@aiwayds/dsh-topics-memory 0.19.0 → 0.20.0

package/README.md

@@ -94,6 +94,10 @@ Adopt / record / fallback bands (calibrated offline on a 383-case gold corpus; t
 | Consolidation prefilter | The cluster goes to the LLM in full, exactly as today | none |
 | Fast-lane shadow | Records the outcome only; zero behavior impact | none |
 
+**Setting up laya**: the bundled skill `dsh-topics-memory-laya` walks through install (venv, mirror acceleration for CN networks, checkpoint download), starting laya-serve, wiring `jevLayaFallback`, and verifying via decisions.jsonl. Ask the agent about laya, or read `skills/dsh-topics-memory-laya/SKILL.md` directly.
+
+**Local laya pace-maker (experimental, off by default).** With `jevLayaFallback: true` and a local [laya](https://github.com/NandhaKishorM/laya) serve running, every jev call fires a parallel local laya request: the primary answer drives the decision, and the laya answer is logged alongside it — a permanently running laya-vs-jev comparison on identical questions. laya is TELEMETRY ONLY: on real queries its within-batch ranking agreed with the primary 0/14 and its negative scores sit at 0.63-0.76 (absolute scores unusable), so a primary failure falls open to the legacy path exactly as without the pace-maker — the comparison rows are the value, not a degraded takeover. The consolidation prefilter deliberately does not ride the pace-maker. laya not running costs a refused connection in milliseconds.
+
 Every call and verdict lands in `~/.dsh/topics/meta/decisions.jsonl` — local-only and redacted: slugs, pair hashes, question types, probabilities, latency and token counts; never conversation text or conclusion bodies. It has no config key (it stops together with `jevEnabled: false`); `/topics status` shows a 30-day summary line.
 
 
@@ -203,6 +207,8 @@ First-time setup belongs to `/topics onboard`; day-to-day tuning is `/topics set
 | `jevModel` | empty (per backend: `jev-1.13-free` / `jev-1.13.0` / `typesafe/jev-1.13`) | Version-pinned decision model; upgrading is an explicit action |
 | `jevTimeoutMs` | `3000` | Per-request decision timeout; single attempt, no retry |
 | `jevSecretFile` | none | External secret list for the outbound secret gate; re-read when hot-changed |
+| `jevLayaFallback` | `false` | Local laya pace-maker: fire a parallel laya call on every jev request; its answer takes over (degraded, relative-ranking only) when the primary fails |
+| `jevLayaUrl` | `http://127.0.0.1:8000/v1/systemone` | laya-serve endpoint for the pace-maker |
 
 ## Acknowledgements
 

package/README.zh.md

@@ -193,8 +193,16 @@ dsh plugin --profile <name> remove @aiwayds/dsh-topics-memory
 | `jevModel` | 空（按 backend：`jev-1.13-free` / `jev-1.13.0` / `typesafe/jev-1.13`） | 版本钉定的决策模型；升级是显式动作 |
 | `jevTimeoutMs` | `3000` | 单次决策请求超时；单发不重试 |
 | `jevSecretFile` | 无 | 出网 secret gate 的外部清单路径；热更换路径即重读 |
+| `jevLayaFallback` | `false` | 本地 laya 陪跑：每次 jev 调用并发一路 laya 请求；laya 为纯遥测对照（不接管决策，真实负载排序一致率 0/14） |
+| `jevLayaUrl` | `http://127.0.0.1:8000/v1/systemone` | 陪跑用的 laya-serve 端点 |
 | `jevDebug` | 关 | 诊断日志开关（走宿主 logger，默认静默） |
 
+### laya 本地陪跑（实验，默认关）
+
+`jevLayaFallback: true` 且本地跑着 [laya](https://github.com/NandhaKishorM/laya) serve 时，每次 jev 调用并发一路本地 laya 请求：主路答案驱动决策，laya 答案并排落账——一组永久运行的 laya vs jev 同题对照。laya 是**纯遥测**（真实负载实测批内排序与主路一致率 0/14、绝对分不可用），主路失败照旧回退旧路径；laya 没起 = 毫秒级连接拒绝。
+
+**laya 安装配置**：随包 skill `dsh-topics-memory-laya` 完整覆盖安装（venv / 国内镜像加速 / checkpoint 下载）、启动 laya-serve、配置 `jevLayaFallback`、用 decisions.jsonl 验证生效。对 agent 说 laya 即可触发，或直接读 `skills/dsh-topics-memory-laya/SKILL.md`。
+
 ## Acknowledgements
 
 本项目的形态直接受以下项目的启发与支撑：

package/lib/config.d.ts

@@ -89,6 +89,14 @@ export declare const TopicsConfig: z<Schemastery.ObjectS<NoInfer<{
     /** External secret list for the outbound secret gate (JEV_SECRET_FILE
      *  semantics, design §3.3); swapping the path hot reloads the list. */
     jevSecretFile: z<string, string, "volatile-defined">;
+    /** Local laya pace-maker (design §3.4, 09-27): fire a parallel laya call on
+     *  every jev request; its verdict is logged for comparison and takes over
+     *  ONLY when the primary backend fails (degraded). Default off — zero
+     *  behavior change, and laya-serve not running costs a refused connection. */
+    jevLayaFallback: z<boolean, boolean, "volatile-defined">;
+    /** laya-serve systemone endpoint for the pace-maker (default: the local
+     *  laya-serve bind from the laya repo's serve mode). */
+    jevLayaUrl: z<string, string, "volatile-defined">;
 }>>, Schemastery.ObjectT<NoInfer<{
     /** GitHub repo `owner/name`; empty = local-only mode (ADR 0008). */
     repo: z<string, string, "volatile-defined">;
@@ -173,6 +181,14 @@ export declare const TopicsConfig: z<Schemastery.ObjectS<NoInfer<{
     /** External secret list for the outbound secret gate (JEV_SECRET_FILE
      *  semantics, design §3.3); swapping the path hot reloads the list. */
     jevSecretFile: z<string, string, "volatile-defined">;
+    /** Local laya pace-maker (design §3.4, 09-27): fire a parallel laya call on
+     *  every jev request; its verdict is logged for comparison and takes over
+     *  ONLY when the primary backend fails (degraded). Default off — zero
+     *  behavior change, and laya-serve not running costs a refused connection. */
+    jevLayaFallback: z<boolean, boolean, "volatile-defined">;
+    /** laya-serve systemone endpoint for the pace-maker (default: the local
+     *  laya-serve bind from the laya repo's serve mode). */
+    jevLayaUrl: z<string, string, "volatile-defined">;
 }>>, "plain">;
 export type TopicsConfigValue = {
     repo: string;
@@ -210,8 +226,10 @@ export type TopicsConfigValue = {
     jevModel?: string;
     jevTimeoutMs?: number;
     jevSecretFile?: string;
+    jevLayaFallback?: boolean;
+    jevLayaUrl?: string;
 };
-export declare const CONFIG_KEYS: readonly ["repo", "autoInject", "injectDedup", "suppressEcho", "topK", "perTopicBudget", "totalBudget", "matchThreshold", "tagBoost", "injectMode", "qualityLane", "graphDepth", "recencyWindowDays", "autoObserve", "includeSubagents", "observationMaxChars", "distillEveryTurns", "distillOnSessionEnd", "distillProvider", "distillModel", "distillBatchSize", "distillMaxModelCalls", "consolidateCadence", "deprecatedTtlDays", "usageBoost", "pushDebounceSeconds", "jevEnabled", "jevBackend", "jevModel", "jevTimeoutMs", "jevSecretFile"];
+export declare const CONFIG_KEYS: readonly ["repo", "autoInject", "injectDedup", "suppressEcho", "topK", "perTopicBudget", "totalBudget", "matchThreshold", "tagBoost", "injectMode", "qualityLane", "graphDepth", "recencyWindowDays", "autoObserve", "includeSubagents", "observationMaxChars", "distillEveryTurns", "distillOnSessionEnd", "distillProvider", "distillModel", "distillBatchSize", "distillMaxModelCalls", "consolidateCadence", "deprecatedTtlDays", "usageBoost", "pushDebounceSeconds", "jevEnabled", "jevBackend", "jevModel", "jevTimeoutMs", "jevSecretFile", "jevLayaFallback", "jevLayaUrl"];
 export type ConfigKey = (typeof CONFIG_KEYS)[number];
 /** CamelCase → dash-display (topK → top-k) for command output. */
 export declare function displayKey(key: string): string;

