lilialla · lilialla · Jun 17, 2026
diff --git a/skills/request-basis/SKILL.md b/skills/request-basis/SKILL.md
@@ -15,12 +15,16 @@ This skill routes PRC civil request-right basis work. It is not a shortcut to a
 4. For each request basis analyzed in detail, apply the request lifecycle: right arises, not extinguished, exercisable, defenses/counter-defenses.
 5. Use controlled concept cards for professional terms when available; create a temporary concept card when absent.
 6. Use rule-obligation groups as scaffolding, not as verified law.
-7. If no local rule-obligation group fits, generate a temporary candidate group and label it `model_inference`.
+7. If no local rule-obligation group fits, generate a temporary candidate group and label it `model_inference`; when neither a rule-obligation group nor a request-basis seed matches, build the schema from the main norm via `reference.generation_protocol`.
 8. Search local materials only for targeted snippets.
 9. For new materials, extract atomic source assertions before merging them into concept cards or rule-obligation groups.
 10. For reference integration work, use source synthesis to group sources by request-basis family before changing runtime rules.
 11. Output uncertainty, verification needs, and excluded paths explicitly.
 12. For real case pilots, use the real-case-testing reference and keep reusable lessons abstract.
+13. In litigation/exam work, hold the fact–question–norm valve and two user-confirmation gates (candidate list, lifecycle outline) before drafting: `reference.staged_gates`.
+14. When an element reaches an indeterminate concept (过错 / 合理 / 显失公平 / 善意), route it through `reference.interpretation_toolkit` instead of asserting it.
+15. Match argument depth to how contested each element is: `reference.proportionality`.
+16. Before drafting or converting to a report, pass the output pre-check gate: `reference.output_precheck`.
 
 ## Scenario Router
 
@@ -72,6 +76,7 @@ Keep schema details, resource manifest mechanics, generated indexes, candidate b
 - Do not suppress a plausible path just because it is absent from seed data. Keep it as a labeled candidate if trigger facts, norm function, evidence target, and verification need can be explained.
 - Do not store real case facts as skill examples; convert only abstract, repeatable failure patterns into evals or references.
 - Without MCP or another authoritative source, do not state current law as verified.
+- Resolve concrete tools through `reference.capability_slots` ([SLOT-LEGAL-DB], [SLOT-DOC-PARSE], [SLOT-WEB-SEARCH], [SLOT-WORD-OUTPUT]); when a slot is unset, degrade through the source labels rather than guessing.
 - Without case-law research, do not say courts usually or generally hold a position.
 - Do not treat party statements as found facts.
 - Every analyzed request basis must visibly pass through: right arises, not extinguished, exercisable, defenses/counter-defenses.
@@ -111,6 +116,12 @@ Read only what is needed, resolved by logical role:
 - `reference.retrieval`: searching local source materials without loading large files.
 - `reference.methodology`: request-basis method, three layers/four steps, practice-vs-study distinction.
 - `reference.output_templates`: reusable output skeletons.
+- `reference.generation_protocol`: build a review schema from norms when no seed group or base matches (查字典→会造句 fallback).
+- `reference.staged_gates`: fact–question–norm valve, two user-confirmation gates, and parallel-drafting premise injection.
+- `reference.proportionality`: match argument depth to how contested each element is.
+- `reference.interpretation_toolkit`: indeterminate-concept hook routing to interpretation ranks and the legal-interpretation-argument / analogical-reasoning skills.
+- `reference.output_precheck`: pre-drafting gate for coverage, citation verification, delivery-format conformance, and one-pass revision.
+- `reference.capability_slots`: tool-to-slot mapping for portability; degradation routes through the source labels.
 
 ## Data Files
 

diff --git a/skills/request-basis/references/capability_slots.md b/skills/request-basis/references/capability_slots.md
@@ -0,0 +1,45 @@
+# Capability Slots
+
+Use this reference to keep the request-basis method portable across tool stacks. The skill describes *what* capability is needed at each step (a slot); the slot table maps each slot to the concrete tool in the current environment. Methodology files refer to slots by `[SLOT-X]` placeholders, never to a tool name. To move the skill to another stack, change only the slot table below — not the method.
+
+## Why slots
+
+- Portability: methodology, source labels, and lifecycle gates stay stable; swapping `mineru-ocr` for another parser is a one-line table edit.
+- Honest degradation: a missing slot does not silently fail. It downgrades through this skill's own discipline (five-tier source label + verification task), not through ad-hoc inline text.
+- Single source of tool truth: do not hardcode tool names inside other references; resolve them here.
