A source-grounded, moderated roundtable across the world's major religious and philosophical traditions — as a drop-in skill for AI coding agents.
以原典為據、由中立主持人調度的「跨傳統哲學議會」——可直接掛載到 AI coding agent 的 skill。
English · 繁體中文
Religion Council turns your AI agent into a neutral moderator that convenes a fair, source-grounded dialogue among religious and philosophical traditions — without flattening them into one mushy consensus.
Ask "does life have meaning?" and you hear Buddhism, Taoism, and Confucianism each argue from its own texts. Every claim is tagged as either a [Text] quotation (with a real locator) or an [Interpretation], and the moderator surfaces the genuine tensions instead of forcing agreement.
Version v0.12.6 supports three execution modes:
- Claude Code only — 37 specialized Claude agents (1 moderator + 36 voices).
- Codex only — a portable Codex skill, with native Codex subagents when requested.
- Claude moderator + Codex panelists — a deterministic Python MCP controller manages persistent Codex threads, barriers, retries, and audit records.
All three modes share one quote-admissibility policy (quote-admissibility/v2). In the
hybrid controller, v0.5.0 added opt-in B1b structured claims: debate_start can accept a
retrieval evidence_envelope, ask panelists for a religion-council/claim/v1 payload,
schema-check it, repair or drop malformed payloads, and bind valid claims to B1a evidence
seeds. Version v0.6.0 adds opt-in B2 validation for that same structured path: with
verify_claims=true, quotation claims are checked against curated evidence snapshots and
stored with per-claim runtime-validated or failed outcomes. Version v0.7.0 adds opt-in
B3 fail-closed boundary enforcement with fail_closed=true: only affirmatively admitted
claims may pass the response boundary, while unknown claim types, unverified [Text],
renderer bypasses, and unsupported protocols are default-denied. Claude-only and portable
modes stay instruction-enforced. See the
assurance matrix and
ADR 0002.
Version v0.8.0 added the A1 corpus-enrichment metadata foundation: retrieve.py can merge
curated representation_kind, rendering_mode, provenance, and rights from a portable
presentation.json sidecar, with seed curation for the existing Chinese Qur'an
meaning-renderings. Bulk public-domain excerpt expansion remains source- and rights-reviewed
content work.
v0.9.0 adds the strict-finalization capstone for the hybrid controller. profile="strict"
requires the structured → verify → fail-closed graph and an evidence envelope, then makes
finalization a separate, explicit step. See ADR 0004
and the v0.9.0 changelog.
v0.11.0 adds stable occurrence identity and a rights-scoped corpus baseline as A2/A3 readiness only. It does not add an index, vector store, or RAG backend; that work remains gated on the retriever-fork contract and benchmark decisions.
v0.12.0 establishes the retriever fork for that next stage: the portable retriever remains stdlib-only and file-based, while the project retriever has the same versioned retrieval envelope and shared conformance suite. It does not select an index, vector store, RAG backend, or edition-backed assurance; those remain gated on benchmark evidence. See ADR 0006.
v0.12.1 defines that benchmark and its adoption gates: a candidate must both beat the lexical baseline and preserve stable occurrence identity, the retrieval contract, artifact lifecycle, rights boundaries, and assurance honesty. See Retrieval Benchmark v1.
v0.12.2 runs the first retrieval-v1 lexical baseline. The benchmark now has a frozen 18-query fixture set, graded relevance judgments, a deterministic offline runner, and committed JSON/Markdown baseline reports. The baseline identifies two main weaknesses of the current lexical retriever: broad thematic recall and no-answer discrimination. No backend is selected.
v0.12.3 evaluates an experiment-only lexical confidence threshold and adds GitHub issue templates for public feedback intake. Thresholds 2 and 3 eliminate the benchmark's no-answer false-support cases without answerable-query regression, while threshold 5 regresses q007 and q010. No threshold or backend is adopted yet.
v0.12.4 evaluates an experiment-only BM25-style lexical ranking candidate. BM25 improves some ranking metrics, including MRR and nDCG@5, while preserving exact-span hit rate, but it does not improve no-answer discrimination or broad thematic recall. No backend is selected and default retrieval remains unchanged.
v0.12.5 evaluates an experiment-only BM25 + lexical confidence threshold candidate. BM25 supplies ranking gains while the threshold supplies no-answer discrimination: MRR improves from 0.938 to 0.969, nDCG@5 from 0.902 to 0.919, exact-span hit remains 1.000, no-answer correctness reaches 1.000, and false-support falls to 0.000. Broad thematic recall remains weak on q010, so no backend is selected and default retrieval remains unchanged.
v0.12.6 publishes ADR 0007, the retrieval backend decision from the retrieval-v1 evidence. It selects BM25 + lexical-confidence threshold (t2/t3) as the project retriever's ranking + no-answer policy, while gating the default BM25 ranking flip on the ≥2-judge + κ step. It still selects no RAG, vector, hybrid, built-index, or network backend, and default retrieval remains unchanged until a follow-up implementation.