package/lib/config.js

@@ -89,6 +89,14 @@ export const TopicsConfig = z.object({
     /** External secret list for the outbound secret gate (JEV_SECRET_FILE
      *  semantics, design §3.3); swapping the path hot reloads the list. */
     jevSecretFile: z.string().default('').volatile(),
+    /** Local laya pace-maker (design §3.4, 09-27): fire a parallel laya call on
+     *  every jev request; its verdict is logged for comparison and takes over
+     *  ONLY when the primary backend fails (degraded). Default off — zero
+     *  behavior change, and laya-serve not running costs a refused connection. */
+    jevLayaFallback: z.boolean().default(false).volatile(),
+    /** laya-serve systemone endpoint for the pace-maker (default: the local
+     *  laya-serve bind from the laya repo's serve mode). */
+    jevLayaUrl: z.string().default('http://127.0.0.1:8000/v1/systemone').volatile(),
 });
 export const CONFIG_KEYS = [
     'repo',
@@ -122,6 +130,8 @@ export const CONFIG_KEYS = [
     'jevModel',
     'jevTimeoutMs',
     'jevSecretFile',
+    'jevLayaFallback',
+    'jevLayaUrl',
 ];
 /** CamelCase → dash-display (topK → top-k) for command output. */
 export function displayKey(key) {
@@ -198,6 +208,14 @@ export function parseConfigValue(key, raw) {
         case 'distillModel':
         case 'jevModel':
         case 'jevSecretFile':
+        case 'jevLayaUrl':
             return raw;
+        case 'jevLayaFallback': {
+            if (raw === 'on' || raw === 'true')
+                return true;
+            if (raw === 'off' || raw === 'false')
+                return false;
+            return { error: 'jev-laya-fallback 取值 on|off' };
+        }
     }
 }

package/lib/index.d.ts

@@ -118,6 +118,8 @@ export declare const Config: import("@deepseek-ai/schemastery").default<Schemast
     jevModel: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
     jevTimeoutMs: import("@deepseek-ai/schemastery").default<number, number, "volatile-defined">;
     jevSecretFile: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
+    jevLayaFallback: import("@deepseek-ai/schemastery").default<boolean, boolean, "volatile-defined">;
+    jevLayaUrl: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
 }>>, Schemastery.ObjectT<NoInfer<{
     repo: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
     autoInject: import("@deepseek-ai/schemastery").default<boolean, boolean, "volatile-defined">;
@@ -150,6 +152,8 @@ export declare const Config: import("@deepseek-ai/schemastery").default<Schemast
     jevModel: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
     jevTimeoutMs: import("@deepseek-ai/schemastery").default<number, number, "volatile-defined">;
     jevSecretFile: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
+    jevLayaFallback: import("@deepseek-ai/schemastery").default<boolean, boolean, "volatile-defined">;
+    jevLayaUrl: import("@deepseek-ai/schemastery").default<string, string, "volatile-defined">;
 }>>, "plain">;
 /** Live config reference as a 0.1.7 host passes it (cron / mcp-adapter
  * spelling): `.get()` snapshots the current value — the host swaps the

package/lib/index.js

@@ -174,13 +174,18 @@ function readConfigValue(slot) {
 // --- Bundled skill -----------------------------------------------------------
 /** Provider name under `ctx.skills`; doubles as the skill name. */
 const SKILL_PROVIDER_NAME = 'dsh-topics-memory-config';
-/** Packaged skill body; `../skills/` resolves to the package root from both lib/ and src/. */
+/** Packaged skill bodies; `../skills/` resolves to the package root from both lib/ and src/. */
 const SKILL_BODY_URL = new URL('../skills/dsh-topics-memory-config/SKILL.md', import.meta.url);
-/** Resource base served with the skill so its relative links resolve. */
+const LAYA_SKILL_BODY_URL = new URL('../skills/dsh-topics-memory-laya/SKILL.md', import.meta.url);
+/** Resource bases served with the skills so their relative links resolve. */
 const SKILL_RESOURCE_BASE = {
     kind: 'directory',
     path: fileURLToPath(new URL('../skills/dsh-topics-memory-config/', import.meta.url)),
 };
+const LAYA_SKILL_RESOURCE_BASE = {
+    kind: 'directory',
+    path: fileURLToPath(new URL('../skills/dsh-topics-memory-laya/', import.meta.url)),
+};
 const SKILL_INVOCATION = { modelInvocable: true, userInvocable: true };
 /** Routing description; must stay identical to the SKILL.md frontmatter (asserted in tests). */
 const SKILL_DESCRIPTION = 'dsh 记忆插件（@aiwayds/dsh-topics-memory）使用与配置指南。凡涉及 dsh 记忆/话题库/GitHub 同步/蒸馏/整理，或要配置 topics 时先读本指南：设置页 `dsh-topics-memory` 条目全部键（repo/autoInject/topK/注入预算/蒸馏/整理/观察/图游走等）、/topics 命令族（onboard/status/distill/consolidate/stats/list/show/history/graph/sync/config/set）、首次配置 ask_user_question 向导（local-only 或绑 GitHub 仓、蒸馏模型路由、注入档位、自动观察）、注入形态 pointer/digest。触发词：topics、记忆、topic、蒸馏、distill、整理、consolidate、合并重复、deprecatedTtl、usageBoost、使用加成、autoInject、记忆库、include-subagents。';
@@ -194,18 +199,35 @@ const SKILL_CANDIDATE = {
     rank: BUNDLED_SKILL_RANK,
     locator: SKILL_BODY_URL,
 };
+/** The laya setup skill rides the same provider: installing/configuring the
+ *  local pace-maker is part of operating this plugin's decision layer. */
+const LAYA_SKILL_NAME = 'dsh-topics-memory-laya';
+const LAYA_SKILL_DESCRIPTION = '为 dsh-topics-memory 的 jev 决策层安装并配置本地 laya 陪跑（laya-serve）。当用户要开启 jevLayaFallback、安装 laya、启动 laya-serve、配置 laya 端点、排查陪跑不生效或 decisions.jsonl 没有 laya 行时读本指南：venv 创建、国内三镜像加速（清华 PyPI / hf-mirror / 禁 Xet）、checkpoint 下载、laya-serve 启动、jevLayaFallback/jevLayaUrl 两键、真机验证（decisions.jsonl 出现 backend=laya 行）。触发词：laya、laya-serve、陪跑、本地决策、jevLaya、离线、8000 端口。';
+const LAYA_SKILL_CANDIDATE = {
+    name: LAYA_SKILL_NAME,
+    description: LAYA_SKILL_DESCRIPTION,
+    invocation: SKILL_INVOCATION,
+    provider: SKILL_PROVIDER_NAME,
+    source: 'bundled',
+    resourceBase: LAYA_SKILL_RESOURCE_BASE,
+    rank: BUNDLED_SKILL_RANK,
+    locator: LAYA_SKILL_BODY_URL,
+};
 const skillProvider = {
     name: SKILL_PROVIDER_NAME,
-    list: () => Promise.resolve([SKILL_CANDIDATE]),
-    async get(_candidate) {
+    list: () => Promise.resolve([SKILL_CANDIDATE, LAYA_SKILL_CANDIDATE]),
+    async get(candidate) {
+        const isLaya = candidate.name === LAYA_SKILL_NAME;
+        const bodyUrl = isLaya ? LAYA_SKILL_BODY_URL : SKILL_BODY_URL;
+        const resourceBase = isLaya ? LAYA_SKILL_RESOURCE_BASE : SKILL_RESOURCE_BASE;
         return {
-            name: SKILL_CANDIDATE.name,
-            description: SKILL_CANDIDATE.description,
-            invocation: SKILL_CANDIDATE.invocation,
-            provider: SKILL_CANDIDATE.provider,
-            source: SKILL_CANDIDATE.source,
-            resourceBase: SKILL_RESOURCE_BASE,
-            content: stripFrontmatter(await readFile(SKILL_BODY_URL, 'utf8')),
+            name: candidate.name,
+            description: candidate.description,
+            invocation: candidate.invocation,
+            provider: SKILL_PROVIDER_NAME,
+            source: candidate.source,
+            resourceBase,
+            content: stripFrontmatter(await readFile(bodyUrl, 'utf8')),
         };
     },
 };

