1-H 修复 2026-10-08 第一性审核] 零参数数值核 (vision/speech 前端, 非已训 checkpoint 结构层)
// 不得标 structural — 旧实现带 structural 短路豁免, 白名单化后移除。
reg.register({
kind: "vision", name: "Vision 感知前端", role: "视觉域感知: 强度网格 → 真实 LoG/Sobel 结构响应 (自研数值)", stage: "deferred", enabled: true, status: "idle",
roleTag: "perception",
justification: "自研视觉前端 (frontends.ts visionFront): max|Lap|/mean|Sobel| 不连续度集中度 → 归一化置信度; 非写死 (宪章第1条); 域消融 v2 过门槛 (pass 证据)",
evidence: { metric: "视觉域结构检测 (锐利/棋盘 vs 平滑, 32×32, 3 种子×200)", baseline: 0, result: 0.8929, status: "pass", note: "2026-10-05 消融 v2 (vision-ablation-v2.js): 分离度 0.8929≥0.2 (meanA=0.9214 vs meanB=0.0285), 阴性C=0.0000<0.05, 单调 C<B<A 3/3, 含噪声鲁棒子类; v1 (8×8+噪声) 0.129 未过记录在案; P1-H: 零参数数值核, pass 证据直接过门槛, 无需 structural 豁免" },
compute: {
unitId: "vision:frontend", depth: 1, cost: 1, chains: ["vision"],
forward: async (task, state) => {
const input = (task.input as number[] | undefined) ?? [];
const rep = visionFront(input.length > 0 ? input : [0, 0]);
state.data[`z:vision:feat:${task.id}`] = rep.summary;
state.data["z:act:vision:frontend"] = rep.confidence;
return rep.confidence;
},
},
});
// 声学域感知前端: 声学时序 → RMS/过零率/谱通量 (自研数值, 非外部模型/写死)
// 输出神经:整合所有中央神经的 logits 到最终输出向量 (MoE 聚合头风格)
// [P1-1 修复 2026-10-13 第一性审核] 本单元为零参数整合层 (共享底座最深层 logits 经黑板落槽),
// 无真实消融数字 (旧 evidence baseline=0/result=1.0 为机制一致性自证, 非对照组对比) →
// P1-1 数值门槛下不得 enabled (宪章第7条「数字说话」), 保持货架待补「output:aggregate vs 单头纯骨干」
// 真实缺口消融 (top-1 对照) 回填 evidence.status="pass" + 真实数字后重启用。
reg.register({
kind: "output", name: "输出神经", role: "多神经协同的最终输出聚合器 (整合所有中央神经的 logits 到最终输出向量)", stage: "phase3", enabled: false, status: "shelf",
roleTag: "output",
justification: "输出聚合:整合所有中央神经的 logits 到最终输出向量,支持动态路由门控与负载均衡;使用骨干最深层 logits 作为最终输出 (零参数整合层, 无独立参数)",
evidence: { metric: "输出整合一致性 (输出神经聚合 vs 单头纯骨干 top-1 对照)", baseline: 0, result: 0, status: "pending", note: "2026-10-13 第一性审核 P1-1: 旧 baseline=0/result=1.0 为机制一致性自证 (logits 切路已验证), 非真实消融数字; 零参数整合层无独立 checkpoint 证据物, 不得 structural 豁免; 待补「output:aggregate 经黑板落槽 vs 直接读骨干 logits」真实 top-1 对照消融 (3 种子同配置) 过 +1.0pt 门槛后回填 pass (宪章第7条)" },
compute: {
unitId: "output:aggregate", depth: 5, cost: 1, chains: ["text"],
forward: async (task, state) => {
// 调用骨干最深层 logits 作为聚合输出 (共享底座自然聚合所有前面神经的计算结果)
const logits = this.peekLogits(4);
state.data[`o:${task.id}`] = logits;
return this.model.confidence(logits);
},
},
});
// 宪章第7条: 域消融过门槛 (谐波 vs 静默 分离度 0.2450 ≥ 0.2, 3 种子×200, vision-speech-ablation.js)
// [P1-H 修复 2026-10-08 第一性审核] 零参数数值核, 不得标 structural (pass 证据直接过门槛)
reg.register({
kind: "speech", name: "Speech 感知前端", role: "声学域感知: 声学时序 → RMS/过零率/谱通量 (自研数值)", stage: "deferred", enabled: true, status: "idle",
roleTag: "perception",
justification: "自研声学前端 (frontends.ts speechFront): 时序样本 → RMS 能量 + 过零率 + 一阶谱通量近似 → 归一化置信度; 非写死 (宪章第1条); 域消融过门槛 (pass 证据)",
