@geoly-ai/social-hub-cli 0.3.18 → 0.3.20

package/skills/reddit-phase-brief/SKILL.md

@@ -1,7 +1,7 @@
 ---
 name: reddit-phase-brief
 description: |
-  Reddit Phase 4 Brief skill for the matrix SOP. Use after reddit-phase-planning completes Phase 4.0.6 and before reddit-content-writing starts. It assigns the target flair first, then establishes the flair-level subreddit baseline, then locks ContentFunction and StructureFamily, then the outline; it also creates board rationale, content schedule, per-Key title/direction/outlines, brand/SKU placement strategy, prewriting control fields (AssignedFlair / PostingTrigger), structure pre-assignment (OpeningMove / EvidenceShape / BrandEntryMode / EndingMode / LengthBand / SampleSetID / RecentCollisionCheck), and the approval gate. Trigger when the user says "确认进入 Phase 4 Brief", "开始 Brief", "产出 Brief 汇报稿", or when planning is complete but drafting has not started.
+  Reddit Phase 4 Brief skill for the matrix SOP. Use after reddit-phase-planning completes Phase 4.0.6 and before reddit-content-writing starts. It locks the target flair first by verifying planning's `ProposedFlair` from the scheduling ledger, then establishes the flair-level subreddit baseline, then locks ContentFunction and StructureFamily, then the outline; it also creates board rationale, content schedule, per-Key title/direction/outlines, brand/SKU placement strategy, prewriting control fields (AssignedFlair / PostingTrigger / PersonaCardRef), structure pre-assignment (OpeningMove / EvidenceShape / BrandEntryMode / EndingMode / LengthBand / SampleSetID / RecentCollisionCheck), and the approval gate. Trigger when the user says "确认进入 Phase 4 Brief", "开始 Brief", "产出 Brief 汇报稿", or when planning is complete but drafting has not started.
 ---
 
 # Reddit Phase Brief
@@ -17,6 +17,8 @@ Read the shared contracts first:
 - `~/.cursor/skills/reddit-strategy-shared/references/context-contract.md`
 - `~/.cursor/skills/reddit-strategy-shared/references/content-type-routing-contract.md` (mandatory)
 - `~/.cursor/skills/reddit-strategy-shared/references/content-diversity-contract.md` (mandatory)
+- `~/.cursor/skills/reddit-strategy-shared/references/persona-card-contract.md` (mandatory — this skill **owns** the Persona Card)
+- `~/.cursor/skills/reddit-strategy-shared/references/native-copy-profile.md` only when the batch runs `AcceptanceProfile: NativeCopy`
 
 Then verify these project memory files exist:
 
@@ -28,6 +30,7 @@ Then verify these project memory files exist:
 - `memory/phase4-执行表.md`
 - `memory/phase4-SKU品牌词映射.md`
 - `memory/phase4-撰写批次map.md`
+- `memory/phase4-排期决策台账.md` with a `ProposedFlair` for every main-post Key in this batch (D2 output; source of `AssignedFlair`)
 - `memory/phase0.5-board-analysis.md`
 - `memory/phase1-VOC竞品缺口.md`
 - `memory/phase2-缺口分析.md`
@@ -59,7 +62,7 @@ guess the prototype from how the code reads.
 Read these local references before producing the brief:
 
 - `references/phase4-brief-workflow.md` for the full B-0 to B-3 workflow, templates, and gate checklist.
-- `references/brand-SKU-PGC-rules.md` before deciding brand/SKU placement for any P-B, C-B, or PGC row.
+- `references/brand-SKU-PGC-rules.md` before deciding brand/SKU placement for any P-B, C-B, or PGC row. Read §2.0 (placement **source** = the brand-first-mention histogram; §2 A/B/C/D are risk **ceilings**) and §2.5 (comparable-sample semantics; `BrandMentionSampleCount = 0` means *no comparable evidence*, not *titles disallowed*) before A/B/C/D.
 