package/lib/jev/backends.d.ts

@@ -9,6 +9,9 @@
  * @module jev/backends
  */
 export type JevBackendName = 'zen' | 'native' | 'openrouter';
+/** The local laya pace-maker writes its own call-layer rows; not a config
+ *  backend (no entry in BACKENDS) — it rides jevAsk's overrides hook. */
+export type JevPaceMakerName = 'laya';
 export type JevFamily = 'systemone' | 'decisions';
 export interface JevBackend {
     readonly name: JevBackendName;

package/lib/jev/client.d.ts

@@ -41,6 +41,13 @@ export interface JevAskArgs {
     fallback?: boolean;
     /** Host logger (optional, warn-only, best-effort). */
     logger?: GateLogger;
+    /** Dual-run pace-maker hook (jev/dual.ts): target this endpoint/model with
+     *  no auth header and tag the call-layer row backend='laya'. Never set by
+     *  user config — internal to the dual wrapper. */
+    overrides?: {
+        endpoint: string;
+        model: string;
+    };
 }
 export type JevAnswer = Record<string, unknown>;
 export type JevAskOk = {

package/lib/jev/client.js

@@ -119,9 +119,22 @@ export function normalizeQuestions(questions, family) {
  *  resolves to a structured failure (never throws) so the calling lane can
  *  fail open. Writes the decisions.jsonl call-layer row on EVERY path. */
 export async function jevAsk(args) {
-    const backend = resolveBackend(args.config.jevBackend);
+    // Pace-maker override: a synthetic laya backend (no key — the auth header
+    // carries a placeholder the local server ignores; the secret gate and the
+    // call-layer row still run, tagged backend='laya').
+    const backend = args.overrides
+        ? {
+            name: 'laya',
+            family: 'systemone',
+            endpoint: args.overrides.endpoint,
+            defaultModel: args.overrides.model,
+            keyEnv: '',
+            keychainService: '',
+            keychainByDefault: false,
+        }
+        : resolveBackend(args.config.jevBackend);
     // '' is the sentinel; undefined (bare harnesses) resolves to the default too
-    const model = args.config.jevModel ? args.config.jevModel : backend.defaultModel;
+    const model = args.overrides ? args.overrides.model : args.config.jevModel ? args.config.jevModel : backend.defaultModel;
     const fallback = args.fallback === true; // absent => false: the column is call-time-invariant (fail-open = outcome !== ok)
     const state = typeof args.state === 'string' ? args.state : '';
     const started = Date.now();
@@ -157,8 +170,9 @@ export async function jevAsk(args) {
     if (hit !== null) {
         return refuse('secret_gate_blocked', `outbound payload matched secret pattern '${hit.pattern}' at offset ${hit.offset}; NOTHING was sent. Remove the secret from the request body (model/state/questions).`);
     }
-    // 3. API key — absent key fails before any network activity.
-    const key = await resolveApiKey(backend);
+    // 3. API key — absent key fails before any network activity. The laya
+    //    pace-maker skips resolution entirely (local, no auth).
+    const key = args.overrides ? 'laya-local' : await resolveApiKey(backend);
     if (key === '') {
         const hint = backend.keychainByDefault
             ? ` (set ${backend.keyEnv}, or store the key in the macOS keychain service '${backend.keychainService}')`

package/lib/jev/dual.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Dual-run pace-maker (design §3.4, added 09-27): fire the configured
+ * backend AND the local laya-serve endpoint in parallel on every jev call.
+ *
+ *  - the PRIMARY answer drives the decision whenever it is ok;
+ *  - the laya answer is always logged (its own call-layer row tagged
+ *    backend='laya' plus verdict rows) — a free, permanently running
+ *    laya-vs-jev comparison on identical questions;
+ *  - TELEMETRY-ONLY: laya NEVER drives a decision. On real data its
+ *    within-batch top-1 agreed with the primary 0/14 and its negatives sit in
+ *    the same band as its positives (0.63-0.76), so both its relative ranking
+ *    and its absolute scores are unusable (4-way bench, 09-27). A primary
+ *    FAILURE (cold-start spike, timeout, 5xx…) is therefore a failure even
+ *    when laya answered: the result stays ok:false and the caller fails open
+ *    (ADR 0018). `layaResult` still carries the comparison answers so the
+ *    caller can log the laya verdict rows;
+ *  - laya not running = a refused connection in milliseconds — the switch
+ *    can stay on unconditionally.
+ *
+ * The consolidation prefilter deliberately does NOT ride this wrapper: laya's
+ * Chinese noul is the weak spot, and a wrong veto silently skips real
+ * consolidation work. Prefilter keeps its plain fail-open.
+ *
+ * @module jev/dual
+ */
+import type { JevAskArgs, JevAskConfig, JevAskResult } from './client.ts';
+/** Default local laya-serve endpoint (laya repo: `laya[serve]` binds 8000). */
+export declare const DEFAULT_LAYA_URL = "http://127.0.0.1:8000/v1/systemone";
+export interface JevDualConfig extends JevAskConfig {
+    /** Pace-maker switch (default OFF — zero extra requests when false). */
+    jevLayaFallback: boolean;
+    /** Local laya-serve systemone endpoint. */
+    jevLayaUrl: string;
+}
+export interface JevDualArgs extends Omit<JevAskArgs, 'config'> {
+    config: JevDualConfig;
+}
+export type JevDualSource = 'primary' | 'laya';
+export type JevDualResult = JevAskResult & {
+    /** Which backend ANSWERED for the record: 'primary' when the primary
+     *  answered (the only case that may drive), 'laya' when it failed and the
+     *  pace-maker covered for telemetry only. */
+    source: JevDualSource;
+    /** The pace-maker's own result when it ran (null when the switch is off),
+     *  always logged by jevAsk via the overrides hook. */
+    layaResult: JevAskResult | null;
+};
+export declare function jevAskDual(args: JevDualArgs): Promise<JevDualResult>;

package/lib/jev/dual.js

@@ -0,0 +1,47 @@
+/**
+ * Dual-run pace-maker (design §3.4, added 09-27): fire the configured
+ * backend AND the local laya-serve endpoint in parallel on every jev call.
+ *
+ *  - the PRIMARY answer drives the decision whenever it is ok;
+ *  - the laya answer is always logged (its own call-layer row tagged
+ *    backend='laya' plus verdict rows) — a free, permanently running
+ *    laya-vs-jev comparison on identical questions;
+ *  - TELEMETRY-ONLY: laya NEVER drives a decision. On real data its
+ *    within-batch top-1 agreed with the primary 0/14 and its negatives sit in
+ *    the same band as its positives (0.63-0.76), so both its relative ranking
+ *    and its absolute scores are unusable (4-way bench, 09-27). A primary
+ *    FAILURE (cold-start spike, timeout, 5xx…) is therefore a failure even
+ *    when laya answered: the result stays ok:false and the caller fails open
+ *    (ADR 0018). `layaResult` still carries the comparison answers so the
+ *    caller can log the laya verdict rows;
+ *  - laya not running = a refused connection in milliseconds — the switch
+ *    can stay on unconditionally.
+ *
+ * The consolidation prefilter deliberately does NOT ride this wrapper: laya's
+ * Chinese noul is the weak spot, and a wrong veto silently skips real
+ * consolidation work. Prefilter keeps its plain fail-open.
+ *
+ * @module jev/dual
+ */
+import { jevAsk } from "./client.js";
+/** Default local laya-serve endpoint (laya repo: `laya[serve]` binds 8000). */
+export const DEFAULT_LAYA_URL = 'http://127.0.0.1:8000/v1/systemone';
+export async function jevAskDual(args) {
+    if (args.config.jevLayaFallback !== true) {
+        const r = await jevAsk(args);
+        return { ...r, source: 'primary', layaResult: null };
+    }
+    const layaArgs = {
+        ...args,
+        config: { ...args.config, jevModel: 'laya-rl-agent' },
+        overrides: { endpoint: args.config.jevLayaUrl || DEFAULT_LAYA_URL, model: 'laya-rl-agent' },
+    };
+    const [primary, laya] = await Promise.all([jevAsk(args), jevAsk(layaArgs)]);
+    if (primary.ok)
+        return { ...primary, source: 'primary', layaResult: laya };
+    // REVISED 09-27 (real-data evidence: within-batch top-1 agreement 0/14):
+    // a primary failure is a FAILURE even when the pace-maker answered — laya
+    // is telemetry-only. The caller fails open (ADR 0018); layaResult carries
+    // the comparison answers so the caller can still log laya verdict rows.
+    return { ok: false, outcome: primary.outcome, message: primary.message, latencyMs: primary.latencyMs, source: 'laya', layaResult: laya };
+}