evidence: { metric: "声学域活跃度 (谐波 vs 静默 类间分离度, 3 种子×200)", baseline: 0, result: 0.245, status: "pass", note: "2026-10-05 消融 (vision-speech-ablation.js): 分离度 0.2450≥0.2 门槛 (meanA=0.5821 vs meanB=0.3368), 过控 = 谐波信号显著高于静默噪声; 真实数值, 无写死; P1-H: 零参数数值核, pass 证据直接过门槛, 无需 structural 豁免" },
compute: {
unitId: "speech:frontend", depth: 1, cost: 1, chains: ["speech"],
forward: async (task, state) => {
const input = (task.input as number[] | undefined) ?? [];
const rep = speechFront(input.length > 0 ? input : [0]);
state.data[`z:speech:feat:${task.id}`] = rep.summary;
state.data["z:act:speech:frontend"] = rep.confidence;
return rep.confidence;
},
},
});
// [P0-2 配套 2026-10-13] 注入经济学的真实 model.step 上界 (R1-B1 口径):
// 全链全跑 = 注册链可达单元数 (同深度取一 ≈ 最大 depth) + 1 (推进 advance)。
// 与 Engine 构造器内同名推导一致, 避免会话工作站沿用默认值失真。
if (economics) {
const textUnits = reg.computeUnits("text");
const maxDepth = textUnits.reduce((m, u) => Math.max(m, u.depth), 0);
const reachable = textUnits.filter((u) => u.cost > 0).length;
economics.setMaxRealForwards(Math.max(reachable, maxDepth) + 1);
}
}
private options: { temperature: number; topK: number } = { temperature: 0.1, topK: 3 };
async generate(opts: GenerateOptions): Promise<GenerateResult> {
const { prompt, maxLen, budget } = opts;
this.l0Probe.reset(); // L0 一致性探针: 单次生成分母干净 (各 run 独立统计)
this.options = { temperature: opts.temperature, topK: opts.topK };
if (opts.seed !== undefined) this.habit = mulberry32(opts.seed);
// 关键:prompt 必须先推进真实状态,模型才能"看到"完整上下文
this.state = this.model.newState();
this.resetSpike(); // SNN 膜电位重置 (生成开始时)
if (this.multiNeuro?.spike) this.multiNeuro.setSpikeInferenceMode(true); // 推理侧不累积 BPTT 训练缓存
const promptIds = this.tokenizer.encode(prompt);
for (const id of promptIds) this.advance(id);
this.ctxWin = promptIds.slice(-this.model.cfg.ctx);
const routes: CharRoute[] = [];
let text = prompt;
for (let i = 0; i < maxLen; i++) {
const ctx = this.ctxWin.slice();
const ctxFp = Engine.ctxFingerprint(ctx, this.model.cfg.vocabSize); // [P0-B] 上下文指纹 (稳定哈希, 非 seed)
const task: Task = { id: `${i}:${ctxFp}`, type: "text", input: ctx, complexity: Math.min(1, 0.4 + (i / Math.max(maxLen, 1)) * 0.4), budget };
const recalled = this.recallAt(ctx);
// [P0-B 修复 2026-10-08 第一性审核] mem: 键绑定「上下文指纹」而非任务序号 —
// 旧键 `mem:m${i}` 在 generate() 入口从 m0 重编号, 黑板不清空 → 服务器常驻进程
// 必然用上一请求残留键制造假命中 (审核实测 m1/m2 本轮零召回却报 L0 命中)。
// 新键 `mem:${ctxFp}` 仅当「相同上下文」时命中, 不同上下文的残留键天然失效。
if (recalled) this.station.blackboard.write(`mem:${ctxFp}`, recalled.char);
// L0 一致性探针 (宪章第7/8条): 每步记任务数; L0 命中时记 L0 直出字 vs 同上下文深链 top-1
this.l0Probe.noteTask();
if (recalled) {
const deep = this.deepTop1At();
this.l0Probe.record({
ctxPreview: this.tokenizer.decode(ctx).slice(-4),
chL0: recalled.char,
chDeep: deep.char,
deepP: deep.p,