 ## Workflow
 
@@ -72,10 +75,18 @@ outline before the structure is decided from evidence.
 2. Confirm the number of Brief items in this batch.
 3. Load or initialize `memory/phase4-内容排期表.md`.
 4. Check historical `memory/phase4-撰写brief.md` for overlap before creating a new batch.
-5. **Assign `AssignedFlair` (required, per main-post Key) before sampling.** flair 同时
-   决定合规资格、披露要求与触达，且决定 baseline 的采样口径。flair 必须取自该 subreddit
-   的现行 flair 列表（不得自造）；板块无 flair 或允许不选时用哨兵值 `NoFlairAvailable` /
-   `NoFlairRequired`（不得留空、不得用评论的 `N/A`）；缺失 → `BlockingGap`。评论 / 回复填 `N/A`。
+5. **Lock `AssignedFlair` (required, per main-post Key) before sampling — by verifying the
+   planning-side `ProposedFlair`, not by choosing a new flair.** Read `ProposedFlair` for the
+   Key from `memory/phase4-排期决策台账.md` (produced by `reddit-phase-planning` D2, contract
+   `reddit-phase-planning/references/scheduling-decision-procedure.md`), then verify it against
+   the subreddit's **currently selectable** flair list and the account's eligibility. Verified →
+   `AssignedFlair = ProposedFlair`. Not verifiable → `FlairVerificationStatus: Rejected` +
+   `BlockingGap` back to planning for a new `ProposedFlair`; the Brief must **not** substitute a
+   flair of its own. The single exception is `ProposedFlair: Unknown(NoSRP)` (planning ran in
+   `SRPMode: Fallback`), where the Brief picks from the live flair list and writes the choice and
+   reason back to the ledger (`FlairVerificationStatus: FallbackAssigned`). Flairless boards use
+   the sentinels `NoFlairAvailable` / `NoFlairRequired` (never blank, never the comment `N/A`);
+   a missing ledger row is a `BlockingGap`. Comments / replies use `N/A`.
 6. **Establish or refresh the flair-level subreddit baseline** for every target subreddit
    and assign a `SampleSetID` (see `references/phase4-brief-workflow.md` step B-1.0).
    优先按 `AssignedFlair` 细分取 exact-flair 样本；记录 `BaselineScope` /
@@ -85,9 +96,29 @@ outline before the structure is decided from evidence.
    Reuse a valid, unexpired baseline of the same flair coverage; only re-acquire when it is missing or stale.
 7. **Lock the project diversity window** (`WindowCount` 8–12, `WindowDays` 30–60) if not already locked, and read the rolling window from `memory/phase4-叙事指纹索引.md`.
 8. Produce board-selection rationale with evidence URLs, then the board overview table.
-9. For every post Key: resolve `客户内容原型` → `MappingRuleID`, lock `ContentFunction`, confirm the `EvidenceShape` actually in hand, narrow `StructureFamily` candidates with the ordered filter in `content-type-routing-contract.md`, pre-assign `OpeningMove` / `BrandEntryMode` / `CaveatMode` / `EndingMode` / `LengthBand` / `AnchorSlot`, and the `PostingTrigger`（受控值 `ConcreteFailure` / `PurchaseDecision` / `MeasurementResult` / `CompatibilityGap` / `UnsolicitedTestimony`；`UnsolicitedTestimony` 仅当 `BrandEntryMode: 不出现` 且有真实非征集非品牌控制的 testimony 证据），and **only then** run `RecentCollisionCheck` on the complete triple. On collision, first swap `OpeningMove` / `EndingMode` within the same family; reselect a family only if that fails. **Only then** write the outline.
+8b. **Run the material-modality preflight (G4, step B-1.4) before locking any structure.** Every selling point / material carries `MaterialID` + `OriginalModality` (`Required` / `Preferred` / `Optional` / `Forbidden`) + `EffectiveModality` + `ModalitySource`. Upgrading modality is blocked unless an `AuthoritativeSource` (`ClientContract` / `SubredditRule` / `ComplianceRequirement` / an already-authorized `ApprovedBrief`) backs it via `ModalityUpgradeRef`, with `ApprovedBy` ≠ the generating agent. **A material the user explicitly marked `Optional` is never upgraded** — go back to the user instead. Missing or conflicting modality is fail-closed. Disclosure/compliance requirements are not selling-point modality.
+9. For every post Key: resolve `客户内容原型` → `MappingRuleID`, lock `ContentFunction`, confirm the `EvidenceShape` actually in hand, narrow `StructureFamily` candidates with the ordered filter in `content-type-routing-contract.md`, pre-assign `OpeningMove` / `CaveatMode` / `EndingMode` / `LengthBand` / `AnchorSlot` (**`BrandEntryMode` is deliberately excluded here — it is assigned in 9b, after its gate**), and **record (not yet validate)** the `PostingTrigger` candidate — its cross-constraint with `BrandEntryMode` can only be checked after 9b（受控值 `ConcreteFailure` / `PurchaseDecision` / `MeasurementResult` / `CompatibilityGap` / `UnsolicitedTestimony`；`UnsolicitedTestimony` 仅当 `BrandEntryMode: 不出现` 且有真实非征集非品牌控制的 testimony 证据），and **only then** — after 9b has supplied `BrandEntryMode` — run `RecentCollisionCheck` on the complete triple. On collision, first swap `OpeningMove` / `EndingMode` within the same family; reselect a family only if that fails. **Only then** write the outline. Under `AcceptanceProfile: NativeCopy` a client requirement about how posts end narrows **the set this step may assign from** — it is never a question appended to a finished draft, since `EndingMode` is a locked field and one of the three collision fields. If the narrowed set leaves no collision-free option, that is a `BlockingGap` back to planning (`../reddit-strategy-shared/references/native-copy-profile.md`).