For a strict run, the lifecycle is:
debate_start(profile="strict", evidence_envelope=…)
→ collect
→ finalization_required=true / finalized=false
→ debate_finalize
→ finalized=true
debate_finalize produces two deliberately different answer surfaces:
- Surface A — textual authority.
AuthorityRenderUnitvalues are minted only by the canonical builder, independently trace-validated, then deterministically serialized. Quotation text comes from the canonical snapshot span; representation metadata is system-authoritative. - Surface B — interpretation. Panelist analysis and inference are framed as non-authoritative interpretation. Rejected claims are separated into audit input and are never passed to the ordinary answer render input.
Strict-finalized responses provide end-to-end machine-enforced construction and traceability of the textual-authority surface (Surface A). Interpretation prose (Surface B) remains explicitly non-authoritative and instruction-bounded.
This does not mean every word is machine-verified. Interpretation prose can still be
misleading; the complete answer is not semantically fail-closed; the default hybrid path without
debate_finalize is unchanged and not finalized; and the capability-token mint guard is not a
Python sandbox. Strict mode is opt-in and does not change the default hybrid workflow.
Run the offline deterministic example from the repository root:
python3 examples/strict-finalization/run_example.pyExpected lifecycle:
debate_start(profile="strict", evidence_envelope=…)
→ debate_collect
→ finalization_required=true / finalized=false
→ debate_finalize
→ finalized=true
This example demonstrates the strict finalization lifecycle and the separation between the textual-authority surface (Surface A) and the interpretation surface (Surface B). Strict mode is opt-in and does not change the default workflow.
Run the offline, deterministic strict end-to-end example to inspect the complete path and its assertions.
- Grounding discipline comes first.
[Text]vs[Interpretation]labels on every line; no invented chapter / verse / sūtra / hadith locators; the Chinese Qur'an is always marked as a rendering of meaning, never the Arabic original; the skill never claims to channel a prophet, deity, or thinker. - Three zoom levels. Compare whole traditions, branch into denominations / schools, or stage historical-thinker debates.
- Neutral moderation, no forced winner. Set the question → opening positions → cross-examination → synthesis. It separates shared concerns from irreducible differences and never crowns a "winner" unless you give an explicit criterion.
- Claim-level pressure, not performative hostility. Openings commit to a thesis; rebuttals target a concrete claim, premise, or counterexample and return a cross-examination question. Practical overlap is not mislabeled as consensus.
- A real secular-liberal voice, with a safe foil as fallback. Religious rosters lean one
way, so the council ships
council-secular-humanistandcouncil-mill(grounded in On Liberty / Utilitarianism) as full panelists that can actually rebut. Only when a roster still leans one way does the moderator fall back to a controller-routed contrast proposition — debate framing only: not source evidence, not a participant claim, never an instruction to execute. - Three execution modes, one corpus. Claude-only, Codex-only, or Claude moderating persistent Codex MCP panelists, all using the same curated references.
- Built for a RAG future. Retrieval lives behind one stable contract
(
scripts/retrieve.py), so the corpus can grow from curated snippets to a full, vector-backed 典籍 store without touching the personas. See the Roadmap.
| Level | Members |
|---|---|
| Traditions (8) | Christianity · Islam · Hinduism · Buddhism · Taoism · Legalism* · Confucianism* · Mohism* |
| Denominations | Christianity: Catholic · Orthodox · Protestant — Islam: Sunni · Shia |
| Thinkers | Christianity: Jesus · Augustine · Aquinas · Luther · Calvin — Islam: Muhammad · al-Ghazali · Ibn Rushd — Buddhism: Shakyamuni · Nāgārjuna · Vasubandhu · Pure Land — Hinduism: Krishna · Shankara · Rāmānuja · Madhva — Pre-Qin China: Confucius · Mencius · Xunzi · Laozi · Zhuangzi |
| Non-religious | Secular Humanism · J. S. Mill (liberty / utilitarianism) — philosophical / ethical stances, not religions; outside the default eight, added on request |
* Legalism, Confucianism, and Mohism are philosophical / intellectual traditions, not religions, and the skill labels them as such (and distinguishes philosophical from religious Taoism where it matters).
1. Claude Code only (uses the 37 custom agents):
git clone https://github.com/F-e-u-e-r/religion-council.git
cd religion-council
claude # open Claude Code in the repo rootThen, in the conversation:
用議會討論:人生有沒有意義?請佛教、道教、儒家三家先各自陳述,再交叉辯論。
The main agent hands off to council-moderator, which frames the question, dispatches
each council-* voice, relays their arguments to one another, and synthesizes.
2. Codex only (portable, self-contained):
Ask Codex to install skills/religion-council/ from this repository, restart Codex, then invoke:
Use $religion-council to convene a sourced roundtable on whether life has meaning.
skills/religion-council/SKILL.md is fully self-contained and degrades gracefully to a
single-context dialogue when sub-agents aren't available.
