diff --git a/plugins/jsc-mod-companion/.claude-plugin/plugin.json b/plugins/jsc-mod-companion/.claude-plugin/plugin.json
index 624b2be..ae8c5c7 100644
--- a/plugins/jsc-mod-companion/.claude-plugin/plugin.json
+++ b/plugins/jsc-mod-companion/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "jsc-mod-companion",
-  "version": "0.0.1",
+  "version": "0.0.2",
   "description": "小說角色陪伴：人設注入、語氣、心情與好感、互動記憶、主動關心與待辦，並可用 ingest skill 解析小說建立角色",
   "license": "MIT",
   "author": {
diff --git a/plugins/jsc-mod-companion/README.md b/plugins/jsc-mod-companion/README.md
index 8ec9e5e..a9d60fe 100644
--- a/plugins/jsc-mod-companion/README.md
+++ b/plugins/jsc-mod-companion/README.md
@@ -82,6 +82,16 @@
 /jsc-mod-companion:ingest chapter1.md 章節編號 1
 ```
 
+## 人格 18 參數（與第 9 章的差異）
+
+人格參數依 wiki 9.2 實作於 `hooks/companion/core/params.ts`：
+
+- 三層值：角色包基準（`core.json` 的 `params`）＋存檔偏移（每次 ±1，上限 ±15）＋今日偏移（N(0,5)，上限 ±10），總值限制在 0–100。
+- 衍生值：有效心防、有效害羞、有效口是心非，公式同 9.2；信賴與親密以 0–100 輸入。
+- Big Five 保留為推導值：有 `params` 時，神經質（心情半衰期）與外向性（關心頻率）改由 18 參數推導，加權為設計啟發式；沒有 `params` 的舊角色沿用 `neuroticism`、`extraversion`。ingest 的 Big Five 推論與證據機制不變。
+- 姓名、性別、年齡、核心信念不可由 AI 改動，`filterAiPatch` 會剔除這些欄位。
+- 目前只提供計算與儲存欄位（`state.json` 的 `params`）；每日抽樣、存檔偏移累積的觸發時機留待後續工作項目。
+
 ## 評估標籤
 
 使用者每句話會對應到一個評估標籤，再由 `rules.ts` 的規則表換成心情增減量；模型只負責分類，不給數值。
diff --git a/plugins/jsc-mod-companion/hooks/companion/core.ts b/plugins/jsc-mod-companion/hooks/companion/core.ts
index ba1438c..47cf70c 100644
--- a/plugins/jsc-mod-companion/hooks/companion/core.ts
+++ b/plugins/jsc-mod-companion/hooks/companion/core.ts
@@ -4,6 +4,7 @@ import { readJson, writeJson } from './fs'
 import type { Clock, Fs } from './fs'
 import { applyDelta, decay, keywordEvent, MOOD_EVENTS, moodLabel } from './mood'
 import type { MoodEvent } from './mood'
+import { extraversionOf, neuroticismOf } from './core/params'
 import { extractTag } from './tags'
 import { createCare } from './care'
 import type { Care, CareProfile, CareSettings, CareEngine, Signals } from './care'
@@ -79,7 +80,7 @@ export function createCompanion({ fs, clock, root, builtins = [FUBUKI] }: Compan
     const st = await getState(active)
     if (!core || !st) return null
     const now = clock.now()
-    const mood = decay(st.mood, core.baseline, now - st.mood.at, core.neuroticism)
+    const mood = decay(st.mood, core.baseline, now - st.mood.at, neuroticismOf(core, st))
     const aff = idleDecay(st.affinity, now)
     const todos = (await readJson<{ done: boolean }[]>(fs, `${dir(active)}/todos.json`))?.filter(t => !t.done).length ?? 0
     return { core, mood, ...moodLabel(mood), affinity: aff.value, stage: STAGES[aff.stage]!, todos }
@@ -96,7 +97,7 @@ export function createCompanion({ fs, clock, root, builtins = [FUBUKI] }: Compan
 
   async function moodEvent(ev: MoodEvent) {
     await mutate((core, st, now) => {
-      const cur = decay(st.mood, core.baseline, now - st.mood.at, core.neuroticism)
+      const cur = decay(st.mood, core.baseline, now - st.mood.at, neuroticismOf(core, st))
       return { ...st, mood: { ...applyDelta(cur, MOOD_EVENTS[ev]), at: now } }
     })
   }
@@ -201,10 +202,11 @@ export function createCompanion({ fs, clock, root, builtins = [FUBUKI] }: Compan
