ワンクリックでManusで任意のスキルを実行

shadow-mentor

スター1

フォーク0

更新日2026年7月20日 18:30

5-voice AI compliance council for regulated loan origination. 11 MCP tools (loan_council, loan_council_typed, risk_tools, recall, calibration, scenarios, traceability, verify_attestation, banking_profile, disparity, size_position). FICO < 700 is a hardcoded JS conditional with pinned tests. Strict-JSON enum verdicts. AA01–AA05 adverse-action codes per CFPB Circular 2026-03. MCPTox / OX Security 2026 named-threat coverage mechanically tested.

インストール

Codex または Claude でインストールこの Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。

Manusで実行

ソース

alex-jb

alex-jb/shadow-mentor

GitHub リポジトリを開く Creator のリポジトリを見る

ダウンロード

Manusで実行

shadow-mentor

5-voice AI compliance council for regulated lending. Drop into Claude Desktop / Cursor / Zed / OpenCode / OpenClaw via MCP. Mid-tier US bank pitch: $1,800 / compliance-officer seat / year. Open source MIT.

What it does

Eleven MCP tools that turn an LLM chat into a procurement-defensible loan-origination compliance surface:

shadow_loan_council — Deterministic 5-voice verdict (block / escalate / approve) + per-voice rationale + risk packet + thresholds applied. FICO < 700 is a hardcoded if with a pinned test.
shadow_loan_council_typed — same council with a typed request/response contract
shadow_risk_tools — Institutional risk primitives: VaR (historical / parametric / MC), Expected Shortfall, concentration (HHI / Gini), sector exposure, correlation (Pearson / Spearman / EWMA), factor exposures, beta decomposition
shadow_recall — Cross-session memory recall keyed by persona + scenario
shadow_calibration — Per-persona Brier calibration stats (for SR 26-2 model risk monitoring)
shadow_scenarios — Surface enumeration (5 personas × 4 scenarios × 4 device clients × 2 providers)
shadow_traceability — Source attribution for any threshold (BRD vs Addendum vs Risk Appetite Note) per CFPB / ECOA / SR 26-2
shadow_verify_attestation — verify a signed request→output attestation (Ed25519 / HMAC)
shadow_banking_profile — check a bundle against the Banking Evidence Profile ("is this credit decision auditable?"), with optional anchor-trust + examiner packet
shadow_disparity — Fair-Lending disparity math (SolasAI-aligned)
shadow_size_position — position-sizing risk primitive

All tools run in-process. No network call from inside the tool body. AA01–AA05 adverse-action codes match CFPB Circular 2026-03.

Why a bank's procurement team can grep it in 10 minutes

Three source files + four test files:

lib/audit-guardrail.js — 12-pattern regex output gate (Schema-Layer Safety)
lib/run-loan-council.js — if (loan.fico < CREDIT_THRESHOLDS.FICO_FLOOR) return { verdict: "block", ... } (Determinism Floor)
installer/tools.json — frozen install-target × scope catalog (EMA-ready)

Test surface:

test/mcptox-canary.test.js — 28 contract tests covering MCPTox §3 attack categories + MosaicLeaks coverage
test/oauth-scaffold.test.js + test/oauth-loan-council-wiring.test.js — Enterprise Managed Auth scope catalog + live wiring
test/glm-call.test.js — Multi-provider contract (Anthropic + GLM-5.2)
test/tools-catalog.test.js — Catalog drift gate

1592/1595 tests passing. Shadow Agentic Score 87 ± 3 (n=6).