package/lib/jev/log.d.ts

@@ -52,6 +52,14 @@ export interface JevVerdictRecord {
     probability: number;
     band: JevBand;
     agree: 'hit' | 'nearFloor' | 'gate-blocked' | 'wouldBlock' | 'n/a';
+    /** Which backend produced this verdict: 'primary' (default — the configured
+     *  backend) or 'laya' (the local pace-maker). Design §3.4. */
+    backend?: string;
+    /** True when this row's source DROVE the decision because the primary
+     *  backend was unavailable (degraded mode). Comparison rows from the
+     *  pace-maker are degraded=false; laya rows are excluded from ECE either
+     *  way (uncalibrated probabilities). */
+    degraded?: boolean;
 }
 export declare function decisionsFile(root: string): string;
 /** digest(state, questionId, instructions): `h1:` + first 16 hex of

package/lib/jev/log.js

@@ -115,5 +115,7 @@ export async function logVerdicts(recs) {
         probability: rec.probability,
         band: rec.band,
         agree: rec.agree,
+        backend: rec.backend ?? 'primary',
+        degraded: rec.degraded === true,
     })));
 }

package/lib/quality.js

@@ -22,7 +22,7 @@
  * @module quality
  */
 import { searchTopics } from "./retrieval.js";
-import { jevAsk } from "./jev/client.js";
+import { jevAskDual } from "./jev/dual.js";
 import { band as jevBand } from "./jev/thresholds.js";
 import { digestFor, logVerdicts } from "./jev/log.js";
 export const PENDING_TTL_MS = 10 * 60_000;
@@ -88,6 +88,33 @@ function noulProbability(answer) {
     const p = answer.noul;
     return typeof p === 'number' && Number.isFinite(p) && p >= 0 && p <= 1 ? p : undefined;
 }
+/** Pace-maker comparison rows for the slow rerank (design §3.4, REVISED
+ *  09-27): laya is TELEMETRY ONLY, so its rows are recorded band='record'
+ *  (its absolute scores never claim adopt) and degraded=false whether or not
+ *  the primary answered — landing them costs no behavior. An unparseable
+ *  answer gets no fabricated row. */
+function layaRerankRows(at, state, metas, answers) {
+    const recs = [];
+    for (const m of metas) {
+        const probability = noulProbability(answers?.[m.qid]);
+        if (probability === undefined)
+            continue;
+        recs.push({
+            at,
+            lane: 'slowlane-rerank',
+            questionId: m.qid,
+            qtype: 'noul',
+            ref: `slug:${m.slug}`,
+            digest: digestFor(state, m.qid, m.instructions),
+            probability,
+            band: 'record',
+            agree: 'n/a',
+            backend: 'laya',
+            degraded: false,
+        });
+    }
+    return recs;
+}
 /** The config slice jevAsk needs, tolerating bare harnesses (§4: consumers
  *  treat undefined as the documented default). */
 function jevConfigOf(cfg) {
@@ -96,6 +123,8 @@ function jevConfigOf(cfg) {
         jevModel: cfg.jevModel ?? '',
         jevTimeoutMs: typeof cfg.jevTimeoutMs === 'number' && cfg.jevTimeoutMs > 0 ? cfg.jevTimeoutMs : 3000,
         jevSecretFile: cfg.jevSecretFile ?? '',
+        jevLayaFallback: cfg.jevLayaFallback === true,
+        jevLayaUrl: cfg.jevLayaUrl ?? 'http://127.0.0.1:8000/v1/systemone',
     };
 }
 /** Race `p` against a timeout; `onTimeout` fires so callers can abort upstream work. */
@@ -242,37 +271,48 @@ export class SlowLane {
         }
         if (metas.length === 0)
             return;
-        const result = await jevAsk({
+        const dual = await jevAskDual({
             state,
             questions,
             config: jevConfigOf(this.service.cfg),
             lane: 'fastgate-shadow',
             fallback: false, // pure shadow — no behavior to fall back
         });
-        if (!result.ok)
-            return;
         const at = new Date().toISOString();
+        // Verdict rows for BOTH sources when they answered: the primary row set
+        // and the laya pace-maker comparison set (backend='laya', excluded from
+        // ECE, §6.2). Pure shadow — nothing gates here, so the laya rows land
+        // whenever laya answered, primary success or not.
+        const sources = [];
+        if (dual.ok)
+            sources.push({ answers: dual.answers, backend: 'primary', degraded: false });
+        if (dual.layaResult?.ok)
+            sources.push({ answers: dual.layaResult.answers, backend: 'laya', degraded: false });
         const recs = [];
-        for (const m of metas) {
-            const probability = noulProbability(result.answers[m.qid]);
-            if (probability === undefined)
-                continue; // unparseable answer: no fabricated row
-            const b = jevBand(probability, 'rerank');
-            // agree reconciliation (§6.2): the ilog disposition, with wouldBlock for
-            // the reverse disagreement (lexical let it through, jev lands in the veto
-            // band — the other half of the calibration goldmine).
-            const agree = m.disposition === 'hit' && b === 'fallback' ? 'wouldBlock' : m.disposition;
-            recs.push({
-                at,
-                lane: 'fastgate-shadow',
-                questionId: m.qid,
-                qtype: 'noul',
-                ref: `slug:${m.slug}`,
-                digest: digestFor(state, m.qid, m.instructions),
-                probability,
-                band: b,
-                agree,
-            });
+        for (const src of sources) {
+            for (const m of metas) {
+                const probability = noulProbability(src.answers?.[m.qid]);
+                if (probability === undefined)
+                    continue; // unparseable answer: no fabricated row
+                const b = jevBand(probability, 'rerank');
+                // agree reconciliation (§6.2): the ilog disposition, with wouldBlock for
+                // the reverse disagreement (lexical let it through, jev lands in the veto
+                // band — the other half of the calibration goldmine).
+                const agree = m.disposition === 'hit' && b === 'fallback' ? 'wouldBlock' : m.disposition;
+                recs.push({
+                    at,
+                    lane: 'fastgate-shadow',
+                    questionId: m.qid,
+                    qtype: 'noul',
+                    ref: `slug:${m.slug}`,
+                    digest: digestFor(state, m.qid, m.instructions),
+                    probability,
+                    band: b,
+                    agree,
+                    backend: src.backend,
+                    degraded: src.degraded,
+                });
+            }
         }
         await logVerdicts(recs);
     }
@@ -477,7 +517,7 @@ export class SlowLane {
             questions[qid] = { type: 'noul', instructions, criteria: candidateCriteria(c) };
             metas.push({ qid, slug: c.slug, instructions });
         }
-        const result = await jevAsk({
+        const dual = await jevAskDual({
             state,
             questions,
             config: jevConfigOf(this.service.cfg),
@@ -488,17 +528,28 @@ export class SlowLane {
             // stat. Fail-open occurrence = outcome !== 'ok' (this lane's fallback
             // to the legacy LLM rerank is deterministic per ADR 0018).
         });
-        if (!result.ok)
+        // REVISED 09-27 (real-data evidence): the laya pace-maker is TELEMETRY
+        // ONLY. Its within-batch top-1 agreed with the primary 0/14 on real
+        // queries (and its negatives sit at 0.63-0.76 — absolute scores unusable),
+        // so the earlier degraded-ranking takeover is REVERTED: on primary
+        // failure the lane falls open to the legacy LLM rerank exactly as
+        // ADR 0018 says. The pace-maker still answers, so its comparison rows
+        // land FIRST here (pure telemetry), then the legacy rerank takes over.
+        if (!dual.ok) {
+            if (dual.layaResult?.ok) {
+                await logVerdicts(layaRerankRows(new Date().toISOString(), state, metas, dual.layaResult.answers));
+            }
             return undefined;
+        }
         const scored = [];
         for (const m of metas) {
-            const probability = noulProbability(result.answers[m.qid]);
+            const probability = noulProbability(dual.answers[m.qid]);
             if (probability === undefined)
                 return undefined; // 坏答案 → fail-open (§1)
             scored.push({ ...m, probability, band: jevBand(probability, 'rerank') });
         }
         const at = new Date().toISOString();
-        await logVerdicts(scored.map((m) => ({
+        const recs = scored.map((m) => ({
             at,
             lane: 'slowlane-rerank',
             questionId: m.qid,
@@ -508,7 +559,16 @@ export class SlowLane {
             probability: m.probability,
             band: m.band,
             agree: 'n/a',
-        })));
+            backend: dual.source,
+            degraded: false,
+        }));
+        // Pace-maker comparison rows (design §3.4): when the primary drove and
+        // laya also answered, log laya's verdicts alongside — the permanently
+        // running laya-vs-jev comparison. band='record': laya's absolute scores
+        // never claim adopt.
+        if (dual.layaResult?.ok)
+            recs.push(...layaRerankRows(at, state, metas, dual.layaResult.answers));
+        await logVerdicts(recs);
         return scored
             .filter((m) => m.band === 'adopt')
             .sort((a, b) => b.probability - a.probability)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@aiwayds/dsh-topics-memory",
-  "version": "0.19.0",
+  "version": "0.20.0",
   "description": "dsh-plugin: quality-first topic memory for dsh — a topic keeps the starting question, conclusion, impact and dependencies, not the process; local-first git-traced OKF bundle, budgeted LLM-free injection",
   "type": "module",
   "main": "lib/index.js",
@@ -42,7 +42,7 @@
     "build": "tsc -p tsconfig.json",
     "check": "tsc --noEmit -p tsconfig.json",
     "pretest": "tsc -p tsconfig.json",
-    "test": "node --test test/yaml.test.mjs test/okf.test.mjs test/retrieval.test.mjs test/digest.test.mjs test/ilog.test.mjs test/store.test.mjs test/paths.test.mjs test/git.test.mjs test/sync.test.mjs test/tools.test.mjs test/dedup.test.mjs test/commands.test.mjs test/onboard.test.mjs test/delegation.test.mjs test/observer.test.mjs test/distill.test.mjs test/consolidate.test.mjs test/viz.test.mjs test/corpus.test.mjs test/llm-scope.test.mjs test/quality.test.mjs test/replay-gate.test.mjs test/depends-repair.test.mjs test/skill.test.mjs test/fast-lane.test.mjs test/legacy-import.test.mjs test/jev-config.test.mjs test/jev-gate.test.mjs test/jev-client.test.mjs test/jev-log.test.mjs test/jev-fastgate.test.mjs",
+    "test": "node --test test/yaml.test.mjs test/okf.test.mjs test/retrieval.test.mjs test/digest.test.mjs test/ilog.test.mjs test/store.test.mjs test/paths.test.mjs test/git.test.mjs test/sync.test.mjs test/tools.test.mjs test/dedup.test.mjs test/commands.test.mjs test/onboard.test.mjs test/delegation.test.mjs test/observer.test.mjs test/distill.test.mjs test/consolidate.test.mjs test/viz.test.mjs test/corpus.test.mjs test/llm-scope.test.mjs test/quality.test.mjs test/replay-gate.test.mjs test/depends-repair.test.mjs test/skill.test.mjs test/fast-lane.test.mjs test/legacy-import.test.mjs test/jev-config.test.mjs test/jev-gate.test.mjs test/jev-client.test.mjs test/jev-log.test.mjs test/jev-dual.test.mjs test/jev-fastgate.test.mjs",
     "smoke": "node scripts/smoke-boot.mjs"
   },
   "dependencies": {},