     const v = await view()
     if (!v) return null
     const u = await getUser()
+    const st = await getState(v.core.id)
     const qs = hourOf(u.quiet.start)
     const qe = hourOf(u.quiet.end)
     ctx = {
-      profile: { attribute: v.core.attribute ?? 'gentle', extraversion: v.core.extraversion ?? 50, stage: CARE_STAGES[STAGES.indexOf(v.stage as (typeof STAGES)[number])] ?? 'stranger', nickname: u.nickname, freq: v.core.freq },
+      profile: { attribute: v.core.attribute ?? 'gentle', extraversion: extraversionOf(v.core, st ?? undefined), stage: CARE_STAGES[STAGES.indexOf(v.stage as (typeof STAGES)[number])] ?? 'stranger', nickname: u.nickname, freq: v.core.freq },
       settings: { dailyCap: u.careLimit, tzOffsetMin: tzMinutes(u.tz), quietHours: qs !== null && qe !== null ? { start: qs, end: qe } : undefined },
     }
     if (bundle?.id !== v.core.id) {
diff --git a/plugins/jsc-mod-companion/hooks/companion/core/params.test.ts b/plugins/jsc-mod-companion/hooks/companion/core/params.test.ts
new file mode 100644
index 0000000..a2bf32d
--- /dev/null
+++ b/plugins/jsc-mod-companion/hooks/companion/core/params.test.ts
@@ -0,0 +1,120 @@
+import { describe, test } from 'claude-code/testing'
+import assert from '../assert'
+
+import { applyInference, initPersonality } from './personality'
+import {
+  baseFor, deriveBig5, effective, extraversionOf, filterAiPatch, initLayers, neuroticismOf, nudgeSave, PARAMS, rollToday, value, values,
+} from './params'
+
+const near = (a: number, b: number, eps = 1) => assert.ok(Math.abs(a - b) <= eps, `${a} 應接近 ${b}`)
+
+describe('18 參數三層值', () => {
+  test('共 18 個參數，無屬性時基準為 50', () => {
+    assert.equal(PARAMS.length, 18)
+    assert.ok(Object.values(baseFor([])).every(v => v === 50))
+  })
+
+  test('存檔偏移每次 ±1，上限 ±15', () => {
+    let l = initLayers(baseFor([]))
+    for (let i = 0; i < 40; i++) l = nudgeSave(l, 'shy', 1)
+    assert.equal(l.save.shy, 15)
+    assert.equal(value(l, 'shy'), 65)
+    l = nudgeSave(l, 'shy', -1)
+    assert.equal(l.save.shy, 14)
+    for (let i = 0; i < 60; i++) l = nudgeSave(l, 'guard', -1)
+    assert.equal(l.save.guard, -15)
+  })
+
+  test('今日偏移限制在 ±10，且可由 rng 重現', () => {
+    const seq = (s: number) => () => ((s = (s * 16807) % 2147483647) / 2147483647)
+    const a = rollToday(initLayers(baseFor([])), seq(7))
+    const b = rollToday(initLayers(baseFor([])), seq(7))
+    assert.deepEqual(a.today, b.today)
+    assert.ok(PARAMS.every(p => Math.abs(a.today[p]!) <= 10))
+    // 極端抽樣也被截斷
+    const c = rollToday(initLayers(baseFor([])), (() => { let n = 0; return () => (n++ % 2 === 0 ? 1e-12 : 0) })())
+    assert.ok(PARAMS.every(p => Math.abs(c.today[p]!) <= 10))
+    assert.ok(PARAMS.some(p => Math.abs(c.today[p]!) === 10))
+  })
+
+  test('今日偏移的標準差約為 5', () => {
+    let s = 12345
+    const rng = () => ((s = (s * 16807) % 2147483647) / 2147483647)
+    const xs: number[] = []
+    for (let i = 0; i < 300; i++) xs.push(...Object.values(rollToday(initLayers(baseFor([])), rng).today) as number[])
+    const mean = xs.reduce((a, b) => a + b, 0) / xs.length
+    const sd = Math.sqrt(xs.reduce((a, b) => a + (b - mean) ** 2, 0) / xs.length)
+    near(mean, 0, 0.3)