Quick install

git clone https://github.com/alex-jb/shadow-mentor
cd shadow-mentor
npm install
node bin/install.mjs                          # see which MCP hosts are detected
node bin/install.mjs --host cursor --dry-run  # preview the merged config
node bin/install.mjs --host cursor            # write it

Defends against named 2026 MCP threats

OX Security MCP STDIO supply-chain advisory (May 2026) — Shadow's tool bodies call only frozen lib/ modules; no untrusted shell input reaches a tool body
MCPTox benchmark (arXiv 2508.14925) — Shadow returns strict-JSON enum verdicts, not narrative; poisoned descriptions cannot widen the response surface beyond schema
MosaicLeaks multi-turn information leakage — canary-token contract tests pin the invariant that an attacker cannot exfiltrate a canary across the tool boundary
MCP Enterprise OAuth (EMA) — opt-in SHADOW_REQUIRE_BEARER=1 enforces shadow:council scope on /api/loan-council; works with OAuth2 / Azure AD claim shapes

Links

Repo: https://github.com/alex-jb/shadow-mentor
Release: https://github.com/alex-jb/shadow-mentor/releases/tag/v1.2.0
Comparison vs Anthropic FS Agents: docs/positioning-vs-anthropic-fs.md
Principles: docs/principles/schema-layer-safety.md · docs/principles/determinism-floor.md
License: MIT

このリポジトリの他の Skills

同じリポジトリ

shadow-aml-kyc-investigator

alex-jb/shadow-mentor

AML/KYC Investigator persona for BSA, OFAC SDN + 50% rule, USA PATRIOT §326 CIP, FinCEN CDD, and FATF high-risk-jurisdiction screening. Hard-blocks on sanctions hit or unverified identity. Escalates on structuring, PEP, opaque beneficial ownership, GTOs. Every finding cites the specific rule. ACAMS 2026 procurement-lane persona.

2026-07-101

shadow-attestation-verify

alex-jb/shadow-mentor

Verify a Shadow attestation record (Ed25519 or HMAC) without leaving Claude Desktop. Confirms request wasn't tampered, response wasn't tampered, the exact pinned model ran, and the deployment key material matches. Use when a bank auditor hands you a persisted JSON response and asks "is this real." Works cross-language — Python shadow-verify library produces byte-identical results for the same inputs on Python 3.9–3.13.

2026-07-101

shadow-compliance-officer

alex-jb/shadow-mentor

Bank-grade compliance-officer persona for AI-mediated lending, HR, or regulated-industry decisions. Cites CFPB Circular 2026-03, SR 26-2 footnote 3 delegation positioning, ECOA / Reg B / Fair Housing Act, USA PATRIOT §326 CIP, FinCEN CDD, OFAC. Refuses to approve without human review. Never invents thresholds not in checked-in policy.

2026-07-101

shadow-customer-advocate

alex-jb/shadow-mentor

Customer-Advocate persona for adverse-action notice quality control. Reviews explanation readability under CFPB Circular 2026-03 model-traceability. Escalates on any adverse-action reason present. Never blocks — human-review path, not veto path.

2026-07-101

shadow-ds-govern

alex-jb/shadow-mentor

Run a 5-voice data-science governance council on an ML artifact reference. Data Steward (drift), Model Validator (calibration), Fair-ML Auditor (EEOC 80% rule, always blocks on violation), Reproducibility Critic (metadata completeness), Ops Realist (p95 latency). Returns SHIP / REWORK / BLOCK with per-voice rationale. Pure computation — no LLM. Use before shipping any ML pipeline that will make consumer-facing decisions.

2026-07-101

shadow-macro-contrarian

alex-jb/shadow-mentor

Macro-Contrarian persona for sector-cycle + recession-sensitivity dissent. Provides counter-narrative even when other voices approve. Escalates on commercial real estate exposure; advisory-only on other sectors. Never blocks — dissent voice, not veto voice.