3. Claude moderator + Codex panelists:
Authenticate Codex, start Claude Code at the repo root, approve the project MCP server in
/mcp, then ask Claude to use religion-council-controller with
orchestrator/panelists/religion-8.json. The controller preserves each Codex threadId
across rounds and refuses to advance before the current round reaches its barrier.
→ Full per-platform setup: INSTALL.md
# Cross-tradition
Convene a council: where does meaning come from? Buddhism, Taoism, and Confucianism,
opening statements then cross-examination.
# Denomination debate (Level 1)
How do Catholic, Orthodox, and Protestant views on salvation differ?
遜尼與什葉對「先知之後誰該領導」的分歧是什麼?
# Thinker debate (Level 2)
儒家內部對人性的看法一致嗎?請孔子、孟子、荀子各自說明,再交叉辯論。
Nāgārjuna vs Vasubandhu on emptiness vs consciousness.
Self-power or other-power for liberation? Pure Land vs Zen / Theravāda.
al-Ghazali vs Ibn Rushd on faith and reason.
# Secular vs religious (non-religious tier)
Must life's meaning come from religion or an afterlife? Secular humanism vs Christianity and Buddhism.
在不傷害他人、不違法並承擔後果下,個人是否仍有義務按宗教或傳統美德生活?請彌爾與儒家、天主教辯論。
┌──────────────────────────────────────────────────────────────┐
│ SKILL.md — the operating manual │
│ question layers · grounding rules · anti-quote-mining · │
│ roundtable flow (set → open → cross-examine → synthesize) │
└──────────────────────────────────────────────────────────────┘
│ every voice reads it before speaking
▼
┌──────────────────────────────────────────────────────────────┐
│ references/<tradition>.md — the personas │
│ voice & tone · core concepts · quotable snippets WITH │
│ citations · cross-tradition tension points │
└──────────────────────────────────────────────────────────────┘
│ Claude Code only
▼
┌──────────────────────────────────────────────────────────────┐
│ council-moderator ──dispatches──▶ council-* sub-agents │
│ (sub-agents can't talk directly, so the moderator relays) │
└──────────────────────────────────────────────────────────────┘
│ optional hybrid mode
▼
┌──────────────────────────────────────────────────────────────┐
│ debate_controller.py ──MCP──▶ codex mcp-server │
│ persistent thread IDs · barriers · retries · audit records │
└──────────────────────────────────────────────────────────────┘
│ retrieval seam (stable contract)
▼
┌──────────────────────────────────────────────────────────────┐
│ scripts/retrieve.py → { text, tradition, school, work, │
│ v0.1: lexical references locator, language, version, │
│ later: vector search category } │
└──────────────────────────────────────────────────────────────┘
The personas, moderation, and citation rules sit above the retrieval seam and are already stable. Everything below it — curated snippets today, a vector store tomorrow — can evolve without rewriting a single voice.
religion/
├── README.md # you are here
├── INSTALL.md # per-platform setup (Codex / Claude Code / other)
├── CONTRIBUTING.md # add a tradition / denomination / thinker
├── DISCLAIMER.md # sourcing rules + religious-sensitivity statement
├── LICENSE # MIT — skill logic, agents, scripts, config
├── LICENSE-CONTENT # CC BY 4.0 — references & corpus
├── VERSION # current release: v0.12.6
├── .mcp.json # Claude → deterministic Codex controller
│
├── skills/religion-council/ # ▸ PORTABLE skill (Codex & any agent)
│ ├── SKILL.md # English operating manual (self-contained)
│ ├── references/ # 16 persona files (snippets + citations)
│ ├── scripts/retrieve.py # dependency-free lexical retrieval
│ └── agents/openai.yaml # Codex interface metadata
│
├── .claude/ # ▸ CLAUDE CODE distribution
│ ├── agents/council-*.md # 37 sub-agents (1 moderator + 36 voices)
│ └── skills/religion-council/
│ ├── SKILL.md # 繁中 operating manual
│ ├── USAGE.md # how to convene a council
│ ├── references/ # 16 persona files (+ 延伸語料 corpus pointers)
│ └── scripts/retrieve.py # lexical retrieval (stable {text+metadata} contract)
│
├── orchestrator/ # ▸ CLAUDE MODERATOR + CODEX PANELISTS
│ ├── debate_controller.py # MCP server, barriers, retries, persistence
│ ├── render_types.py # Surface A/B/audit render types and mint guard
│ ├── render_finalizer.py # strict finalizer + trace validator + serializer
│ └── panelists/ # 8-member and 30-member example rosters
├── examples/strict-finalization/ # offline strict workflow fixture and runnable example
├── scripts/smoke_codex_mcp.py # opt-in authenticated create/reply MCP check
├── tests/ # controller persistence and pagination tests
├── docs/CORPUS.md # corpus & RAG-roadmap notes
├── docs/ORCHESTRATION.md # controller architecture and limitations
└── 01-基督宗教/ … 08-墨家/ # 典籍清單.md + 思想概要.md (corpus seed)
The two distributions are intentionally different:
skills/is dependency-free and portable, while.claude/is project-integrated (its references point back into the01–08corpus folders and it orchestrates real sub-agents).