+    near(sd, 4.7, 0.6) // 截斷在 ±2σ 使標準差略小於 5
+  })
+
+  test('參數值 = 基準 + 存檔 + 今日，並限制在 0–100', () => {
+    let l = initLayers({ ...baseFor([]), shy: 95 })
+    for (let i = 0; i < 15; i++) l = nudgeSave(l, 'shy', 1)
+    l = { ...l, today: { shy: 8 } }
+    assert.equal(value(l, 'shy'), 100)
+    assert.equal(values(l).blush, 50)
+  })
+})
+
+describe('有效參數', () => {
+  const l = initLayers({ ...baseFor([]), guard: 50, shy: 80, tsun: 70 })
+  test('有效心防：心防 × (1 − 0.5 × 信賴)，心防 50、信賴 90 約 28', () => {
+    near(effective(l, 90, 0).guard, 27.5, 0.01)
+    assert.equal(effective(l, 0, 0).guard, 50)
+  })
+  test('有效害羞：害羞 × (1 − 0.3 × 親密)', () => {
+    near(effective(l, 0, 100).shy, 56, 0.01)
+  })
+  test('有效口是心非：口是心非 × (1 − 0.4 × 親密)', () => {
+    near(effective(l, 0, 50).tsun, 56, 0.01)
+    near(effective(l, 0, 100).tsun, 42, 0.01)
+  })
+  test('信賴與親密超出範圍時被截斷', () => {
+    assert.equal(effective(l, 500, 500).guard, 25)
+  })
+})
+
+describe('Big Five 推導值與 ingest 相容', () => {
+  test('心防、忌妒越高，推導的神經質越高', () => {
+    const calm = deriveBig5(values(initLayers(baseFor(['onee']))))
+    const edgy = deriveBig5(values(initLayers(baseFor(['yandere']))))
+    assert.ok(edgy.N > calm.N)
+  })
+  test('主動度與玩心越高，推導的外向性越高', () => {
+    assert.ok(deriveBig5(values(initLayers(baseFor(['genki'])))).E > deriveBig5(values(initLayers(baseFor(['dandere'])))).E)
+  })
+  test('推導值落在 0–100', () => {
+    for (const a of ['tsundere', 'genki', 'kuudere', 'dandere', 'yandere', 'onee']) {
+      assert.ok(Object.values(deriveBig5(values(initLayers(baseFor([a]))))).every(x => x >= 0 && x <= 100))
+    }
+  })
+  test('ingest 的 Big Five 推論機制不受影響', () => {
+    const p = applyInference(initPersonality(['tsundere']), { N: { value: 99, evidence: [{ chapter: '1', quote: 'x' }] } })
+    assert.equal(p.dims.N.value, 85)
+  })
+})
+
+describe('核心設定保護', () => {
+  test('姓名、性別、年齡、核心信念不可由 AI 改動', () => {
+    const r = filterAiPatch({ name: '新名', gender: 'm', age: 30, coreBelief: 'x', shy: 60 })
+    assert.deepEqual(r.rejected.sort(), ['age', 'coreBelief', 'gender', 'name'])
+    assert.deepEqual(r.allowed, { shy: 60 })
+  })
+})
+
+describe('Big Five 推導值接上半衰期與關心頻率', () => {
+  const core = { neuroticism: 40, extraversion: 60 }
+  test('沒有參數時沿用舊欄位', () => {
+    assert.equal(neuroticismOf(core), 40)
+    assert.equal(extraversionOf(core), 60)
+  })
+  test('有參數時改由三層值推導，存檔偏移會影響結果', () => {
+    const params = baseFor(['tsundere'])
+    const a = neuroticismOf({ ...core, params })
+    const b = neuroticismOf({ ...core, params }, { params: { save: { guard: 15, jealous: 15 }, today: {} } })
+    assert.notEqual(a, 40)
+    assert.ok(b > a)
+    assert.equal(extraversionOf({ ...core, params }), deriveBig5(params).E)
+  })
+})
diff --git a/plugins/jsc-mod-companion/hooks/companion/core/params.ts b/plugins/jsc-mod-companion/hooks/companion/core/params.ts
new file mode 100644
index 0000000..d37efb3
--- /dev/null
+++ b/plugins/jsc-mod-companion/hooks/companion/core/params.ts
@@ -0,0 +1,127 @@
+import type { Vec } from './personality'
+
+/** 第 9.2 章的 18 個人格參數，0–100，越高越多 */
+export const PARAMS = [
+  'shy', 'blush', 'clingy', 'tsun', 'tone',
+  'jealous', 'depend', 'guard', 'recover', 'miss',