source: recalled.source,
});
}
// L0 Titans 记忆直出 — 在 executor 之前直接检查(避免黑屏时序问题)
let ch = "";
let memoryHit = false;
// const titansHit = this.getTitansOutput(ctx);
// if (titansHit !== null && !this.isDegenerate(titansHit)) {
// ch = titansHit;
// memoryHit = true;
// const mid = this.tokenizer.idOf(ch);
// if (mid !== undefined) {
// this.advance(mid);
// this.ctxWin.push(mid);
// this.ctxWin = this.ctxWin.slice(-this.model.cfg.ctx);
// this.titansMem.write(Float64Array.from(this.model.embedAvg(ctx)), ch);
// }
// routes.push({ char: ch, depth: 0, confidence: 1.0, units: ["titans"], memoryHit: true, memorySource: "repeat" as any });
// this.propose(ch);
// text += ch;
// continue;
// }
const res: TaskResult = await this.station.executor.execute(task);
memoryHit = res.memoryHit ?? false;
// [P0-A 修复 2026-10-08 第一性审核] L0 记忆直出真接: res.output (executor 返回的
// 记忆字符) 必须直接作为输出字, 不得丢弃后走 peekLogits(4) 兜底深链 (旧实现 res.output
// 全仓从未被读取, L0 命中步仍跑 4 次完整前向, 审计记 depth=0/units=[]/conf=1 而真实
// 输出是深链结果 — 宪章第 1 条「假输出」)。
// 不变量: L0 命中步的最终输出字 === res.output (记忆字), 由 l0Probe 记录供审计。
let l0Advanced = false; // [P0-A 补漏] L0 命中步在本分支完成推进 → 下方「推进真实状态」块跳过, 防同字双推进
if (res.memoryHit && res.depth === 0 && res.output !== undefined && res.output !== "") {
ch = String(res.output);
// 推进状态 (零算力: 仅 advance + titansMem 写, 不跑深链前向)。
// [P0-A 补漏] L0 命中步在此完成推进, 跳过下方「推进真实状态」块, 避免同字双推进。
const hitId = this.tokenizer.idOf(ch);
if (hitId !== undefined) {
this.advance(hitId);
this.ctxWin.push(hitId);
this.ctxWin = this.ctxWin.slice(-this.model.cfg.ctx);
this.ctxWinWindow.add({ type: "char", char: ch, id: hitId }, 1);
this.titansMem.write(Float64Array.from(this.model.embedAvg(ctx)), ch);
l0Advanced = true;
}
}
// 深度路径: 从 blackboard 取 logits 并采样 (仅 L0 未命中时执行)
if (ch === "" && res.depth > 0) {
const zLogits = this.station.blackboard.read(`z:logits:${task.id}`) as Float64Array | undefined;
if (zLogits) {
// SNN 脉冲支路融合: 若 SNN 单元已运行, 将 spike logits 加法融合进骨干 logits
const spikeLogits = this.station.blackboard.read(`z:spike:${task.id}`) as Float64Array | undefined;
if (spikeLogits && spikeLogits.length === zLogits.length) {
for (let k = 0; k < zLogits.length; k++) zLogits[k] += spikeLogits[k];
}
const probs = this.model.logitsToProbs(zLogits);
ch = this.tokenizer.charOf(sampleToken(this.penalized(probs), this.options.topK, this.options.temperature, this.habit));
}
}
// [P0-A] 兜底仅用于 L0 未命中且深链也未产出的情况 (保证有输出, 不伪造零算力)
if (ch === "") {
const fallbackLogits = this.peekLogits(4);
const probs = this.model.logitsToProbs(fallbackLogits);
ch = this.tokenizer.charOf(sampleToken(this.penalized(probs), this.options.topK, this.options.temperature, this.habit));
}
// 路由深度 → 真实状态推进深度: 2026-10-08 第一性审核修复 — 旧实现主状态恒按
// depth 4 全深推进, MoD 预算裁剪只作用于探针副本, 算力与路由结果脱钩。