package/skills/dsh-topics-memory-laya/SKILL.md

@@ -0,0 +1,84 @@
+---
+name: dsh-topics-memory-laya
+description: "为 dsh-topics-memory 的 jev 决策层安装并配置本地 laya 陪跑（laya-serve）。当用户要开启 jevLayaFallback、安装 laya、启动 laya-serve、配置 laya 端点、排查 laya 连接失败/陪跑不生效/decisions.jsonl 没有 laya 行时读本指南。覆盖：venv 创建、国内三镜像加速（清华 PyPI / hf-mirror / 禁 Xet）、checkpoint 下载、laya-serve 启动与自启、jevLayaFallback/jevLayaUrl 两键配置、真机验证（decisions.jsonl 出现 backend=laya 行）。触发词：laya、laya-serve、陪跑、本地决策、jevLaya、离线、8000 端口。"
+---
+
+# laya 本地陪跑安装与配置
+
+> laya 是开源的 System One 决策引擎（本地推理，jev 的同范式替身）。topics-memory 的
+> `jevLayaFallback: true` 会把每次 jev 调用并发一路本地 laya 请求作对照遥测。本指南把
+> 安装配置一次讲完；全程不涉及付费 API。
+
+## 前置检查（先判断要不要装）
+
+1. `curl -s --max-time 3 -o /dev/null -w "%{http_code}" -X POST http://127.0.0.1:8000/v1/systemone -H "content-type: application/json" -d '{"state":"w","questions":{"q":{"type":"noul","instructions":"w","criteria":{"true":"t","false":"f"}}}}'`
+   返回 200 = laya-serve 已在跑，**跳到「插件配置」**；连接拒绝 = 继续安装。
+2. `python3 --version` 需要 ≥ 3.10（3.13 以下优先，太新的版本 torch 轮子可能缺）。
+3. 磁盘 ≥ 5GB（torch + 两个 checkpoint）。
+
+## 安装（国内网络三件套必带，否则要么 100KB/s 要么挂死）
+
+```bash
+mkdir -p ~/laya-tryout && cd ~/laya-tryout
+uv venv --python 3.12 .venv            # 没有 uv: brew install uv；或 python3 -m venv .venv
+uv pip install --python .venv/bin/python \
+  --index-url https://pypi.tuna.tsinghua.edu.cn/simple "laya[serve]" hf_transfer
+```
+
+下载 checkpoint（两个共 ~1.5GB）——三个环境变量一个都不能少：
+
+```bash
+export HF_ENDPOINT=https://hf-mirror.com    # HF 直连国内常 <50KB/s
+export HF_HUB_DISABLE_XET=1                 # 新版 huggingface_hub 默认 Xet 协议，镜像不支持，会 0 字节挂死
+export HF_HUB_ENABLE_HF_TRANSFER=1          # 并行分块，断流自动续传
+.venv/bin/python -c "
+from laya import Router
+r = Router()
+r.preload(['english', 'multilingual'])
+print('checkpoints ok')
+"
+```
+
+## 启动 laya-serve
+
+```bash
+LAYA_PRELOAD=1 LAYA_MODELS=english,multilingual ~/laya-tryout/.venv/bin/laya-serve
+# binds 0.0.0.0:8000；进程前台跑，挂 nohup/tmux/launchd 由用户自选
+```
+
+- `LAYA_MODELS` 不设会尝试预载第三个 checkpoint（typed-decisions，再 +1.7GB）——陪跑用不到，别下。
+- 自启（macOS launchd）可选：写 `~/Library/LaunchAgents/com.laya.serve.plist` 指向上面命令；不开自启就要接受「重启后陪跑静默失效」（无 laya = 毫秒级连接拒绝，无害但没对照数据）。
+
+## 插件配置（在 profile patch 的 `dsh-topics-memory` 条目）
+
+```yaml
+config:
+  jevEnabled: true          # 决策层总开关（本身默认关）
+  jevLayaFallback: true     # laya 陪跑开关
+  # jevLayaUrl 默认就是 http://127.0.0.1:8000/v1/systemone，不改端口不用写
+```
+
+前置：`jevEnabled: true` 且已按主配置指南配好 jev 后端 key（zen 免费档默认；key 走 patch `env:` 块或 `JEV_KEYCHAIN`——dsh 会清洗 ambient KEY 变量）。
+
+## 验证（三步）
+
+1. `curl -s -X POST http://127.0.0.1:8000/v1/systemone -H "content-type: application/json" -d '{"state":"w","model":"laya-rl-agent","questions":{"q":{"type":"noul","instructions":"w","criteria":{"true":"t","false":"f"}}}}' | jq .answers` — 返回 noul 概率即 laya 侧通。
+2. 跑一轮真实会话（任意问题），然后：
+   `jq -r 'select(.backend=="laya") | [.ref,.probability] | @tsv' ~/.dsh/topics/meta/decisions.jsonl | head` — 出现 `backend:"laya"` 的判定行 = 陪跑生效。
+3. `~/.dsh/topics/meta/decisions.jsonl` 里同一 `digest` 应有成对的 primary/laya 行——这就是永久运行的 laya-vs-jev 对照数据。
+
+## 语义边界（必须向用户讲清）
+
+- laya 是**纯遥测对照，不参与决策**：真实负载实测其批内排序与 jev 一致率 0/14、负例分 0.63-0.76（绝对分不可用），主路失败照旧回退旧路径。
+- 判定质量上 zen/typesafe 的 jev 系远胜 laya；laya 的价值是免费对照数据流 + 离线可用。
+- laya 换语言（中→英）首次调用会触发 checkpoint 重载（数百 ms），陪跑场景无感。
+
+## 常见坑速查
+
+| 症状 | 原因 | 解法 |
+|---|---|---|
+| 下载 0 字节挂死 | Xet 协议 + 镜像不兼容 | `HF_HUB_DISABLE_XET=1` |
+| checkpoint 下载 <100KB/s | HF 直连 | `HF_ENDPOINT=https://hf-mirror.com` |
+| `z.string().volatile is not a function` 类插件报错 | （与 laya 无关）dsh 版本漂移 | 钉 0.1.7-rc 线，勿裸解析 |
+| decisions.jsonl 无 backend=laya 行 | laya-serve 没起 / jevLayaFallback 没开 / jevEnabled 没开 | 按「验证」步骤逐层查 |
+| 首次调用慢（秒级） | checkpoint 冷加载 | 预热一发即稳态（32-300ms） |