+9b. **Only now assign `BrandEntryMode`, and only through `BrandPlacementGate` (G2):** for **P-B / PGC posts**, placement is decided by the brand-first-mention histogram plus the §2 risk ceiling (`SelectedPlacement` = the later of the two, histogram bucket 结论 mapping to 后段; the domain is 标题/首段/中段/后段, and a class-D board or `NoBrandMentionRate ≥ 0.8` that leaves no viable spot is a `BlockingGap` back to planning, never a silent drop to 不出现/评论) — **never derived from the P-B label**. **P-NB posts** go straight to `不出现` without running the formula; **C-B comments** use `CommentPlacementDecision` plus the §2 ceiling and complement rules. Fill the placement evidence first, **branching by row type**: P-B / PGC posts need a complete `BrandPlacementBasis`; P-NB posts take `N/A(P-NB · 主帖不露品牌)`; C-B comments take `InheritedBrandPlacementBasis` + `CommentPlacementDecision` instead and are never required to carry their own. Legitimate title exposure is preserved for `PGCApproved` (precedent check passed, board allows vendor/official, flair/disclosure/frequency met) and `BrandReviewPost` (genuinely a post about that brand, judged from `PrimaryPostType` / `WritingVariant` plus topic evidence — never by inventing a `ContentFunction: Review` value). This gate governs *what decides placement only* — brand quota semantics (P-B / P-NB / C-B / C-NB / PGC) are untouched.
+9c. **Narrative layering (G1, step B-1.6) before writing title directions or outlines:** every post and comment carries `InternalCoverageIntent` (FAQ ids, fanout queries, keywords, gap codes, `prompt_id` — internal only) and a non-empty `ReaderVisiblePremise`. The gate scans only reader-facing fields (核心角度 / 标题方向 / 大纲 / 撰写逻辑一句话 / 评论任务方向 / outbound Reddit payload); Routing Decision, TraceFields and 方向选取理由 keep their internal labels. Converting FAQ evidence into a real reader question is the intended outcome — only raw internal labels in the narrative are blocked.
+9d. **Assign the Persona Card for every main-post Key (contract:
+   `../reddit-strategy-shared/references/persona-card-contract.md`).** The Card is the
+   third `PrewritingControlField` and this skill owns it: record `PersonaCardRef` +
+   `PersonaCardVersion` + `PersonaCardHash`, the capability fields
+   (`knowledgeAccess` — exactly one of `research` / `owned` / `used` / `observed` /
+   `supplied_by_brand`; `permittedExperienceClaims` — a possibly-empty subset of
+   `{ownership, delivery, wear, service}`; `attentionFocus`; `questionIntent`), and the
+   per-draft fields (`postingWhy`, `knows[]`, `doesNotKnow[]`, `permittedFactRefs`,
+   `brandRelationshipDisclosed`). `postingWhy` must be able to coexist with the assigned
+   `PostingTrigger` — they are different fields and neither substitutes for the other.
+   **The writer never authors a Card**; letting it do so would let the draft grant itself
+   whatever boundary it already needs. Comments / self-comments / external comments take
+   `N/A` — this train does not put comments under a Card. `permittedExperienceClaims` is
+   authorised from submitted context, never inferred from the brand relationship: a
+   partner may genuinely own the product, and an empty set is a normal Card. The Card is a
+   **declared posting context, not a verified identity** — it is `LocalOnly`, is never
+   sent to Hub, and never implies the publishing account was resolved.
 10. Create one Brief block for every execution-table Key in the selected batch, including posts, self-comments, and external comments.
-11. For all P-B/C-B/PGC rows, record brand/SKU placement rationale using `references/brand-SKU-PGC-rules.md`.
+11. For all P-B/C-B/PGC rows, record brand/SKU placement rationale using `references/brand-SKU-PGC-rules.md` (§3's five sample-learning classes are *rationale records only* — they never write `SelectedPlacement` directly nor bypass §2.0's mechanical rules).
 12. Write a `Planned` fingerprint row for every briefed post Key into `memory/phase4-叙事指纹索引.md` (`DraftHash: PENDING`), including the two `PrewritingControlFields` (`AssignedFlair`, `PostingTrigger`).
 13. Write or update the memory files listed below, then show the Brief gate checklist.
 