The project develops along two low-coupling axes that meet at one versioned seam. Architecture stages are A0–A3 (corpus & retrieval) and B0–B3 (citation admissibility & enforcement). PR numbers are delivery history, not stage names. Normative detail: ADR 0002 (nomenclature + enforcement ladder) and ADR 0003 (the retrieval→evidence adapter).
Axis A — corpus & retrieval. Pivots on one seam: the retrieval envelope contract. Keep it stable and the 36 personas never change as the backend grows from files to a vector service.
| Stage | What | Retrieval |
|---|---|---|
| A0 — Curated council | Voices quote hand-picked, cited snippets in references/. Offline, any agent. |
File parse + lexical ranking. |
| A1 — Deeper corpus (metadata foundation, v0.8.0) | Adds the presentation.json provenance/rights sidecar and seeds existing Qur'an meaning-renderings; broader public-domain excerpt expansion remains rights-reviewed curation work. |
Still file-based. |
| A2 — Full 典籍 + local index | Store complete open scriptures in-repo, chunked; benchmark lexical / cross-lingual / dense / hybrid; build the chosen index. | Local index — same envelope contract. |
| A3 — RAG server | Index behind a retrieval service; retrieve.py becomes a thin client. Optionally expose the council as an API/app. |
Networked — same contract. |
(The v0.1 deterministic hybrid panel — Claude moderating persistent Codex panelists — is delivery history, not an Axis-A stage: it did not change retrieval.)
Axis B — citation admissibility & enforcement. Pivots on the structured claim / quote-admissibility policy, escalating from prompt-only to a fail-closed boundary.
| Stage | What | Enforcement |
|---|---|---|
| B0 — Unified policy (done, v0.2.0) | One policy source → four surfaces; memory alone never supports [Text]; packets are untrusted data. |
Instruction-enforced; not fail-closed. |
| B1 — Structured claims + evidence seam (done, v0.5.0) | Hybrid opt-in mode parses religion-council/claim/v1, rejects malformed payloads (retry → repair → drop), and binds valid claims to RetrievalEvidenceAdapterV1 evidence seeds. Initial verification is always unverified. |
Schema rejection only; not B2 validation or B3 fail-closed. |
| B2 — Claim-level validation (done, v0.6.0) | Hybrid opt-in verify_claims=true validates bound [Text] quotation edges against curated snapshots, validates source-bound summaries by evidence edge, removes failed support edges, and downgrades all-failed [Text] claims to [Unverified citation]; the council still completes. |
Claim-level runtime validation; not B3 fail-closed. |
| B3 — Fail-closed boundary (done, v0.7.0) | Hybrid opt-in fail_closed=true runs a response-boundary gate after B2: unknown claim types, unvalidated [Text], renderer bypasses, and unsupported protocols are default-denied before rendering. |
Hybrid runtime-enforced / fail-closed. |
| ADR 0004 capstone — strict finalization (done, v0.9.0) | profile="strict" requires the complete graph and debate_finalize builds a deterministic, trace-validated Surface A from admitted claims only, with a non-removable Surface B frame and separate audit input. |
End-to-end machine-enforced construction and traceability of Surface A for strict-finalized responses only. |
The enforcement ladder keeps three rejections distinct: B1 rejects malformed
structure (retry/repair/drop) · B2 removes a failed [Text] support edge (may keep a
non-supporting [Unverified citation]) and continues · B3 default-denies at the
response boundary. Claude-only and portable modes stay instruction-enforced. From v0.7.0,
hybrid structured runs surface a response-level enforcement mode:
structured-schema-enforced when payloads bind successfully, or
structured-claim-validated when verify_claims=true and B2 validation ran, or
structured-fail-closed when fail_closed=true and the B3 boundary gate ran. Per-claim
runtime assurance and renderability remain in the claim verification and boundary decision
records.
Where the axes meet. Axis A's retrieval envelope is converted by the B1 adapter into immutable, content-addressed Artifact/Span identity — so A can swap retrieval backends without rewriting personas, and B can raise enforcement without first building a vector DB:
retrieve_envelope() → RetrievalEvidenceAdapterV1 → Artifact + Span → ClaimEvidenceEdge → validator → renderer
Distribution split. The portable skills/ stays snippet-based, file-based,
instruction-enforced — the demo anyone can clone and run. The project-integrated
.claude/ + orchestrator/ + retrieval service grows the full corpus, structured
protocol, validator, and hybrid fail-closed enforcement. Both share the core policy, not
dependencies or enforcement guarantees. The portable and project retrievers are now intentionally
forked (ADR 0006): they share one retrieval-envelope
contract, proven by a conformance suite (tests/retrieval_contract/) rather than by byte-parity.