package/src/config.ts

@@ -91,6 +91,14 @@ export const TopicsConfig = z.object({
   /** External secret list for the outbound secret gate (JEV_SECRET_FILE
    *  semantics, design §3.3); swapping the path hot reloads the list. */
   jevSecretFile: z.string().default('').volatile(),
+  /** Local laya pace-maker (design §3.4, 09-27): fire a parallel laya call on
+   *  every jev request; its verdict is logged for comparison and takes over
+   *  ONLY when the primary backend fails (degraded). Default off — zero
+   *  behavior change, and laya-serve not running costs a refused connection. */
+  jevLayaFallback: z.boolean().default(false).volatile(),
+  /** laya-serve systemone endpoint for the pace-maker (default: the local
+   *  laya-serve bind from the laya repo's serve mode). */
+  jevLayaUrl: z.string().default('http://127.0.0.1:8000/v1/systemone').volatile(),
 })
 
 export type TopicsConfigValue = {
@@ -129,6 +137,8 @@ export type TopicsConfigValue = {
   jevModel?: string
   jevTimeoutMs?: number
   jevSecretFile?: string
+  jevLayaFallback?: boolean
+  jevLayaUrl?: string
 }
 
 export const CONFIG_KEYS = [
@@ -163,6 +173,8 @@ export const CONFIG_KEYS = [
   'jevModel',
   'jevTimeoutMs',
   'jevSecretFile',
+  'jevLayaFallback',
+  'jevLayaUrl',
 ] as const
 
 export type ConfigKey = (typeof CONFIG_KEYS)[number]
@@ -235,6 +247,12 @@ export function parseConfigValue(key: ConfigKey, raw: string): boolean | number
     case 'distillModel':
     case 'jevModel':
     case 'jevSecretFile':
+    case 'jevLayaUrl':
       return raw
+    case 'jevLayaFallback': {
+      if (raw === 'on' || raw === 'true') return true
+      if (raw === 'off' || raw === 'false') return false
+      return { error: 'jev-laya-fallback 取值 on|off' }
+    }
   }
 }

package/src/index.ts

@@ -288,14 +288,19 @@ interface UserMessageData {
 /** Provider name under `ctx.skills`; doubles as the skill name. */
 const SKILL_PROVIDER_NAME = 'dsh-topics-memory-config'
 
-/** Packaged skill body; `../skills/` resolves to the package root from both lib/ and src/. */
+/** Packaged skill bodies; `../skills/` resolves to the package root from both lib/ and src/. */
 const SKILL_BODY_URL = new URL('../skills/dsh-topics-memory-config/SKILL.md', import.meta.url)
+const LAYA_SKILL_BODY_URL = new URL('../skills/dsh-topics-memory-laya/SKILL.md', import.meta.url)
 
-/** Resource base served with the skill so its relative links resolve. */
+/** Resource bases served with the skills so their relative links resolve. */
 const SKILL_RESOURCE_BASE = {
   kind: 'directory',
   path: fileURLToPath(new URL('../skills/dsh-topics-memory-config/', import.meta.url)),
 } as const
+const LAYA_SKILL_RESOURCE_BASE = {
+  kind: 'directory',
+  path: fileURLToPath(new URL('../skills/dsh-topics-memory-laya/', import.meta.url)),
+} as const
 
 const SKILL_INVOCATION = { modelInvocable: true, userInvocable: true } as const
 
@@ -313,18 +318,36 @@ const SKILL_CANDIDATE: SkillCandidate = {
   locator: SKILL_BODY_URL,
 }
 
+/** The laya setup skill rides the same provider: installing/configuring the
+ *  local pace-maker is part of operating this plugin's decision layer. */
+const LAYA_SKILL_NAME = 'dsh-topics-memory-laya'
+const LAYA_SKILL_DESCRIPTION = '为 dsh-topics-memory 的 jev 决策层安装并配置本地 laya 陪跑（laya-serve）。当用户要开启 jevLayaFallback、安装 laya、启动 laya-serve、配置 laya 端点、排查陪跑不生效或 decisions.jsonl 没有 laya 行时读本指南：venv 创建、国内三镜像加速（清华 PyPI / hf-mirror / 禁 Xet）、checkpoint 下载、laya-serve 启动、jevLayaFallback/jevLayaUrl 两键、真机验证（decisions.jsonl 出现 backend=laya 行）。触发词：laya、laya-serve、陪跑、本地决策、jevLaya、离线、8000 端口。'
+const LAYA_SKILL_CANDIDATE: SkillCandidate = {
+  name: LAYA_SKILL_NAME,
+  description: LAYA_SKILL_DESCRIPTION,
+  invocation: SKILL_INVOCATION,
+  provider: SKILL_PROVIDER_NAME,
+  source: 'bundled',
+  resourceBase: LAYA_SKILL_RESOURCE_BASE,
+  rank: BUNDLED_SKILL_RANK,
+  locator: LAYA_SKILL_BODY_URL,
+}
+
 const skillProvider: SkillProvider = {
   name: SKILL_PROVIDER_NAME,
-  list: () => Promise.resolve([SKILL_CANDIDATE]),
-  async get(_candidate): Promise<SkillDefinition> {
+  list: () => Promise.resolve([SKILL_CANDIDATE, LAYA_SKILL_CANDIDATE]),
+  async get(candidate): Promise<SkillDefinition> {
+    const isLaya = candidate.name === LAYA_SKILL_NAME
+    const bodyUrl = isLaya ? LAYA_SKILL_BODY_URL : SKILL_BODY_URL
+    const resourceBase = isLaya ? LAYA_SKILL_RESOURCE_BASE : SKILL_RESOURCE_BASE
     return {
-      name: SKILL_CANDIDATE.name,
-      description: SKILL_CANDIDATE.description,
-      invocation: SKILL_CANDIDATE.invocation,
-      provider: SKILL_CANDIDATE.provider,
-      source: SKILL_CANDIDATE.source,
-      resourceBase: SKILL_RESOURCE_BASE,
-      content: stripFrontmatter(await readFile(SKILL_BODY_URL, 'utf8')),
+      name: candidate.name,
+      description: candidate.description,
+      invocation: candidate.invocation,
+      provider: SKILL_PROVIDER_NAME,
+      source: candidate.source,
+      resourceBase,
+      content: stripFrontmatter(await readFile(bodyUrl, 'utf8')),
     }
   },
 }

package/src/jev/backends.ts

@@ -10,6 +10,9 @@
  */
 
 export type JevBackendName = 'zen' | 'native' | 'openrouter'
+/** The local laya pace-maker writes its own call-layer rows; not a config
+ *  backend (no entry in BACKENDS) — it rides jevAsk's overrides hook. */
+export type JevPaceMakerName = 'laya'
 export type JevFamily = 'systemone' | 'decisions'
 