@@ -103,6 +134,14 @@ Baseline cards may live in `memory/phase4-撰写brief.md` or a dedicated baselin
 file; either way each is addressed by its `SampleSetID` and is reusable across
 Keys and batches within its validity period.
 
+Persona Cards may likewise live inside `memory/phase4-撰写brief.md` or in a
+dedicated `memory/phase4-persona-card.md`; either way each is addressed by its
+`PersonaCardRef` and carries its own `PersonaCardVersion` / `PersonaCardHash`.
+Unlike a baseline card a Persona Card is **per post Key, not shared** — its
+per-draft fields belong to one submission. Do not record Cards in
+`memory/phase4-叙事指纹索引.md`: the index has a fixed schema and a fixed hash
+definition, and the Card is not a fingerprint field.
+
 Optional Feishu Brief documents require an explicit user-provided folder link.
 
 ## Gate
@@ -110,10 +149,15 @@ Optional Feishu Brief documents require an explicit user-provided folder link.
 Writing may start only after:
 
 - Every target subreddit has a valid, unexpired baseline card with a `SampleSetID`, with flair coverage recorded (`BaselineScope` / `FlairCoverageStatus`); `FlairCoverageStatus` is `Exact` or `ApprovedFallback` (never `Insufficient`).
-- Every **main-post** Key carries a non-empty `AssignedFlair` — drawn from the subreddit's live flair list, or the sentinel `NoFlairAvailable` / `NoFlairRequired` for flairless boards — and its `AssignedFlair` matches the baseline card's covered flair (or an approved fallback).
+- Every **main-post** Key carries a non-empty `AssignedFlair` — verified against the ledger's `ProposedFlair` and the subreddit's live flair list (`FlairVerificationStatus: Verified` or `FallbackAssigned`; no `Rejected` left open), or the sentinel `NoFlairAvailable` / `NoFlairRequired` for flairless boards — and its `AssignedFlair` matches the baseline card's covered flair (or an approved fallback).
 - Every **main-post** Key's `客户内容原型` resolves to an `Approved` mapping row (`MappingRuleID` / `MappingVersion` recorded in the Routing Decision). Self-comments inherit the paired post's value; external comments use `—` and need no mapping of their own.
-- Every post Brief row carries the nine `LockedFingerprintFields` (`ContentFunction`, `StructureFamily`, `OpeningMove`, `EvidenceShape`, `BrandEntryMode`, `CaveatMode`, `EndingMode`, `LengthBand`, `AnchorSlot`), the two `PrewritingControlFields` (`AssignedFlair`, `PostingTrigger` — the latter satisfying the `PostingTrigger × BrandEntryMode` cross-constraint), and the four `TraceFields` (`SampleSetID`, `RecentCollisionCheck`, `MappingRuleID`, `MappingVersion`).
+- Every post Brief row carries the nine `LockedFingerprintFields` (`ContentFunction`, `StructureFamily`, `OpeningMove`, `EvidenceShape`, `BrandEntryMode`, `CaveatMode`, `EndingMode`, `LengthBand`, `AnchorSlot`), the three `PrewritingControlFields` (`AssignedFlair`, `PostingTrigger` — satisfying the `PostingTrigger × BrandEntryMode` cross-constraint — and `PersonaCardRef`), and the four `TraceFields` (`SampleSetID`, `RecentCollisionCheck`, `MappingRuleID`, `MappingVersion`).
+- Every **main-post** Key carries a non-empty `PersonaCardRef` with its `PersonaCardVersion` / `PersonaCardHash`, a single `knowledgeAccess`, a `permittedExperienceClaims` set drawn only from `{ownership, delivery, wear, service}` (empty is valid), non-empty `postingWhy`, and a `postingWhy` that can coexist with the assigned `PostingTrigger`. Comments take `N/A`. Role labels, if present, are non-authoritative and no gate reads them. Any Card with a non-empty `permittedExperienceClaims` also carries `DeclaredBy` (never the generating agent) and `ClaimAuthorizationRef`; without both, the claim set is reported as a Brief declaration rather than an authorised set.
 - `RecentCollisionCheck` is `Pass` for every row, or `Exception` with an approved `ExceptionRef`.
+- Every **P-B / PGC post** row carries a complete `BrandPlacementBasis` and a `SelectedPlacement` consistent with it (an already-planned P-B/PGC row may only land on 标题/首段/中段/后段 — moving the brand to a comment or dropping it is a quota change, so it goes back to planning); **C-B comments** carry `InheritedBrandPlacementBasis` + `CommentPlacementDecision` instead and never build their own histogram. No row derives title-brand exposure from the P-B label; `NoComparableEvidence` rows went through re-sampling / StopAndAsk / a recorded `BrandTitleExposureException`, not a default; `Sparse` rows meet the three §2.5 conditions.
+- Every post and comment Brief carries `InternalCoverageIntent` + a non-empty `ReaderVisiblePremise`, and no reader-facing field leaks internal labels.
+- The material-modality ledger is complete, with no unauthorized upgrades or downgrades (and nothing the user marked `Optional` upgraded).
+- Baseline cards cover all **10** dimensions (the 10th is the em-dash / parenthesis usage distribution).
 - `memory/phase4-叙事指纹索引.md` has exactly one active `Planned` row for every briefed **post** Key (comments are out of scope while `WindowScope: PostOnly`).
 - `memory/phase4-撰写brief.md` exists and covers every planned Key for the batch.
 - `brief_approval_mode` is recorded.
@@ -138,8 +182,15 @@ Writing may start only after:
 - FlairCoverageStatus: Exact / ApprovedFallback / Insufficient   # batch-level = worst status across cards (Insufficient > ApprovedFallback > Exact)
 - ClientPrototypeMappingStatus: AllApproved / BlockingGap
 - StructureFieldsAssigned: Yes / No
-- PrewritingControlFieldsAssigned: Yes / No   # AssignedFlair + PostingTrigger per main-post Key
+- PrewritingControlFieldsAssigned: Yes / No   # AssignedFlair + PostingTrigger + PersonaCardRef per main-post Key
 - PostingTriggerCrossConstraintOK: Yes / No    # UnsolicitedTestimony ⇒ BrandEntryMode 不出现
+- PersonaCardCoverage: <n>/<N>                # main-post Keys with a non-empty PersonaCardRef
+- PersonaCardTransportStatus: LocalOnly(ServerContractMissing)   # the Card is never sent to Hub
+- PersonaCardAuthorizationStatus: Authorized / DeclaredOnly / NotApplicable   # worst status across Keys; propagate to writing and delivery
+- AcceptanceProfile: Default / NativeCopy
+- NativeCopyProfileStatus: NotSelected / Unavailable(ServerContractMissing)
+- LocalReviewIntent: accountBinding=<account_bound|persona_less>, brandTreatment=<managed_brand|organic>   # local planner declaration; propagate to writing and delivery, never sent to Hub
+- HubReviewScopeStatus: Unavailable(ServerContractMissing)   # no server-side scope exists to report
 - FingerprintPlannedRows: <n>/<N>
 - DiversityWindow: WindowCount=<n>, WindowDays=<n>
 - RecentCollisionCheck: Pass / BlockingCollision / Exception(<ExceptionRef>)

package/skills/reddit-phase-brief/references/brand-SKU-PGC-rules.md

@@ -17,11 +17,90 @@ P-B / C-B 只表示该内容需要承载品牌词，不代表品牌名、SKU、
 
 禁止把所有 P-B 主帖批量设置为“标题或第一段露出品牌 / SKU”。
 
+> **上列因素的分工见 §2.0：** 「板块规则 / 推广容忍度 / 帖型 / 合规风险」是**风险边界**，
+> 「样本（GEOly 引用帖、VOC 帖、同板块高存活帖）」是**位置来源**。二者不可互换，
+> 且 **P-B 标签本身不是任何一方**。
+
 ---
 