+  'initiative', 'assertive', 'playful', 'caring', 'gentle',
+  'cute', 'health', 'imagine',
+] as const
+export type Param = (typeof PARAMS)[number]
+export type ParamVec = Record<Param, number>
+
+export const PARAM_LABELS: Record<Param, string> = {
+  shy: '害羞傾向', blush: '臉紅傾向', clingy: '撒嬌傾向', tsun: '口是心非度', tone: '語尾與顏文字使用度',
+  jealous: '忌妒敏感度', depend: '依賴傾向', guard: '心防', recover: '釋懷速度', miss: '想念累積速度',
+  initiative: '主動度', assertive: '自我主張', playful: '玩心', caring: '照顧欲', gentle: '溫柔度',
+  cute: '服裝可愛偏好', health: '體質', imagine: '想像力',
+}
+
+export const SAVE_STEP = 1
+export const SAVE_LIMIT = 15
+export const TODAY_SIGMA = 5
+export const TODAY_LIMIT = 10
+
+/** 三層值：角色包基準（本機 core.json）、存檔偏移與今日偏移（state.json） */
+export type ParamLayers = { base: ParamVec; save: Partial<ParamVec>; today: Partial<ParamVec> }
+
+const clamp = (x: number, lo: number, hi: number) => Math.min(hi, Math.max(lo, x))
+
+const NEUTRAL = Object.fromEntries(PARAMS.map(p => [p, 50])) as ParamVec
+
+/** 屬性標籤對 18 參數的先驗（設計啟發式）；未列出的參數維持 50 */
+export const ARCHETYPE_PARAMS: Record<string, Partial<ParamVec>> = {
+  tsundere: { shy: 75, blush: 75, tsun: 85, guard: 65, assertive: 65, gentle: 35, jealous: 65, recover: 40 },
+  genki: { shy: 25, tone: 80, initiative: 85, playful: 75, depend: 55, guard: 25, recover: 75, health: 75, gentle: 60 },
+  kuudere: { shy: 30, blush: 25, tone: 20, initiative: 25, guard: 65, tsun: 40, playful: 25, gentle: 40 },
+  dandere: { shy: 85, blush: 80, tone: 30, initiative: 20, guard: 70, depend: 60, assertive: 25, gentle: 65 },
+  yandere: { jealous: 90, depend: 85, guard: 75, recover: 20, miss: 85, tsun: 35, caring: 75, assertive: 70 },
+  onee: { gentle: 85, caring: 85, shy: 25, blush: 25, playful: 65, guard: 40, recover: 70, assertive: 60 },
+}
+
+/** 以屬性標籤產生角色包基準：多個標籤取平均 */
+export function baseFor(archetypes: string[]): ParamVec {
+  const known = archetypes.filter(a => ARCHETYPE_PARAMS[a])
+  const out = { ...NEUTRAL }
+  if (!known.length) return out
+  for (const p of PARAMS) out[p] = known.reduce((s, a) => s + (ARCHETYPE_PARAMS[a]![p] ?? 50), 0) / known.length
+  return out
+}
+
+export function initLayers(base: ParamVec): ParamLayers {
+  return { base: { ...base }, save: {}, today: {} }
+}
+
+/** 存檔偏移：每次 ±1，上限 ±15 */
+export function nudgeSave(l: ParamLayers, p: Param, dir: 1 | -1): ParamLayers {
+  const next = clamp((l.save[p] ?? 0) + dir * SAVE_STEP, -SAVE_LIMIT, SAVE_LIMIT)
+  return { ...l, save: { ...l.save, [p]: next } }
+}
+
+/** 每日抽樣今日偏移：N(0, 5)，限制 ±10；rng 回傳 [0,1)，可注入以便測試 */
+export function rollToday(l: ParamLayers, rng: () => number = Math.random): ParamLayers {
+  const today: Partial<ParamVec> = {}
+  for (const p of PARAMS) {
+    const u1 = Math.max(rng(), 1e-12)
+    const z = Math.sqrt(-2 * Math.log(u1)) * Math.cos(2 * Math.PI * rng())
+    today[p] = clamp(z * TODAY_SIGMA, -TODAY_LIMIT, TODAY_LIMIT)
+  }
+  return { ...l, today }
+}
+
+export function value(l: ParamLayers, p: Param): number {
+  return clamp(l.base[p] + clamp(l.save[p] ?? 0, -SAVE_LIMIT, SAVE_LIMIT) + clamp(l.today[p] ?? 0, -TODAY_LIMIT, TODAY_LIMIT), 0, 100)
+}
+
+export function values(l: ParamLayers): ParamVec {
+  return Object.fromEntries(PARAMS.map(p => [p, value(l, p)])) as ParamVec