2026-07-101

name	shadow-mentor
description	5-voice AI compliance council for regulated loan origination. 11 MCP tools (loan_council, loan_council_typed, risk_tools, recall, calibration, scenarios, traceability, verify_attestation, banking_profile, disparity, size_position). FICO < 700 is a hardcoded JS conditional with pinned tests. Strict-JSON enum verdicts. AA01–AA05 adverse-action codes per CFPB Circular 2026-03. MCPTox / OX Security 2026 named-threat coverage mechanically tested.
version	1.2.0
license	MIT
authors	["Alex Xiaoyu Ji","Loredana C. Levitchi"]
runtime	node
install	["npm install","node bin/install.mjs --host <claude-desktop\|cursor\|zed\|opencode\|openclaw>"]
tools	["shadow_loan_council","shadow_loan_council_typed","shadow_risk_tools","shadow_recall","shadow_calibration","shadow_scenarios","shadow_traceability","shadow_verify_attestation","shadow_banking_profile","shadow_disparity","shadow_size_position"]

shadow-mentor

What it does

Eleven MCP tools that turn an LLM chat into a procurement-defensible loan-origination compliance surface:

shadow_loan_council — Deterministic 5-voice verdict (block / escalate / approve) + per-voice rationale + risk packet + thresholds applied. FICO < 700 is a hardcoded if with a pinned test.
shadow_loan_council_typed — same council with a typed request/response contract
shadow_risk_tools — Institutional risk primitives: VaR (historical / parametric / MC), Expected Shortfall, concentration (HHI / Gini), sector exposure, correlation (Pearson / Spearman / EWMA), factor exposures, beta decomposition
shadow_recall — Cross-session memory recall keyed by persona + scenario
shadow_calibration — Per-persona Brier calibration stats (for SR 26-2 model risk monitoring)
shadow_scenarios — Surface enumeration (5 personas × 4 scenarios × 4 device clients × 2 providers)
shadow_traceability — Source attribution for any threshold (BRD vs Addendum vs Risk Appetite Note) per CFPB / ECOA / SR 26-2
shadow_verify_attestation — verify a signed request→output attestation (Ed25519 / HMAC)
shadow_banking_profile — check a bundle against the Banking Evidence Profile ("is this credit decision auditable?"), with optional anchor-trust + examiner packet
shadow_disparity — Fair-Lending disparity math (SolasAI-aligned)
shadow_size_position — position-sizing risk primitive

All tools run in-process. No network call from inside the tool body. AA01–AA05 adverse-action codes match CFPB Circular 2026-03.

Why a bank's procurement team can grep it in 10 minutes

Three source files + four test files:

lib/audit-guardrail.js — 12-pattern regex output gate (Schema-Layer Safety)
lib/run-loan-council.js — if (loan.fico < CREDIT_THRESHOLDS.FICO_FLOOR) return { verdict: "block", ... } (Determinism Floor)
installer/tools.json — frozen install-target × scope catalog (EMA-ready)

Test surface:

test/mcptox-canary.test.js — 28 contract tests covering MCPTox §3 attack categories + MosaicLeaks coverage
test/oauth-scaffold.test.js + test/oauth-loan-council-wiring.test.js — Enterprise Managed Auth scope catalog + live wiring
test/glm-call.test.js — Multi-provider contract (Anthropic + GLM-5.2)
test/tools-catalog.test.js — Catalog drift gate

1592/1595 tests passing. Shadow Agentic Score 87 ± 3 (n=6).

Quick install

git clone https://github.com/alex-jb/shadow-mentor
cd shadow-mentor
npm install
node bin/install.mjs                          # see which MCP hosts are detected
node bin/install.mjs --host cursor --dry-run  # preview the merged config
node bin/install.mjs --host cursor            # write it

Defends against named 2026 MCP threats

OX Security MCP STDIO supply-chain advisory (May 2026) — Shadow's tool bodies call only frozen lib/ modules; no untrusted shell input reaches a tool body
MCPTox benchmark (arXiv 2508.14925) — Shadow returns strict-JSON enum verdicts, not narrative; poisoned descriptions cannot widen the response surface beyond schema
MosaicLeaks multi-turn information leakage — canary-token contract tests pin the invariant that an attacker cannot exfiltrate a canary across the tool boundary
MCP Enterprise OAuth (EMA) — opt-in SHADOW_REQUIRE_BEARER=1 enforces shadow:council scope on /api/loan-council; works with OAuth2 / Azure AD claim shapes