-## 2. 默认露出位置规则
+## 2. 露出位置规则
+
+### 2.0 位置来源 vs 风险边界（G2 · 读 A/B/C/D 之前必须先读本节）
+
+> **出处（真实返工）：** 需求方质问「**为什么你把 P-B 的标题都统一加上了 aporro 这个
+> 品牌了？**」——P-B 的含义是「帖子中包含品牌词」，**并不代表品牌名必须在标题里出现**，
+> 那样显得很广告。修正后的口径：品牌词出现在 P-B 帖的**正文 body 中**，标题中通常不写
+> 品牌名，**除非是 Brand Review / 专门讨论该品牌的帖子**。
+
+本节确立两条不同性质的输入，**不得互相替代**：
+
+| | 是什么 | 谁给出 |
+|---|---|---|
+| **位置来源（source）** | 该板块真实帖子的**品牌首次出现位置直方图**（baseline 卡第 5 维；或板块级 SRP 的 `brandFirstMentionHistogram`） | 证据 |
+| **风险边界（ceiling）** | 下面 A/B/C/D 四类给出的**允许位置集合 / 最早允许位置** | 版规、合规、板块风险 |
+
+**取值规则（机械化，不留解释空间）：**
+
+```
+SelectedPlacement = 更晚的（§2 A/B/C/D 允许的最早位置, 直方图给出的经验位置）
+
+例外分支：持有有效 BrandTitleExposureException（§2.5）时
+  → 该例外可覆盖直方图与「标题档 < 20%」限制，SelectedPlacement 允许取「标题」
+  → 但仍不得越过 §2 D 类、版规与披露要求
+  → 此时 BrandPlacementBasis.SourceKind = Exception（**Exception 是合法的 SourceKind**）
+
+D 类是终止条件：主帖不露品牌，不参与「取更晚」比较
+
+档位映射（直方图档 → SelectedPlacement 受控值，必须按此换算）：
+  标题 → 标题   首段 → 首段   中段 → 中段   结论 → 后段
+  （直方图只有四档，SelectedPlacement 的「评论」/「不出现」不来自直方图）
+```
+
+**⚠️ 既有 P-B / PGC 行的取值域限制（防止位置改动变成配额改动）：**
+对**规划时已定为 P-B / PGC** 的行，`SelectedPlacement` 只能取
+`标题 / 首段 / 中段 / 后段` —— **不得取「评论」或「不出现」**。因为帖—评互补规则是
+「P-B 主帖含品牌 → 配对自评 C-NB；P-NB 主帖不含品牌 → 配对自评 C-B」，把 P-B 的品牌
+挪到评论或取消，实际上就把这行变成了 P-NB + C-B，**那是配额与行类型变更**。
+正确动作：`BlockingGap: BrandPlacementUnsupported` → 回规划决定。
+「评论」与「不出现」只对**规划时已是 P-NB** 的行合法。
+
+- 直方图允许标题、但 §2 判为 C / D（敏感 / 高风险）→ 按 C / D。
+- §2 判为 A（可标题或首段）、但直方图显示该板块几乎不在标题出现品牌 → **按直方图**。
+- **`SelectedPlacement` 永远不得从「这行标记 P-B」推导出来。** 出现该推导即
+  `BlockingGap: BrandPlacementUnsupported`。
+
+**直方图的经验位置怎么取（不得自由解释「占比低」）：**
+
+- `BrandFirstMentionHistogram` **只有四档：标题 / 首段 / 中段 / 结论**，分母是
+  `BrandMentionSampleCount`（**确实出现品牌的样本**）。「不出现」**不是本直方图的一档**，
+  它由另一个量 `NoBrandMentionRate`（= 1 − `BrandMentionSampleCount` /
+  `ComparableSampleCount`）表示；两者不得混为一谈（详见 §2.5）。
+- 取**众数档**（四档中占比最高的那一档）作为经验位置；
+- **并列时取更晚的那一档**（标题 < 首段 < 中段 < 结论），不得挑更靠前的；
+- 「标题」档占比 **< 20%** 时不得选标题位置，除非持有 §2.5 的合法例外；
+- `NoBrandMentionRate` **≥ 0.8**（该板块普遍不提品牌）**不直接决定位置**，它是把位置往后推
+  一档、或触发 §3.5 StopAndAsk 的信号；**它本身不构成 §2 D 类终止条件**。
+  后推到「后段」仍不可行时：**规划时已是 P-NB 的行**可落到「不出现」；
+  **规划时是 P-B / PGC 的行** → `BlockingGap` 回规划，**不得**静默落到「不出现」。
+  持有有效 `BrandTitleExposureException` 时，该后推信号**可被例外覆盖**
+  （例外仍受 D 类、版规与披露要求约束）。
+  `ComparableSampleCount = 0` 时该比率**未定义**，填 `N/A`，并直接按
+  `DecisionStatus: NoComparableEvidence` 处理。
+- **众数档已是「结论」时没有更晚的一档**：保持「结论」，不再后移（后移的下一步是
+  「评论 / 不出现」，那属于配额语义，见上方取值域限制）。
+
+**A/B/C/D 四类的读法（重要）：** 下面 A/B/C/D 是**风险边界**，即「最早允许到哪」和
+「哪些位置被版规/合规禁掉」，**不是位置来源**。看到某个帖型落在 A 类**不等于**应该把
+品牌放进标题 —— 还要直方图支持。
+
+⚠️ **本节不改动品牌配额语义。** P-B / P-NB / C-B / C-NB / PGC 的配额是否为合同硬约束，
+属另一批工作。本节只回答**位置由什么决定**，不回答**要不要有品牌**：位置不成立时正确
+动作是 `BlockingGap` 或回规划确认，**不得**静默把 P-B 改成 P-NB、不得自动改配对或配额，
+PGC 行也不因标题例外变成 P-B 行。
+
+---
 