 export interface JevBackend {

package/src/jev/client.ts

@@ -51,6 +51,10 @@ export interface JevAskArgs {
   fallback?: boolean
   /** Host logger (optional, warn-only, best-effort). */
   logger?: GateLogger
+  /** Dual-run pace-maker hook (jev/dual.ts): target this endpoint/model with
+   *  no auth header and tag the call-layer row backend='laya'. Never set by
+   *  user config — internal to the dual wrapper. */
+  overrides?: { endpoint: string; model: string }
 }
 
 export type JevAnswer = Record<string, unknown>
@@ -214,9 +218,22 @@ export function normalizeQuestions(questions: Record<string, JevQuestion>, famil
  *  resolves to a structured failure (never throws) so the calling lane can
  *  fail open. Writes the decisions.jsonl call-layer row on EVERY path. */
 export async function jevAsk(args: JevAskArgs): Promise<JevAskResult> {
-  const backend = resolveBackend(args.config.jevBackend)
+  // Pace-maker override: a synthetic laya backend (no key — the auth header
+  // carries a placeholder the local server ignores; the secret gate and the
+  // call-layer row still run, tagged backend='laya').
+  const backend = args.overrides
+    ? ({
+        name: 'laya',
+        family: 'systemone',
+        endpoint: args.overrides.endpoint,
+        defaultModel: args.overrides.model,
+        keyEnv: '',
+        keychainService: '',
+        keychainByDefault: false,
+      } as unknown as ReturnType<typeof resolveBackend>)
+    : resolveBackend(args.config.jevBackend)
   // '' is the sentinel; undefined (bare harnesses) resolves to the default too
-  const model = args.config.jevModel ? args.config.jevModel : backend.defaultModel
+  const model = args.overrides ? args.overrides.model : args.config.jevModel ? args.config.jevModel : backend.defaultModel
   const fallback = args.fallback === true // absent => false: the column is call-time-invariant (fail-open = outcome !== ok)
   const state = typeof args.state === 'string' ? args.state : ''
   const started = Date.now()
@@ -259,8 +276,9 @@ export async function jevAsk(args: JevAskArgs): Promise<JevAskResult> {
     )
   }
 
-  // 3. API key — absent key fails before any network activity.
-  const key = await resolveApiKey(backend)
+  // 3. API key — absent key fails before any network activity. The laya
+  //    pace-maker skips resolution entirely (local, no auth).
+  const key = args.overrides ? 'laya-local' : await resolveApiKey(backend)
   if (key === '') {
     const hint =
       backend.keychainByDefault

package/src/jev/dual.ts

@@ -0,0 +1,73 @@
+/**
+ * Dual-run pace-maker (design §3.4, added 09-27): fire the configured
+ * backend AND the local laya-serve endpoint in parallel on every jev call.
+ *
+ *  - the PRIMARY answer drives the decision whenever it is ok;
+ *  - the laya answer is always logged (its own call-layer row tagged
+ *    backend='laya' plus verdict rows) — a free, permanently running
+ *    laya-vs-jev comparison on identical questions;
+ *  - TELEMETRY-ONLY: laya NEVER drives a decision. On real data its
+ *    within-batch top-1 agreed with the primary 0/14 and its negatives sit in
+ *    the same band as its positives (0.63-0.76), so both its relative ranking
+ *    and its absolute scores are unusable (4-way bench, 09-27). A primary
+ *    FAILURE (cold-start spike, timeout, 5xx…) is therefore a failure even
+ *    when laya answered: the result stays ok:false and the caller fails open
+ *    (ADR 0018). `layaResult` still carries the comparison answers so the
+ *    caller can log the laya verdict rows;
+ *  - laya not running = a refused connection in milliseconds — the switch
+ *    can stay on unconditionally.
+ *
+ * The consolidation prefilter deliberately does NOT ride this wrapper: laya's
+ * Chinese noul is the weak spot, and a wrong veto silently skips real
+ * consolidation work. Prefilter keeps its plain fail-open.
+ *
+ * @module jev/dual
+ */
+
+import { jevAsk } from './client.ts'
+import type { JevAskArgs, JevAskConfig, JevAskResult } from './client.ts'
+
+/** Default local laya-serve endpoint (laya repo: `laya[serve]` binds 8000). */
+export const DEFAULT_LAYA_URL = 'http://127.0.0.1:8000/v1/systemone'
+
+export interface JevDualConfig extends JevAskConfig {
+  /** Pace-maker switch (default OFF — zero extra requests when false). */
+  jevLayaFallback: boolean
+  /** Local laya-serve systemone endpoint. */
+  jevLayaUrl: string
+}
+
+export interface JevDualArgs extends Omit<JevAskArgs, 'config'> {
+  config: JevDualConfig
+}
+
+export type JevDualSource = 'primary' | 'laya'
+
+export type JevDualResult = JevAskResult & {
+  /** Which backend ANSWERED for the record: 'primary' when the primary
+   *  answered (the only case that may drive), 'laya' when it failed and the
+   *  pace-maker covered for telemetry only. */
+  source: JevDualSource
+  /** The pace-maker's own result when it ran (null when the switch is off),
+   *  always logged by jevAsk via the overrides hook. */
+  layaResult: JevAskResult | null
+}
+
+export async function jevAskDual(args: JevDualArgs): Promise<JevDualResult> {
+  if (args.config.jevLayaFallback !== true) {
+    const r = await jevAsk(args as JevAskArgs)
+    return { ...r, source: 'primary', layaResult: null }
+  }
+  const layaArgs: JevAskArgs = {
+    ...args,
+    config: { ...args.config, jevModel: 'laya-rl-agent' },
+    overrides: { endpoint: args.config.jevLayaUrl || DEFAULT_LAYA_URL, model: 'laya-rl-agent' },
+  }
+  const [primary, laya] = await Promise.all([jevAsk(args as JevAskArgs), jevAsk(layaArgs)])
+  if (primary.ok) return { ...primary, source: 'primary', layaResult: laya }
+  // REVISED 09-27 (real-data evidence: within-batch top-1 agreement 0/14):
+  // a primary failure is a FAILURE even when the pace-maker answered — laya
+  // is telemetry-only. The caller fails open (ADR 0018); layaResult carries
+  // the comparison answers so the caller can still log laya verdict rows.
+  return { ok: false, outcome: primary.outcome, message: primary.message, latencyMs: primary.latencyMs, source: 'laya', layaResult: laya }
+}

package/src/jev/log.ts

@@ -70,6 +70,14 @@ export interface JevVerdictRecord {
   probability: number
   band: JevBand
   agree: 'hit' | 'nearFloor' | 'gate-blocked' | 'wouldBlock' | 'n/a'
+  /** Which backend produced this verdict: 'primary' (default — the configured
+   *  backend) or 'laya' (the local pace-maker). Design §3.4. */
+  backend?: string
+  /** True when this row's source DROVE the decision because the primary
+   *  backend was unavailable (degraded mode). Comparison rows from the
+   *  pace-maker are degraded=false; laya rows are excluded from ECE either
+   *  way (uncalibrated probabilities). */
+  degraded?: boolean
 }
 
 const COMPACT_LIMIT = 512 * 1024 // ~512KB cap, same as the ilog sidecars
@@ -178,6 +186,8 @@ export async function logVerdicts(recs: readonly JevVerdictRecord[]): Promise<vo
         probability: rec.probability,
         band: rec.band,
         agree: rec.agree,
+        backend: rec.backend ?? 'primary',
+        degraded: rec.degraded === true,
       }),
     ),
   )

package/src/quality.ts

@@ -28,8 +28,9 @@ import { searchTopics } from './retrieval.ts'
 import type { RingEntry } from './observer.ts'
 import type { QueryBuildShape } from './ilog.ts'
 import type { TopicsConfigValue } from './config.ts'
-import { jevAsk } from './jev/client.ts'
-import type { JevAskConfig, JevQuestion } from './jev/client.ts'
+import { jevAskDual } from './jev/dual.ts'
+import type { JevQuestion } from './jev/client.ts'
+import type { JevDualConfig, JevDualResult } from './jev/dual.ts'
 import { band as jevBand } from './jev/thresholds.ts'
 import type { JevBand } from './jev/thresholds.ts'
 import { digestFor, logVerdicts } from './jev/log.ts'
@@ -125,14 +126,48 @@ function noulProbability(answer: unknown): number | undefined {
   return typeof p === 'number' && Number.isFinite(p) && p >= 0 && p <= 1 ? p : undefined
 }
 