+}
+
+/** 信賴與親密皆為 0–100 */
+export function effective(l: ParamLayers, trust: number, intimacy: number) {
+  const t = clamp(trust, 0, 100) / 100
+  const i = clamp(intimacy, 0, 100) / 100
+  return {
+    guard: value(l, 'guard') * (1 - 0.5 * t),
+    shy: value(l, 'shy') * (1 - 0.3 * i),
+    tsun: value(l, 'tsun') * (1 - 0.4 * i),
+  }
+}
+
+/** Big Five 推導值：沿用半衰期（N）與關心頻率（E）；加權為設計啟發式 */
+export function deriveBig5(v: ParamVec): Vec {
+  const inv = (x: number) => 100 - x
+  const avg = (xs: [number, number][]) => xs.reduce((s, [x, w]) => s + x * w, 0)
+  return {
+    O: avg([[v.imagine, 0.5], [v.playful, 0.3], [v.initiative, 0.2]]),
+    C: avg([[v.caring, 0.4], [v.assertive, 0.3], [inv(v.playful), 0.3]]),
+    E: avg([[v.initiative, 0.35], [v.playful, 0.25], [v.tone, 0.2], [inv(v.shy), 0.2]]),
+    A: avg([[v.gentle, 0.4], [v.caring, 0.3], [inv(v.assertive), 0.15], [inv(v.tsun), 0.15]]),
+    N: avg([[v.guard, 0.25], [v.jealous, 0.25], [v.depend, 0.2], [v.shy, 0.15], [inv(v.recover), 0.15]]),
+  }
+}
+
+/** 核心設定：AI 不可改動 */
+export const LOCKED_FIELDS = ['name', 'gender', 'age', 'coreBelief'] as const
+
+/** 套用 AI 的修改提案：鎖定欄位一律剔除並回報，其餘原樣放行 */
+export function filterAiPatch<T extends Record<string, unknown>>(patch: T): { allowed: Partial<T>; rejected: string[] } {
+  const allowed: Record<string, unknown> = {}
+  const rejected: string[] = []
+  for (const [k, v] of Object.entries(patch)) {
+    if ((LOCKED_FIELDS as readonly string[]).includes(k)) rejected.push(k)
+    else allowed[k] = v
+  }
+  return { allowed: allowed as Partial<T>, rejected }
+}
+
+type WithParams = { params?: ParamVec; neuroticism?: number; extraversion?: number }
+type StateParams = { params?: { save: Partial<ParamVec>; today: Partial<ParamVec> } }
+
+/** 有角色包基準時，用三層值推導 Big Five；否則維持 core.json 的舊欄位 */
+function derived(core: WithParams, st?: StateParams) {
+  if (!core.params) return null
+  return deriveBig5(values({ base: core.params, save: st?.params?.save ?? {}, today: st?.params?.today ?? {} }))
+}
+export const neuroticismOf = (core: WithParams, st?: StateParams) => derived(core, st)?.N ?? core.neuroticism ?? 50
+export const extraversionOf = (core: WithParams, st?: StateParams) => derived(core, st)?.E ?? core.extraversion ?? 50
diff --git a/plugins/jsc-mod-companion/hooks/companion/types.ts b/plugins/jsc-mod-companion/hooks/companion/types.ts
index 4d727cd..1c161b7 100644
--- a/plugins/jsc-mod-companion/hooks/companion/types.ts
+++ b/plugins/jsc-mod-companion/hooks/companion/types.ts
@@ -1,4 +1,5 @@
 import type { Attribute, Trigger } from './care'
+import type { ParamVec } from './core/params'
 
 export type Pad = { p: number; a: number; d: number }
 
@@ -23,6 +24,8 @@ export type CompanionCore = {
   freq?: Partial<Record<Trigger, number>>
   // 神經質 0~100，影響心情半衰期；個性推導屬於後續里程碑，先用預設值
   neuroticism: number
+  // 人格 18 參數的角色包基準；有值時神經質改由參數推導
+  params?: ParamVec
   baseline: Pad
 }
 
@@ -41,6 +44,8 @@ export type CompanionState = {
   mood: Pad & { at: number }
   affinity: Affinity
   todos: number
+  // 人格 18 參數的存檔偏移與今日偏移（今日偏移於 dayKey 當天有效）
+  params?: { save: Partial<ParamVec>; today: Partial<ParamVec>; dayKey: string }
 }
 
 export type UserSettings = {