-### A. 可标题或首段露出
+### A. 风险边界：可标题或首段露出
 
 适用于：
 
@@ -49,7 +128,7 @@ P-B / C-B 只表示该内容需要承载品牌词，不代表品牌名、SKU、
 
 ---
 
-### B. 默认中段露出
+### B. 风险边界：最早中段露出
 
 适用于：
 
@@ -73,7 +152,7 @@ P-B / C-B 只表示该内容需要承载品牌词，不代表品牌名、SKU、
 
 ---
 
-### C. 默认后段或评论露出
+### C. 风险边界：最早后段或评论露出
 
 适用于：
 
@@ -92,20 +171,89 @@ P-B / C-B 只表示该内容需要承载品牌词，不代表品牌名、SKU、
 
 ---
 
-### D. 高风险或禁止推广板块
+### D. 终止条件：高风险或禁止推广板块
 
-适用于：
+适用于（**只由明确规则触发，不由样本分布触发**）：
 
-- 明确禁止 self-promotion、advertising、commercial post 的板块
+- 明确禁止 self-promotion、advertising、commercial post 的板块（版规原文可引用）
 - 医疗、法律、隐私等高敏感板块
-- 样本显示高存活内容通常不直接提品牌的板块
+- 合规 / 法务明确列为高风险的板块
+
+> **v3 变更（G2）：删除「样本显示高存活内容通常不直接提品牌的板块」这一条 D 类触发项。**
+> 那是**样本分布**，属于位置来源，会把 source 直接升级成终止条件，与 §2.0 的分层冲突。
+> 该信号现在记为 `NoBrandMentionRate` 高，作用是**把位置往后推**（最多到「不出现」这一
+> `SelectedPlacement`）或触发 §3.5 StopAndAsk，**不自动把板块判成 D 类**。
+
+处理方式（**按该行规划时的既有类型分叉**）：
+
+- **规划时已是 P-NB 的行**：主帖不露品牌；如规则允许，再用 C-B 评论谨慎补位；
+  如果评论也有推广风险，则该板块不做品牌露出。
+- **规划时是 P-B / PGC 的行**：**不得**在本阶段把它降成「不露品牌」或挪到评论 ——
+  那是配额与行类型变更。正确动作是 `BlockingGap: BrandPlacementUnsupported`，
+  **回规划决定**（换板块 / 改配额 / 撤这行）。
+- 有效 `BrandTitleExposureException` **不能**越过 D 类：D 类的禁止推广规则优先。
+
+---
+
+### 2.5 品牌位置直方图的可比样本口径（G2 · 防止把「没样本」误读成「不允许」）
+
+**最容易犯的错：** baseline 的 5–10 条样本大多是普通 UGC，里面本来就没几条提品牌，
+于是「标题」档占比 = 0 —— 这只说明**没有可比较的品牌样本**，**不等于**「品牌标题被证明
+不允许」。把 0 当成一个保守位置，会把 PGC 与 Brand Review 的合法标题露出一并挡掉。
+
+因此直方图必须分三个量记录，缺一不可：
+
+| 字段 | 含义 |
+|---|---|
+| `ComparableSampleCount` | 该 Scope（subreddit + flair + `PrimaryPostType`/`ContentFunction` + 账号类型 UGC/PGC）下的合格样本总数 |
+| `ComparableScopeStatus` | `Exact`（四维全分层）/ `ApprovedFallback`（放宽了某一维，须记放宽了哪一维 + `ComparableScopeApprovedBy`）/ `Insufficient`（→ 按 `NoComparableEvidence` 处理）。⚠️ **baseline 采样默认只按 subreddit + flair 分层**（见 brief B-1.0），因此帖型 / 账号类型这两维通常需要显式声明：要么在建卡时补分层，要么记 `ApprovedFallback`。**混了不同帖型或 UGC/PGC 的样本，直方图就不是本行的有效位置来源** |
+| `BrandMentionSampleCount` | 其中**确实出现品牌 / 产品**的样本数 —— **直方图的真实分母** |
+| `NoBrandMentionRate` | `1 − BrandMentionSampleCount / ComparableSampleCount`，即该板块**根本不提品牌**的比例。**它不是直方图的一档**，也不构成 D 类终止条件 |
+| `BrandFirstMentionHistogram` | 在 `BrandMentionSampleCount` 这个分母内，首次出现位置落在 **标题 / 首段 / 中段 / 结论** 四档的占比（**没有「不出现」这一档**） |
 
-处理方式：
+**`DecisionStatus` 四值（producer 与 consumer 必须同口径）：**
 
-- 主帖不露品牌
-- 使用 P-NB 主帖
-- 如规则允许，再用 C-B 评论谨慎补位
-- 如果评论也有推广风险，则该板块不做品牌露出
+| 值 | 判据 | 处置 |
+|---|---|---|
+| `Supported` | `BrandMentionSampleCount ≥ 5` | 按 §2.0 取值规则直接使用 |
+| `Sparse` | `1 ≤ BrandMentionSampleCount ≤ 4` | **可用**，但须同时满足三条：①Brief 写明**三个计数字段**（`ComparableSampleCount` / `BrandMentionSampleCount` / `NoBrandMentionRate`）**与四档直方图**；②在众数档与其下一档之间取**更保守（更晚）**的一档（众数已是「结论」则保持「结论」）；③记 `SparseApprovedBy`（显式批准人，≠ 生成本 Brief 的 agent）。三条缺一 → 按 `NoComparableEvidence` 处理 |
+| `NoComparableEvidence` | `BrandMentionSampleCount = 0` | **不得**当成「不允许标题」；走 §3.5 的补样本 / 换来源 / StopAndAsk / 合法例外 |
+| `Exception` | 持有下方**有效**合法例外 | **该状态优先于 `NoComparableEvidence`**：例外成立时 `DecisionStatus` 直接记 `Exception`（不再同时报 `NoComparableEvidence`），须记 `BrandTitleExposureException` 四字段；此时 `SourceKind: Exception`，`SourceRef` 填 `ExceptionRef`（或 `ExceptionBasis` 的引用），**不要求 SampleSetID** |
+
+**零分母时的字段填法（避免必填字段互相打架）：** `BrandMentionSampleCount = 0` 时，
+`BrandFirstMentionHistogram` 与 `NoBrandMentionRate` 均填 `N/A`（`ComparableSampleCount = 0`
+时两者同样 `N/A`），**不得编造占比**；`ComparableSampleCount` 与 `BrandMentionSampleCount`
+本身仍必填（写 0）。
+
+**`NoComparableEvidence` 的正确动作（按序）：** ①扩大时间窗 / 放宽到 `FlairCluster` 重采；
+②改用板块级 SRP 的 `brandFirstMentionHistogram`（契约见
+`../../reddit-strategy-shared/references/subreddit-research-profile.md`；**该板块暂无可用 SRP 实例时
+以 baseline 卡为主来源，不得因其缺失而无条件阻断**）；③仍无可比样本 → **停下来问人**
+（策略师 / 用户），或走下方例外；④都不成立 → `BlockingGap`。
+
+#### ⚠️ 合法例外（必须保留，不得一刀切）
+
+以下两类**标题出现品牌名是合法的**，即使直方图不支持：
+
+| 例外值 | 成立条件（全部满足） |
+|---|---|
+| `PGCApproved` | ①账号类型为官方 / PGC；②已过 §4 PGC account precedent check（**有真实先例**，须给出先例 URL）；③版规允许 brand / vendor / official account 内容；④flair / disclosure / frequency 要求均满足 |
+| `BrandReviewPost` | 该帖**确实**围绕该品牌 / 该产品本身展开。判据用**现有受控字段组合**，两个条件**必须同时成立**：<br>`( PrimaryPostType ∈ {ROUNDUP, OWNER, OPINION} OR WritingVariant ∈ {COMPARE, WORTHIT, DATA} )`<br>**AND** 选题证据表明**内容主体就是评这个品牌 / 产品**（须给出证据）。<br>**只命中枚举、没有选题证据 → 不成立。** 另：**不得**为此新造 `ContentFunction: Review` —— `ContentFunction` 是受控枚举，没有 Review 值，自造即非法；§2 A 里的自然语言「Review」是帖子意图描述，**不是受控字段值** |
+
+**例外不是绕过口：**
+
+- 只有「这是 PGC」「这是 Review」的**自我描述**、拿不出先例或内容并非专门讨论该品牌
+  → 仍是 `BlockingGap`，不得放行；
+- 例外**不得越过 D 类**（禁止 self-promotion / 高风险板块）与任何披露要求；
+- 例外须**拆字段**记录，不得只塞进一段括号文本：
+
+```markdown
+BrandTitleExposureException: None / PGCApproved / BrandReviewPost
+ExceptionBasis:      （PGC：先例 URL + 版规条款；BrandReview：PrimaryPostType/WritingVariant + 选题证据）
+ExceptionScope:      （适用到哪些 Key / 哪个 subreddit / 有效期）
+ExceptionApprovedBy: （批准人，≠ 生成本 Brief 的 agent）
+ExceptionApprovedAt: （ISO-8601）
+```
 