+
+## Prefilled Default Slot Table
+
+This environment is a single fixed tool stack, so the slots are prefilled — no cold-start interview is run. The defaults below apply directly.
+
+| Slot | Capability | Default tool | Fallback |
+| --- | --- | --- | --- |
+| `[SLOT-LEGAL-DB]` | Authoritative law / judicial interpretation / case-law lookup | `yuandian-law-search` | `regulation-search`; 北大法宝 |
+| `[SLOT-DOC-PARSE]` | Convert PDF/Word/image case files to Markdown | `mineru-ocr` | `case-files-to-md-fast` (sensitive / no-upload / offline) |
+| `[SLOT-WEB-SEARCH]` | Open-web search for practice articles, context, leads | built-in WebSearch | — |
+| `[SLOT-WORD-OUTPUT]` | Render a finished report to .docx | `md2word` (legal report uses `rules/common/legal-report-delivery-format.md`, 《法学研究》-style) | output Markdown + note manual conversion |
+| `[SLOT-ACADEMIC]` | Academic paper retrieval | (optional, unset) | route the need through `[SLOT-WEB-SEARCH]` |
+
+Routing notes:
+
+- `[SLOT-LEGAL-DB]` is the verification engine. It is what closes a `verification.md` task from `pending_mcp` to `mcp_verified`. Without it, no statute/interpretation is `verified_law`.
+- `[SLOT-DOC-PARSE]`: default to `mineru-ocr`; when material is confidential or must not be uploaded, use `case-files-to-md-fast` instead.
+- `[SLOT-WORD-OUTPUT]` is only invoked at final delivery; analysis stays in Markdown until then.
+
+## Degradation Rules
+
+When a slot is unset or its tool is unavailable, degrade through this skill's existing machinery — do **not** introduce inline text markers such as `[未经校验]`. The five-tier label schema (`source_policy.md`) is stronger and already covers this.
+
+- `[SLOT-LEGAL-DB]` missing: do not state any current-law claim as `verified_law`. Keep cited statutes/interpretations at `judicial_interpretation` / `source_extracted` with `verification_status: unverified` (or `pending_mcp` once queued), and open a `verification.md` task per `mcp_verification_required` claim. This is the existing Global Rule "Without MCP… do not state current law as verified", expressed through the slot.
+- `[SLOT-DOC-PARSE]` missing: work only from text the user supplied; mark any reconstructed table/figure claim `ocr_risk: true` and label it `source_extracted` / `needs_human_review`. Do not invent case-file content.
+- `[SLOT-WEB-SEARCH]` missing: do not assert court-practice generalizations; keep them `model_inference` and queue a `case_law_reference` verification task.
+- `[SLOT-WORD-OUTPUT]` missing: deliver Markdown and note that conversion is pending; this is a delivery-format gap, not a substance gap, so no label change.
+- `[SLOT-ACADEMIC]` unset: fold the need into `[SLOT-WEB-SEARCH]`; scholarly propositions stay `scholarly_reference` / `academic_paper` and unverified until checked.
+
+In every case the degraded claim still passes through the four request-lifecycle gates and carries an explicit source label and, where verification is required, a verification task. Degradation narrows confidence; it never bypasses the labels.
+
+## Portability Note
+
+When distributing this skill to another environment, clear the prefilled table back to empty slots and either run a first-use slot interview (ask the user which tool fills each slot) or apply the degradation rules above until slots are registered. The methodology and label schema do not change; only the slot table does.
+
+> 能力槽思想借鉴 Youchu-lawhub/gutachten-civil-case（Apache-2.0）；本 skill 改为单栈预填、沿用自有五档证据标签。
diff --git a/skills/request-basis/references/classification.md b/skills/request-basis/references/classification.md
@@ -38,3 +38,52 @@ Only after that, ask:
 ```text
 Which candidate causes of action help retrieve materials, draft pleadings, or compare cases?
 ```
+
+## Five-Level Norm Retrieval
+
+When finding the candidate norm from raw facts, walk five levels (this is the
+"找法" front end to `reference.generation_protocol`):
+
+1. **Party relations** — build the party-pair matrix (A→B, B→A, A→C …); each
+   pair is a separate analysis unit.
+2. **Relief sought** — what each claimant wants (给付 / 确认 / 形成 / 非讼).
+3. **Norm class** — place the relief in one of the norm-basis classes below.
+4. **Specific norm** — the exact 条文 inside that class (verify via `[SLOT-LEGAL-DB]`).