Byte-parity is kept only as a narrow same-artifact check between the two portable retrieve.py copies.
Rights gate (tiered). A1 requires excerpt-level provenance + a rights note per snippet;
A2 requires full operational redistribution clearance (redistributable = true,
jurisdiction notes, review record) before any full text enters the distributable corpus.
This project takes source integrity seriously: no fabricated locators, no presenting a generated translation as a published quotation, the Qur'an in Chinese always marked as a rendering, and the strongest honest version of each position before any critique. For personal crises (self-harm, abuse, medical/legal/financial), it points to professional help first and treats the council as supplementary reflection only.
→ Read the full policy: DISCLAIMER.md
Want to add a tradition, denomination, or thinker — or grow the corpus? The citation rules are the heart of the project, so please read the guide first.
Dual-licensed, by purpose:
| Part | License | Covers |
|---|---|---|
| Skill logic | MIT | SKILL.md, USAGE.md, .claude/agents/*.md, orchestrator/, scripts/*.py, .mcp.json, agents/openai.yaml |
| Written content | CC BY 4.0 | references/, the 01–08 corpus, docs/CORPUS.md |
Quoted primary scriptures are public-domain source texts in their original languages; CC BY 4.0 covers this project's curation, summaries, and renderings. Attribute to "Religion Council contributors."
English · 繁體中文
以原典為據、由中立主持人調度的「跨傳統哲學議會」——可直接掛載到 AI coding agent 的 skill。
多傳統哲學議會(Religion Council) 讓你的 AI agent 化身為中立的主持人,召集一場 以原典為據、公平的跨傳統對話——而不是把各家硬揉成一團含糊的共識。
問一句「人生有沒有意義?」,你會聽到佛教、道教、儒家各自從自己的經典立論。每一句發言都 標注為**〔據典〕(引文+真實出處)或〔詮釋〕**;主持人負責把真正的張力點攤開,而非強行 調和。
目前 v0.12.6 支援三種執行方式:
- 純 Claude Code——附 37 個專屬 agent(1 位主持人 + 36 個聲音)。
- 純 Codex——可攜 Codex skill;明確要求時可用 Codex 原生 subagent。
- Claude 主持 + Codex 議員——Python MCP controller 保存 Codex threadId、執行 barrier、重試及紀錄。
三種模式共用同一份引用可採性政策(quote-admissibility/v2)。混合 controller 在 v0.5.0
加入 opt-in B1b 結構化 claim;v0.6.0 進一步加入 opt-in B2 verify_claims=true,可把 quotation
claim 對 curated evidence snapshot 做執行期驗證,並以每個 claim 的 runtime-validated 或
failed 結果寫入 state。v0.7.0 加入 opt-in B3 fail_closed=true:只有明確 admit 的 claim 可通過
response boundary;unknown claim type、未 runtime-validated 的〔據典〕、renderer bypass 與不支援
protocol 會預設拒絕。純 Claude 與可攜模式仍維持 instruction-enforced。
v0.8.0 已新增 A1 語料加厚的 metadata 基礎建設:retrieve.py 可從可攜的
presentation.json sidecar 合併 curated representation_kind、rendering_mode、
provenance 與 rights,並為既有古蘭經中文釋義片段建立種子標註。大量新增公有領域摘錄仍屬
需來源與 rights review 的人工 curation 工作。
v0.9.0 加入混合 controller 的 strict-finalization capstone。profile="strict" 要求完整的
structured → verify → fail-closed 圖與 evidence envelope，並把 finalization 設成獨立、明確的
步驟。詳見 ADR 0004 與
v0.9.0 變更紀錄。
v0.11.0 新增 stable occurrence identity 與附權利範圍說明的語料基線,僅為 A2/A3 readiness;並未 加入 index、vector store 或 RAG backend。後續工作仍受 retriever-fork contract 與 benchmark 決策把關。
v0.12.0 建立下一階段所需的 retriever fork：portable retriever 仍是 stdlib-only、file-based； project retriever 則以同一份 versioned retrieval envelope 與 shared conformance suite 為約束。這並未 選定 index、vector store、RAG backend 或 edition-backed assurance；它們仍須先有 benchmark evidence。 詳見 ADR 0006。