 ---
 
@@ -121,6 +269,11 @@ P-B / C-B 只表示该内容需要承载品牌词，不代表品牌名、SKU、
 - 同帖型或相近帖型的 Reddit 内容
 - 已移除 / 被删帖样本中的风险信号
 
+> **⚠️ §3 是「样本学习理由的记录格式」，不是第二套位置来源（G2）。**
+> 下面五类**不得直接写入 `SelectedPlacement`**，也**不得绕过** §2.0 的机械取值规则
+> （众数档、并列取更晚、标题档 < 20% 不得选标题、既有 P-B/PGC 的取值域限制）。
+> 归类结论与 `SelectedPlacement` 冲突时，**以 §2.0 的取值为准**，并在 Brief 写明差异原因。
+
 每条 P-B 必须归入以下五类之一：
 
 ### 1. 样本支持前置露出
@@ -155,7 +308,9 @@ Brief 必须写明：
 
 - 中段露出
 - 后段露出
-- 配对评论露出
+- 配对评论露出 —— ⚠️ **仅当该行规划时已是 P-NB**（其配对自评为 C-B）。
+  **规划时是 P-B / PGC 的行不得选此项**（见 §2.0 取值域限制），只能后移到中段 / 后段，
+  否则 `BlockingGap` 回规划
 
 Brief 必须写明：
 
@@ -173,17 +328,19 @@ Brief 必须写明：
 - 板块规则或社区氛围反感推广
 - 样本显示问题讨论比产品推荐更容易存活
 
-允许结果：
+允许结果（⚠️ **区分当前行的既有类型，不得自动重分类**）：
 
-- P-NB 主帖
-- C-B 评论补位
-- 完全不露品牌
+- **本行规划时已是 P-NB** → 可采用 P-NB 主帖 + C-B 评论补位，或完全不露品牌；
+- **本行规划时是 P-B / PGC** → 样本只说明**该主帖的前置位置不被支持**，正确动作是把
+  `SelectedPlacement` 后移到**中段 / 后段**（**不含「评论」，也不含「不出现」**），
+  或 `BlockingGap` **回规划决定**；
+  **不得**在 Brief 阶段把 P-B 静默改成 P-NB，也不得自行改动配对与配额。
 
 Brief 必须写明：
 
-- 为什么主帖不露品牌
+- 为什么主帖不前置品牌
 - 是否由评论补位
-- 如果不补位，品牌配额如何在其他位置补偿
+- 若建议改变行类型或配额：写成**建议 + 回规划确认**，不是已执行的变更
 
 ---
 
@@ -207,27 +364,46 @@ Brief 必须写明：
 
 ---
 
-### 5. 样本不足或不可比
+### 5. 样本不足或不可比 → **StopAndAsk（不再套默认位置）**
 
 使用条件：
 
-- 样本太少
+- 样本太少（`ComparableSampleCount` 不足，或 `BrandMentionSampleCount = 0`）
 - 帖子被删除，看不到正文
 - 样本板块不同、帖型不同或语境不可比
 
-默认处理：
+> **v3 变更（G2）：旧版「样本不足时仍按内容类型给默认品牌位置」已删除。**
+> 那四条默认（Review/owner/price/comparison → 标题或首段；Guide/workflow → 中段；
+> Discussion/sensitive → 后段或评论；高风险 → 不露）正是「**位置由帖型标签推导**」，
+> 与 §2.0 的位置来源规则直接冲突。样本不足时套默认，会产出一个**看起来有依据、
+> 实际没有证据**的位置，而这正是「P-B 标题统一加品牌」那类返工的入口。
+> §2 A/B/C/D 仍然有效，但它们是**风险边界**，只能把位置往后推，**不能在没有证据时
+> 把位置往前提到标题或首段**。
 
-- Review / owner / price / product comparison：允许标题或首段露出
-- Guide / workflow / checklist：默认中段露出
-- Discussion / caution / sensitive scenario：默认后段或评论露出
-- 高风险板块：主帖不露品牌
+**样本不足时的正确动作（按序，不得跳到最后一步）：**
+
+1. **补样本**：扩大时间窗重采 / 放宽到 `FlairCluster` / 取相邻可比板块；
+2. **换来源**：改用板块级 SRP 的 `brandFirstMentionHistogram`
+   （`../../reddit-strategy-shared/references/subreddit-research-profile.md`；
+   该板块暂无可用 SRP 实例时以 baseline 卡为主来源，不得因其缺失而无条件阻断）；
+3. **停下来问人**：向策略师 / 用户说明「本行缺品牌位置证据，请裁决」，
+   给出可选项与各自风险；
+4. 仍不成立 → `BlockingGap: BrandPlacementUnsupported`，
+   `DecisionStatus: NoComparableEvidence`，本行不得进入写作。
+
+**唯二可以在缺证据时仍取标题位置的情况**：§2.5 的 `PGCApproved` 与 `BrandReviewPost`
+两类合法例外（须满足其全部成立条件并记 `BrandTitleExposureException`）。
 