+5. **Review schema** — build the four-gate 检视程式 for that norm
+   (`reference.request_lifecycle` / `reference.generation_protocol`).
+
+## Norm-Basis Classes (taxonomy)
+
+A working class list for level 3. Treat it as a retrieval index, not a closed set;
+add a class if the facts need one.
+
+| # | 类别 | 典型请求/规范 |
+| --- | --- | --- |
+| 1 | 合同 | 履行 / 违约 / 解除返还 |
+| 2 | 准合同（无因管理·不当得利） | 返还 / 费用偿还 |
+| 3 | 物权请求权 | 返还原物 / 排除妨害 / 消除危险 |
+| 4 | 占有保护 | 占有返还 / 妨害除去 |
+| 5 | 人格权 | 停止侵害 / 赔礼道歉 / 损害赔偿 |
+| 6 | 知识产权 | 停止侵害 / 赔偿 |
+| 7 | 侵权 | 损害赔偿（含人身/财产/精神） |
+| 8 | 缔约过失 | 信赖利益赔偿 |
+| 9 | 后合同义务 | 协助 / 保密 / 损害赔偿 |
+| 10 | 婚姻家庭 | 抚养 / 分割 / 损害赔偿 |
+| 11 | 继承 | 继承权确认 / 遗产分割 |
+| 12 | 债的保全 | 代位权 / 撤销权 |
+| 13 | 债权债务移转 | 债权转让 / 债务承担 |
+| 14 | 担保物权 | 优先受偿 / 实现担保物权 |
+| 15 | 确认之诉 | 确认权利/法律关系存在或不存在 |
+| 16 | 形成之诉 | 撤销 / 解除 / 变更 |
+| 17 | 非讼 | 宣告失踪/死亡 / 指定监护 等 |
+| 18 | 程序性 | 管辖 / 保全 / 执行 等（非请求权基础，单列） |
+
+Class #2 bundles 无因管理 and 不当得利 as a retrieval index only; in the outer
+examination order they stay distinct — 无因管理 is checked before 物权, and
+不当得利 is the residual/兜底 basis checked last (`reference.methodology`).
+
+Terminology discipline: 请求权基础 is for 给付之诉; 确认 / 形成 / 非讼 / 程序 rest on
+a 规范基础 that is not a 请求权基础 — keep the words distinct
+(`reference.request_basis_concept`).
+
+> 方法借鉴 Youchu-lawhub/gutachten-civil-case（Apache-2.0），已按本 skill 的双轨入口与术语纪律改写。
diff --git a/skills/request-basis/references/generation_protocol.md b/skills/request-basis/references/generation_protocol.md
@@ -0,0 +1,83 @@
+# Generation Protocol
+
+Use this reference only as a fallback. When `rule_obligation_seed` and `request_basis_seed` both miss — no local rule-obligation group fits the claim and no seed request basis matches — this protocol lets the skill build a review schema (检视程式) from the norm itself instead of stopping.
+
+This is the move from 查字典 (look up a prebuilt schema) to 造句 (construct one from first principles). It does not produce verified law.
+
+## Activation Gate
+
+Run this protocol only when all are true:
+
+- The scenario and claim goal are already fixed (see `reference.methodology`).
+- No `rule_obligation_seed` group bridges the claim to bases/elements/defenses.
+- No `request_basis_seed` entry names a matching base.
+
+If a seed exists, prefer it and use this protocol only to cross-check coverage.
+
+## Output Discipline (non-negotiable)
+
+Every node this protocol generates is a method scaffold, not located law:
+
+```yaml
+source_status: model_inference
+verification_status: pending_mcp
+mcp_verification_required: true
+```
+
+- Carry these labels through to the final output; do not present a generated base, element, or article number as settled law (`reference.source_policy`, `reference.discretion_policy`).
+- For each generated main norm and each defense/auxiliary citation, write a verification task into the queue (`reference.verification`): exact article-number check, current-version check, and any 配套司法解释.
+- A generated schema is a candidate path, never a "located request basis" until the queue clears.
+
+## Step 1 — Find Candidate Main Norms
+
+From the fixed claim goal, locate candidate main norms (请求权基础条文) by structured discovery, not by guessing articles.
+
+1. **Party-pair matrix.** List every party and its legal role (natural person, legal person, consumer/operator, possessor/物权人, guarantor, etc.). Pair them both directions: A→B (forward claim), B→A (defense/counterclaim), and any third-party leg. Keep only pairs joined by a connection point: contract, property/possession, tort, unjust enrichment, negotiorum gestio, or status relationship.