v0.12.1 定義該 benchmark 與採用門檻：候選後端必須同時勝過 lexical baseline，並保住 stable occurrence identity、retrieval contract、artifact lifecycle、rights 邊界與 assurance honesty。詳見 Retrieval Benchmark v1。
v0.12.2 執行第一次 retrieval-v1 lexical baseline。benchmark 現有凍結 18-query fixture set、graded relevance judgments、deterministic offline runner，以及已 commit 的 JSON/Markdown baseline reports。 baseline 識別出 lexical retriever 的兩大弱點：broad thematic recall 與 no-answer discrimination。 未選定任何後端。
v0.12.3 評估 experiment-only lexical confidence threshold，並加入 GitHub issue templates 以承接公開 feedback。threshold 2 與 3 消除了 benchmark 的 no-answer false-support cases，且未造成 answerable-query regression；threshold 5 則使 q007 與 q010 regression。目前仍未採用 threshold 或任何 backend。
v0.12.4 評估 experiment-only BM25-style lexical ranking candidate。BM25 改善部分 ranking metrics， 包含 MRR 與 nDCG@5，並保住 exact-span hit rate；但它沒有改善 no-answer discrimination 或 broad thematic recall。未選定任何 backend，default retrieval 也維持不變。
v0.12.5 評估 experiment-only BM25 + lexical confidence threshold candidate。BM25 提供 ranking 改善，threshold 提供 no-answer discrimination：MRR 從 0.938 提升到 0.969，nDCG@5 從 0.902 提升到 0.919，exact-span hit 維持 1.000，no-answer correctness 達到 1.000，false-support 降到 0.000。q010 的 broad thematic recall 仍然偏弱，因此未選定任何 backend，default retrieval 也維持不變。
v0.12.6 發布 ADR 0007，也就是根據 retrieval-v1 evidence 作出的 retrieval backend decision。 它選擇 BM25 + lexical-confidence threshold (t2/t3) 作為 project retriever 的 ranking + no-answer policy，但把 default BM25 ranking flip 仍然 gate 在 ≥2-judge + κ 步驟之後。它仍未選定 RAG、vector、hybrid、built-index 或 network backend；default retrieval 會等後續實作才改變。
strict run 的生命週期如下：
debate_start(profile="strict", evidence_envelope=…)
→ collect
→ finalization_required=true / finalized=false
→ debate_finalize
→ finalized=true
debate_finalize 產生兩個刻意分離的 answer surface：
- Surface A — textual authority。
AuthorityRenderUnit只能由 canonical builder mint， 經獨立 trace validation 後才 deterministic serialization。quotation 文字取自 canonical snapshot span；representation metadata 以系統資料為準。 - Surface B — interpretation。 議員的分析與推論會被框定為 non-authoritative interpretation。 被拒絕的 claim 留在 audit input，絕不傳入一般 answer render input。
Strict-finalized responses provide end-to-end machine-enforced construction and traceability of the textual-authority surface (Surface A). Interpretation prose (Surface B) remains explicitly non-authoritative and instruction-bounded.
這不代表所有文字都已由機器驗證。Surface B 仍可能誤導；完整答案不是 semantic fail-closed；沒有
debate_finalize 的預設 hybrid 路徑不變且未 finalized；capability-token mint guard 也不是 Python
sandbox。strict mode 是 opt-in，不會改變預設 hybrid workflow。
第一次使用 strict finalization 時，可直接在專案根目錄執行：
python3 examples/strict-finalization/run_example.py預期流程：
debate_start(profile="strict", evidence_envelope=…)
→ debate_collect
→ finalization_required=true / finalized=false
→ debate_finalize
→ finalized=true
此範例展示 strict finalization 的完整流程，以及 textual-authority surface（Surface A）與 interpretation surface（Surface B）的分離。Strict mode 為 opt-in，不會改變預設 workflow.