 Brief 必须写明：
 
-- 样本不足原因
-- 使用了哪条默认规则
+- 样本不足原因与三个计数字段（`ComparableSampleCount` / `BrandMentionSampleCount` /
+  `NoBrandMentionRate`）**加**四档直方图（零分母时后两项填 `N/A`）
+- 走了上述哪一步（补样本 / 换来源 / 问人 / 例外 / BlockingGap）及其结论与批准人
 - 后续写作时需要特别注意什么
 
+⚠️ **不得因为「位置定不下来」就改动配额**：不得把 P-B 静默改成 P-NB，
+也不得删行或把品牌配额挪到别处凑数 —— 那是规划侧决策，须回规划确认。
+
 ---
 
 ## 4. PGC 账号前置验证规则
@@ -257,7 +433,9 @@ Brief 必须写明：
 - 有竞品 PGC 成功样本：该板块可进入 PGC 候选，但仍需符合 flair、披露、频率、账号历史等条件。
 - 有竞品 PGC 但集中在 weekly thread / AMA / vendor flair：只能按该路径发，不能普通发帖。
 - 有品牌名但都是普通用户 UGC：不能直接推导 PGC 可发，只能说明品牌话题可讨论。
-- 没有竞品 PGC 样本：默认按 UGC / P-NB / C-B 保守处理。
+- 没有竞品 PGC 样本：**该板块的 PGC 候选**默认按 UGC / P-NB / C-B 保守处理。
+  ⚠️ 这条只适用于**尚未定为 PGC 的候选**；**已进入执行表的 PGC 行不得据此自动改成
+  P-NB / C-B** —— 那是配额与行类型变更，须回规划确认。
 - 明确禁止 self-promotion / commercial post / company account：即使竞品出现过，也不能直接判定可用，必须标高风险或排除。
 
 ### PGC 可用时的品牌 / SKU 露出
@@ -279,13 +457,15 @@ Brief 必须写明：
 
 每条 P-B 主帖 Brief 必须包含：
 
-- 品牌 / SKU 露出位置：标题 / 第一段 / 中段 / 后段 / 评论
+- 品牌 / SKU 露出位置（`SelectedPlacement`）：标题 / 第一段 / 中段 / 后段
+  —— ⚠️ **P-B / PGC 主帖不含「评论」与「不出现」**（见 §2.0 取值域限制）；
+  确需改到评论 = 行类型/配额变更，须回规划并重跑门禁
 - 露出位置理由
 - subreddit 规则依据
 - 帖型与内容意图依据
 - 样本学习对齐结果
 - 是否存在 GEO 与 Reddit 原生感冲突
-- 如果不前置，具体在什么结构位置或哪条评论露出
+- 如果不前置，具体在正文的什么结构位置露出（P-B / PGC 主帖不得答「放到配对评论里」）
 
 每条 PGC 主帖 Brief 还必须包含：
 
@@ -305,13 +485,20 @@ Brief 必须写明：
 
 - [ ] 没有把所有 P-B 批量设置为标题或第一段露出
 - [ ] 每条 P-B 都单独判断了品牌 / SKU 露出位置
-- [ ] Review / owner / price / comparison 才默认允许标题或首段露出
-- [ ] Guide / workflow / checklist 默认中段露出
+- [ ] **不存在「因为这行标记 P-B，所以标题写品牌」的推导**（G2 硬拦截）
+- [ ] 每条 **P-B / PGC 主帖**均已填满 `BrandPlacementBasis`，`SelectedPlacement` = 更晚的（§2 风险边界允许的最早位置, 直方图经验位置），且对既有 P-B/PGC 行未取「评论」/「不出现」
+- [ ] 每条 **C-B 评论**填 `InheritedBrandPlacementBasis`（继承配对 P Key；配对为 P-NB 时填 `N/A(PairedPostIsP-NB)`；外部评论填 `N/A(ExternalComment)`）+ `CommentPlacementDecision`；**评论不要求自有 Basis、不单独建直方图**
+- [ ] **P-B / PGC 主帖**的三个计数字段齐全（`ComparableSampleCount` / `BrandMentionSampleCount` / `NoBrandMentionRate`）**且**四档 `BrandFirstMentionHistogram`（标题/首段/中段/结论）已给出（零分母填 `N/A`），`ComparableScopeStatus` 与 `DecisionStatus` 已判定 —— **C-B 评论不适用本项**（无自有直方图）
+- [ ] 直方图档位已按「结论 → 后段」映射换算成 `SelectedPlacement` 受控值
+- [ ] 既有 P-B / PGC 行未被 D 类或高 `NoBrandMentionRate` 静默推成「不出现」/「评论」（该情况须 BlockingGap 回规划）
+- [ ] `BrandMentionSampleCount = 0` 的行未被当作「不允许标题」，而是走了 §3.5 的补样本 / 换来源 / 问人 / 例外 / BlockingGap
+- [ ] 选标题位置的行：直方图「标题」档 ≥ 20%，或持有 `BrandTitleExposureException`（`PGCApproved` 须有真实先例，`BrandReviewPost` 须确为专门讨论该品牌的帖）
+- [ ] §2 A/B/C/D 仅用作风险边界（只能把位置往后推），未被当作位置来源
 - [ ] 中段露出没有被机械理解为固定第 2-4 段
-- [ ] Discussion / caution / sensitive scenario 默认后段或评论露出
-- [ ] 高风险板块没有主帖前置品牌
+- [ ] 高风险（D 类）板块没有主帖前置品牌，例外未越过 D 类与披露要求
 - [ ] 样本学习结果已经写入每条 P-B 的 Brief
 - [ ] 若样本与 GEO 需求冲突，已说明最终选择理由
+- [ ] 本文件的判定未改动 P-B / P-NB / C-B / C-NB / PGC 配额语义（位置不成立时走 BlockingGap 或回规划，不静默改配额或改行类型）
 - [ ] 如果使用 PGC，已经完成 PGC account precedent check
 - [ ] PGC 帖已在执行表标注为 PGC，且帖子品牌词为有
 - [ ] PGC 发帖板块已确认允许品牌 / vendor / official account 路径