+2. **Claim type.** For each live pair, classify the claim: 给付之诉 / 确认之诉 / 形成之诉. Inside 给付, separate 原给付 / 次给付 (违约) / 物权请求权 / 损害赔偿 / 返还 (不当得利) / 费用偿还.
+3. **Norm class then article.** Walk the typical outer order (`reference.methodology`): contract → contract-like → negotiorum gestio → property/possession → unjust enrichment → tort. For each plausible class, pull candidate articles. Apply the priorities: 有合同先检合同; 物权优先于债权; 不当得利兜底; flag 竞合 and 聚合.
+
+Output: a candidate list keyed by party-pair and claim, each entry `model_inference`.
+
+## Step 2 — Classify Norm Function
+
+For each candidate, judge its function before building anything (`reference.norm_decomposition` vocabulary):
+
+- `main_norm` / `incomplete_main_norm`: directly supports the claim (incomplete if a legal act or another norm must fill the effect). Only these seed a schema spine.
+- `auxiliary_norm`: defines an element, status, scope, time point, or consequence of a main norm.
+- `defense_norm`: blocks arising, extinguishes, or blocks exercisability.
+- cross-class signals (`counter_defense_norm`, `referral_norm`, `principle_norm`, `burden_norm`): note but do not promote to spine.
+
+Demote anything that only "looks related." Keep only genuine main norms as Step 3 anchors; hold the rest for mounting.
+
+## Step 3 — Build the Lifecycle Schema
+
+Take each confirmed main norm and raise the four-gate review spine (`reference.request_lifecycle`). Do not match an external template library; build the spine directly from this skill's lifecycle.
+
+For the main norm:
+
+1. Load the article, isolate the legal effect ("……的，应当/有权/可以……"), and treat everything before it as positive elements.
+2. Lay the elements under `产生` (right arises). If the base rests on a legal act, do not collapse the gate — identify the act model (unilateral/bilateral/resolution/...) and select formation + effectiveness facts (`reference.request_lifecycle` legal-act gate).
+3. Mount the gates and hang the held norms from Step 2:
+   - `产生`: positive elements + auxiliary norms + 权利障碍抗辩 (无效/可撤销/效力待定; tort: 阻却违法/无过错).
+   - `未消灭`: 清偿/抵销/提存/混同/免除/解除/撤销 consequences (anchor on 债的通用消灭事由).
+   - `可行使`: 诉讼时效 + 同时/先履行/不安抗辩 + 约定条件 + 程序前提.
+   - `抗辩/反抗辩`: counter-defenses, waiver, 时效中断/中止, defense-unavailable.
+4. For 损害赔偿 bases, split 责任成立 from 责任范围 inside `产生`.
+
+Mark each defense/auxiliary with its burden holder. Tag branches the facts do not touch as `[本案不涉及]` and fold them — do not delete, to keep the spine complete and auditable.
+
+## Step 4 — Decompose Elements and Flag Open Concepts
+
+Break each gate's elements into smallest 涵摄 units and hand the soft ones to the concept layer (`reference.concept_cards`).
+
+- Split on connectors: 和/以及/同时 → AND (each independent); 或者/或 → OR (alternative); 但……除外 → carve the 但书 into a defense norm.
+- Surface implied elements the text omits: 因果关系, 过错, 违法性, 行为能力.
+- Tag each element type (主体/行为/客体/主观/因果/结果/程序) and certainty (确定 / 不确定法律概念 / 一般条款).
+- For every 不确定法律概念 (重大、合理、显失公平、善意、根本违约、不可抗力, ...): mark **需法律解释**, name the dispute focus, and route to a concept card — create a temporary card when none exists (`reference.concept_cards`). Do not let the term float on model intuition.
+- Apply the 否认 vs 抗辩 distinction: a norm denying an element fact sits under that element marked 否认; a norm introducing an independent blocking/extinguishing fact stands alone marked 抗辩. Note any 跨阶段双重挂载 (one norm serving two gates).
+
+## Handoff
+
+The product is a candidate schema: four-gate spine + decomposed elements + concept-card hooks + a populated verification queue, all `model_inference` / `pending_mcp`. It feeds the same gates and output skeletons as a seed-sourced base (`reference.output_templates`), and must clear `reference.verification` before any node is stated as law. If a real seed later appears, compare the two and reconcile.
+
+> 方法借鉴 Youchu-lawhub/gutachten-civil-case（Apache-2.0），已按本 skill 的四阶层 lifecycle 与证据标签体系改写。