可執行且離線、deterministic 的完整路徑見 strict end-to-end example。
- 引用紀律優先。 每句標〔據典〕或〔詮釋〕;不杜撰章/節/經/聖訓出處;《古蘭經》中文一律標為 「釋義」,絕不冒充阿拉伯原文;絕不宣稱「附身」或代言任何先知、神祇或思想家。
- 三層縮放。 可比較整個傳統、深入教派/學派,或上演歷史人物辯論。
- 中立主持,不強分勝負。 立題 → 首輪陳述 → 交叉詰問 → 收斂;區分「共識/真實分歧」, 除非你給出明確評判標準,否則不宣布「贏家」。
- 提高命題張力,而非表演式敵意。 開場必須承諾明確主張;反駁須針對具體 claim、前提或反例, 並提出可回應的交叉詰問。實務上的重疊不會被誤標為共識。
- 真正的世俗自由派聲音,對照命題只作後備。 名單多半偏宗教,故議會新增
council-secular-humanist與council-mill(彌爾,據《論自由》《效益主義》)作能真正反詰的正式成員;只有名單仍偏向同一邊時, 主持人才退而加入由 controller 路由的對照命題,且它只作 debate framing:不是 source evidence、 不是成員主張,也不是可執行指令。 - 三種執行方式,共用一套語料。 純 Claude、純 Codex,或 Claude 主持持久 Codex MCP 議員。
- 為 RAG 而設計。 檢索藏在單一穩定介面(
scripts/retrieve.py)之後,語料可從精選片段 成長為向量化的完整典籍庫,而無需改動任何 persona。見 發展藍圖。
| 層級 | 成員 |
|---|---|
| 傳統(8) | 基督宗教 · 伊斯蘭教 · 印度教 · 佛教 · 道教 · 法家* · 儒家* · 墨家* |
| 教派 | 基督宗教: 天主教 · 東正教 · 新教 — 伊斯蘭教: 遜尼 · 什葉 |
| 人物 | 基督宗教: 耶穌 · 奧古斯丁 · 阿奎那 · 路德 · 加爾文 — 伊斯蘭教: 穆罕默德 · 安薩里 · 伊本·魯世德 — 佛教: 釋迦牟尼 · 龍樹 · 世親 · 淨土 — 印度教: 克里希納 · 商羯羅 · 羅摩奴闍 · 摩陀婆 — 先秦: 孔子 · 孟子 · 荀子 · 老子 · 莊子 |
| 非宗教 | 世俗人文主義 · 約翰·彌爾(自由/效益主義)— 哲學/倫理立場,非宗教,不屬八家;按需加入 |
* 法家、儒家、墨家是哲學/思想流派而非宗教,skill 會據此標示(並在必要時區分哲學道家與 宗教道教)。
1. 純 Claude Code:
git clone https://github.com/F-e-u-e-r/religion-council.git
cd religion-council
claude # 在專案根目錄開啟 Claude Code然後在對話中:
用議會討論:人生有沒有意義?請佛教、道教、儒家三家先各自陳述,再交叉辯論。
主對話會把任務交給 council-moderator,由它立題、逐一調度各 council-* 成員、轉述彼此
論點並收斂分歧。
2. 純 Codex(可攜、自足):
把 agent 的 skill 載入路徑指向 skills/religion-council/,然後呼叫:
Use $religion-council to convene a sourced roundtable on whether life has meaning.
skills/religion-council/SKILL.md 完全自足;沒有 sub-agent 時會自動降級為單一脈絡內的對話。
3. Claude 主持 + Codex 議員:
先登入 Codex,在 repo 根目錄啟動 Claude Code,並於 /mcp 批准專案 MCP server;再請 Claude
使用 religion-council-controller 與 orchestrator/panelists/religion-8.json。Controller
會跨輪保存同一批 Codex threadId,並在全員完成前阻止進入下一輪。
→ 各平台完整設定:INSTALL.md
發言前,每個聲音都先讀 SKILL.md(提問層次、引用紀律、防斷章取義、圓桌流程);各傳統的立場與
可引用片段放在 references/<傳統>.md。在 Claude Code 中,council-moderator 透過 Agent
工具調度各 council-* 成員(sub-agent 之間不能直接對話,故由主持人轉述)。檢索則統一走
scripts/retrieve.py 的穩定介面(輸出 text + tradition/school/work/locator/language/ version/category metadata):v0.1 解析並以詞彙比對排序 references 片段,未來換成向量檢索,而上層的
persona 與引用紀律不必更動。
整個專案沿兩條低耦合的軸線發展,並在一個版本化的接縫處交會。架構階段只用 A0–A3(語料與檢索)與 B0–B3(引用可採性與強制力);PR 編號只是交付歷史,不是階段名稱。 規範細節見 ADR 0002(階段命名 + 強制力梯度) 與 ADR 0003(檢索→證據 adapter)。
軸線 A — 語料與檢索。 樞紐是一條接縫:檢索 envelope 契約。守住它,後端由檔案逐步升級為 向量服務時,36 個 persona 都不必改。
| 階段 | 內容 | 檢索 |
|---|---|---|
| A0 — 精選議會 | 各成員引用 references/ 中手選、附出處片段,離線、任何 agent 皆可跑。 |
檔案解析 + 詞彙排序。 |
| A1 — 加厚語料(metadata 基礎,v0.8.0) | 新增 presentation.json provenance/rights sidecar,並為既有古蘭經中文釋義建立種子標註;更大規模的公有領域摘錄擴充仍須逐筆來源與 rights review。 |
仍為檔案式。 |
| A2 — 完整典籍 + 本地索引 | 把開放典籍全文入庫並切分;benchmark lexical / cross-lingual / dense / hybrid;再建所選索引。 | 本地索引——envelope 契約不變。 |
| A3 — RAG server | 索引移到檢索服務,retrieve.py 變薄客戶端;亦可把議會做成 API/應用。 |
網路檢索——仍是同一契約。 |
(v0.1 的可重現混合 panel——Claude 主持持久 Codex 議員——是交付歷史,不是 Axis-A 階段:它沒有改動檢索。)