+/** Pace-maker comparison rows for the slow rerank (design §3.4, REVISED
+ *  09-27): laya is TELEMETRY ONLY, so its rows are recorded band='record'
+ *  (its absolute scores never claim adopt) and degraded=false whether or not
+ *  the primary answered — landing them costs no behavior. An unparseable
+ *  answer gets no fabricated row. */
+function layaRerankRows(
+  at: string,
+  state: string,
+  metas: { qid: string; slug: string; instructions: string }[],
+  answers: Record<string, unknown> | undefined,
+): JevVerdictRecord[] {
+  const recs: JevVerdictRecord[] = []
+  for (const m of metas) {
+    const probability = noulProbability(answers?.[m.qid])
+    if (probability === undefined) continue
+    recs.push({
+      at,
+      lane: 'slowlane-rerank',
+      questionId: m.qid,
+      qtype: 'noul',
+      ref: `slug:${m.slug}`,
+      digest: digestFor(state, m.qid, m.instructions),
+      probability,
+      band: 'record',
+      agree: 'n/a',
+      backend: 'laya',
+      degraded: false,
+    })
+  }
+  return recs
+}
+
 /** The config slice jevAsk needs, tolerating bare harnesses (§4: consumers
  *  treat undefined as the documented default). */
-function jevConfigOf(cfg: TopicsConfigValue): JevAskConfig {
+function jevConfigOf(cfg: TopicsConfigValue): JevDualConfig {
   return {
     jevBackend: cfg.jevBackend ?? 'zen',
     jevModel: cfg.jevModel ?? '',
     jevTimeoutMs: typeof cfg.jevTimeoutMs === 'number' && cfg.jevTimeoutMs > 0 ? cfg.jevTimeoutMs : 3000,
     jevSecretFile: cfg.jevSecretFile ?? '',
+    jevLayaFallback: cfg.jevLayaFallback === true,
+    jevLayaUrl: cfg.jevLayaUrl ?? 'http://127.0.0.1:8000/v1/systemone',
   }
 }
 
@@ -320,35 +355,45 @@ export class SlowLane {
       metas.push({ qid, slug: cand.slug, instructions, disposition: cand.disposition })
     }
     if (metas.length === 0) return
-    const result = await jevAsk({
+    const dual = await jevAskDual({
       state,
       questions,
       config: jevConfigOf(this.service.cfg),
       lane: 'fastgate-shadow',
       fallback: false, // pure shadow — no behavior to fall back
     })
-    if (!result.ok) return
     const at = new Date().toISOString()
+    // Verdict rows for BOTH sources when they answered: the primary row set
+    // and the laya pace-maker comparison set (backend='laya', excluded from
+    // ECE, §6.2). Pure shadow — nothing gates here, so the laya rows land
+    // whenever laya answered, primary success or not.
+    const sources: { answers: Record<string, unknown> | undefined; backend: string; degraded: boolean }[] = []
+    if (dual.ok) sources.push({ answers: dual.answers, backend: 'primary', degraded: false })
+    if (dual.layaResult?.ok) sources.push({ answers: dual.layaResult.answers, backend: 'laya', degraded: false })
     const recs: JevVerdictRecord[] = []
-    for (const m of metas) {
-      const probability = noulProbability(result.answers[m.qid])
-      if (probability === undefined) continue // unparseable answer: no fabricated row
-      const b = jevBand(probability, 'rerank')
-      // agree reconciliation (§6.2): the ilog disposition, with wouldBlock for
-      // the reverse disagreement (lexical let it through, jev lands in the veto
-      // band — the other half of the calibration goldmine).
-      const agree: JevVerdictRecord['agree'] = m.disposition === 'hit' && b === 'fallback' ? 'wouldBlock' : m.disposition
-      recs.push({
-        at,
-        lane: 'fastgate-shadow',
-        questionId: m.qid,
-        qtype: 'noul',
-        ref: `slug:${m.slug}`,
-        digest: digestFor(state, m.qid, m.instructions),
-        probability,
-        band: b,
-        agree,
-      })
+    for (const src of sources) {
+      for (const m of metas) {
+        const probability = noulProbability(src.answers?.[m.qid])
+        if (probability === undefined) continue // unparseable answer: no fabricated row
+        const b = jevBand(probability, 'rerank')
+        // agree reconciliation (§6.2): the ilog disposition, with wouldBlock for
+        // the reverse disagreement (lexical let it through, jev lands in the veto
+        // band — the other half of the calibration goldmine).
+        const agree: JevVerdictRecord['agree'] = m.disposition === 'hit' && b === 'fallback' ? 'wouldBlock' : m.disposition
+        recs.push({
+          at,
+          lane: 'fastgate-shadow',
+          questionId: m.qid,
+          qtype: 'noul',
+          ref: `slug:${m.slug}`,
+          digest: digestFor(state, m.qid, m.instructions),
+          probability,
+          band: b,
+          agree,
+          backend: src.backend,
+          degraded: src.degraded,
+        })
+      }
     }
     await logVerdicts(recs)
   }
@@ -544,7 +589,7 @@ export class SlowLane {
       questions[qid] = { type: 'noul', instructions, criteria: candidateCriteria(c) }
       metas.push({ qid, slug: c.slug, instructions })
     }
-    const result = await jevAsk({
+    const dual = await jevAskDual({
       state,
       questions,
       config: jevConfigOf(this.service.cfg),
@@ -555,27 +600,45 @@ export class SlowLane {
       // stat. Fail-open occurrence = outcome !== 'ok' (this lane's fallback
       // to the legacy LLM rerank is deterministic per ADR 0018).
     })
-    if (!result.ok) return undefined
+    // REVISED 09-27 (real-data evidence): the laya pace-maker is TELEMETRY
+    // ONLY. Its within-batch top-1 agreed with the primary 0/14 on real
+    // queries (and its negatives sit at 0.63-0.76 — absolute scores unusable),
+    // so the earlier degraded-ranking takeover is REVERTED: on primary
+    // failure the lane falls open to the legacy LLM rerank exactly as
+    // ADR 0018 says. The pace-maker still answers, so its comparison rows
+    // land FIRST here (pure telemetry), then the legacy rerank takes over.
+    if (!dual.ok) {
+      if (dual.layaResult?.ok) {
+        await logVerdicts(layaRerankRows(new Date().toISOString(), state, metas, dual.layaResult.answers))
+      }
+      return undefined
+    }
     const scored: { qid: string; slug: string; instructions: string; probability: number; band: JevBand }[] = []
     for (const m of metas) {
-      const probability = noulProbability(result.answers[m.qid])
+      const probability = noulProbability(dual.answers[m.qid])
       if (probability === undefined) return undefined // 坏答案 → fail-open (§1)
       scored.push({ ...m, probability, band: jevBand(probability, 'rerank') })
     }
     const at = new Date().toISOString()
-    await logVerdicts(
-      scored.map((m) => ({
-        at,
-        lane: 'slowlane-rerank' as const,
-        questionId: m.qid,
-        qtype: 'noul' as const,
-        ref: `slug:${m.slug}`,
-        digest: digestFor(state, m.qid, m.instructions),
-        probability: m.probability,
-        band: m.band,
-        agree: 'n/a' as const,
-      })),
-    )
+    const recs: JevVerdictRecord[] = scored.map((m) => ({
+      at,
+      lane: 'slowlane-rerank' as const,
+      questionId: m.qid,
+      qtype: 'noul' as const,
+      ref: `slug:${m.slug}`,
+      digest: digestFor(state, m.qid, m.instructions),
+      probability: m.probability,
+      band: m.band,
+      agree: 'n/a' as const,
+      backend: dual.source,
+      degraded: false,
+    }))
+    // Pace-maker comparison rows (design §3.4): when the primary drove and
+    // laya also answered, log laya's verdicts alongside — the permanently
+    // running laya-vs-jev comparison. band='record': laya's absolute scores
+    // never claim adopt.
+    if (dual.layaResult?.ok) recs.push(...layaRerankRows(at, state, metas, dual.layaResult.answers))
+    await logVerdicts(recs)
     return scored
       .filter((m) => m.band === 'adopt')
       .sort((a, b) => b.probability - a.probability)