// 现按本步路由实际执行到的深度推进骨干 (res.depth=0 无单元执行 → 4 全深兜底)。
const advDepth = (res.depth >= 1 && res.depth <= 4 ? res.depth : 4) as 1 | 2 | 3 | 4;
// ===== 语义反思层 =====
// 在字符进入真实状态之前,进行反思检查
const reflection = reflect(ch, text, {
minSentenceLen: 8,
maxSentenceLen: 30,
maxRepeatCount: 2,
bigramThreshold: 0.001,
punctWeight: 5.0,
enabled: true,
}, this.bigramStats ?? undefined);
// 应用反思结果
if (reflection.action === 'force_end') {
ch = '。';
console.log(`[反思] 强制结束句子: ${reflection.reason}`);
} else if (reflection.action === 'reject') {
ch = reflection.char;
console.log(`[反思] 替换字符: ${reflection.reason}`);
}
// ===== 反思层结束 =====
// 推进真实状态
const id = this.tokenizer.idOf(ch);
if (id !== undefined && !l0Advanced) {
// [P0-D 补漏 2026-10-08] 旧实现恒 `advance(id)` 全深推进 — advDepth 已按本步
// 路由深度算出却从未使用 (MoD 算力与路由脱钩的残留), 现真接。
this.advance(id, advDepth);
this.ctxWin.push(id);
this.ctxWin = this.ctxWin.slice(-this.model.cfg.ctx);
// 更新上下文窗口
this.ctxWinWindow.add({ type: 'char', char: ch, id }, 1);
// 在线写入 Titans 神经记忆
this.titansMem.write(Float64Array.from(this.model.embedAvg(ctx)), ch);
}
routes.push({
char: ch,
depth: res.depth,
confidence: res.confidence,
units: res.units,
memoryHit: memoryHit,
memorySource: recalled?.source,
});
this.propose(ch);
text += ch;
if (!memoryHit && id !== undefined) this.memory.set(ctx.join(","), ch);
if (this.memory.size > 600) {
const first = this.memory.keys().next().value;
if (first !== undefined) this.memory.delete(first);
}
}
this.remember(text.substring(prompt.length));
if (this.multiNeuro?.spike) this.multiNeuro.setSpikeInferenceMode(false); // 生成结束, 恢复训练模式 (训练侧调用可累积 BPTT 缓存)
// [P1-L 修复 2026-10-09 第一性审核] 生成结束, 将探针和经济性数据写入 JSON (每次生成一轮)
if (this.station && this.station.probe) {
this.station.probe.writeToFile && this.station.probe.writeToFile();
}
if (this.station && this.station.economics) {
this.station.economics.report(); // 内部包含写入逻辑
}
return { text, routes, budget };
}
get registry() {
return this.station.registry;
}
get blackboard() {
return this.station.blackboard;
}
get audit() {
return this.station.audit;
}
get probe() {
return this.station.probe;
}
get economics() {
return this.station.economics;
}
get titans() {
return this.titansMem;
}
get cnVectorStore() { return this.cnStore; }
get mathVectorStore() { return this.mathStore; }
/** [P0-1 2026-10-13] SessionEngine 读取共享向量库/大语料统计, 补公开 getter (宪章第10条: 共享只读) */
get vectorStore() { return this.store; }
get bigramStatsGetter() { return this.bigramStats; }
getWorldModel() { return this.worldModel; }
get reasoner() { return this._reasoner; }
/** 获取上下文窗口 */
getContextWindow() {
return this.ctxWinWindow;
}
/** 返回模型总参数数 (CharMultiNeuro 含 MoE/专家头,CharGRU 仅 backbone) */
totalParamCount(): number {
return this.multiNeuro ? this.multiNeuro.paramCount() : this.model.paramCount();
}
}
|