軸線 B — 引用可採性與強制力。 樞紐是結構化 claim / quote-admissibility 政策,由純 prompt 指示逐步升級為 fail-closed 邊界。
| 階段 | 內容 | 強制力 |
|---|---|---|
| B0 — 統一政策(已完成,v0.2.0) | 同一政策來源 → 四個 surface;僅憑記憶永不支持〔據典〕;packet 視為不可信資料。 | instruction-enforced;尚未 fail-closed。 |
| B1 — 結構化 claim + 證據接縫(已完成,v0.5.0) | 混合模式可 opt-in 解析 religion-council/claim/v1,對格式錯誤 payload 執行 retry → repair → drop,並把有效 claim 綁定至 RetrievalEvidenceAdapterV1 evidence seeds。initial verification 恆為 unverified。 |
僅 schema 拒絕;不是 B2 驗證,也不是 B3 fail-closed。 |
| B2 — claim 層驗證(已完成,v0.6.0) | 混合模式可 opt-in verify_claims=true,對已綁定〔據典〕quotation edge 做 curated snapshot 驗證,source-bound summary 以 evidence edge 驗證;失敗 support edge 會移除,全失敗〔據典〕降為〔未驗證引用〕,議會仍可完成。 |
claim 層執行期驗證;不是 B3 fail-closed。 |
| B3 — fail-closed 邊界(已完成,v0.7.0) | 混合模式可 opt-in fail_closed=true,在 B2 之後執行 response-boundary gate:未知 claim type、未驗證〔據典〕、renderer bypass 與 unsupported protocol 會在 render 前預設拒絕。 |
混合模式 runtime-enforced / fail-closed。 |
| ADR 0004 capstone — strict finalization(已完成,v0.9.0) | profile="strict" 要求完整圖；debate_finalize 只從 admitted claim 建立 deterministic、trace-validated 的 Surface A，附有不可移除的 Surface B frame 與分離 audit input。 |
僅 strict-finalized responses 的 Surface A 具有端到端 machine-enforced construction 與 traceability。 |
強制力梯度把三種「拒絕」分清楚:B1 拒絕格式不良的結構(retry/repair/drop)·B2 移除失敗
〔據典〕的 support edge(可保留 non-supporting〔未驗證引用〕)後續行·B3 在 response 邊界預設拒絕。
純 Claude 與可攜模式維持 instruction-enforced。v0.7.0 起,混合模式的結構化回合會顯示
response-level enforcement mode:成功綁定時為 structured-schema-enforced;若同時
verify_claims=true 且 B2 驗證已執行,則為 structured-claim-validated;若再加
fail_closed=true 且 B3 邊界已執行,則為 structured-fail-closed。每個 claim 的真實 runtime
assurance 與可否 render 仍以 claim verification 與 boundary decision records 為準。
兩軸交會處。 軸線 A 的檢索 envelope 由 B1 adapter 轉換為不可變、content-addressed 的 Artifact/Span identity——因此 A 可更換檢索後端而不重寫 persona,B 可提升強制力而不必先建向量庫:
retrieve_envelope() → RetrievalEvidenceAdapterV1 → Artifact + Span → ClaimEvidenceEdge → validator → renderer
發行分工。 可攜的 skills/ 維持片段式、檔案式、instruction-enforced——人人 clone 後即可跑
的 demo。專案整合版 .claude/ + orchestrator/ + 檢索服務則養成完整語料、結構化 protocol、
validator 與混合 fail-closed 強制力。兩者共用核心政策,但不共用依賴或強制力保證。可攜版與專案版檢索器
現已依 ADR 0006 正式分叉:兩者共用同一份檢索 envelope 契約
(由 tests/retrieval_contract/ 的 conformance suite 保證),而非位元組 parity;位元組 parity 僅保留為
兩份可攜 retrieve.py 副本之間的窄同源檢查。
Rights gate(分層)。 A1 要求每個片段具 excerpt 層 provenance + rights note;A2 要求完整的
operational redistribution clearance(redistributable = true、jurisdiction notes、review record),
全文方可進入可分發 corpus。
本專案重視來源誠信:不杜撰出處、不把自動生成的翻譯當成已出版引文、《古蘭經》中文一律標「釋義」、 批評任何立場前先呈現其最強且誠實的版本。遇到個人危機(自傷、受虐、醫療/法律/財務)時,優先指向 專業協助,議會僅作輔助性反思。
→ 完整守則:DISCLAIMER.md
想新增傳統、教派或人物,或擴充語料?引用紀律是本專案的核心價值,動手前請先讀指南。
依用途雙重授權:
| 部分 | 授權 | 範圍 |
|---|---|---|
| Skill 邏輯 | MIT | SKILL.md、USAGE.md、.claude/agents/*.md、orchestrator/、scripts/*.py、.mcp.json、agents/openai.yaml |
| 書面內容 | CC BY 4.0 | references/、01–08 語料、docs/CORPUS.md |
所引原典在其原文中多屬公有領域;CC BY 4.0 涵蓋的是本專案的選編、摘要與釋義。請標註出處 「Religion Council contributors」。