菜单
5-to-6 voice AI compliance council for regulated lending. Deterministic verdict in milliseconds. Signed Ed25519 attestation. Runs in your VPC. Cites CFPB / SR 26-2 / ECOA / Reg B / BSA / OFAC / FinCEN CDD by construction. Pure-compute pass — no LLM in the tool body.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
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.
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.
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.
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.
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.
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.
基于 SOC 职业分类
| name | shadow-loan-council |
| description | 5-to-6 voice AI compliance council for regulated lending. Deterministic verdict in milliseconds. Signed Ed25519 attestation. Runs in your VPC. Cites CFPB / SR 26-2 / ECOA / Reg B / BSA / OFAC / FinCEN CDD by construction. Pure-compute pass — no LLM in the tool body. |
| version | 1.4.0 |
| authors | ["Alex Xiaoyu Ji <xji1@mail.yu.edu>","Loredana C. Levitchi (BRD + Addenda A/B/C + Risk Appetite Note)"] |
| license | MIT |
| repo | https://github.com/alex-jb/shadow-mentor |
| tags | ["banking","regulated-lending","loan-origination","underwriting","compliance","sr-26-2","cfpb","ecoa","reg-b","aml","kyc","hitl","human-in-the-loop","attestation","ed25519","mcp"] |
The pure-compute verdict layer of Shadow's compliance council. Takes a typed loan dict + policy + optional AML/KYC signals; returns a {verdict, voices, adverse_action_codes, traceability, attestation, ...} bundle in milliseconds. NO LLM call inside the tool body — every threshold is hardcoded to a checked-in policy artifact.
Install this skill in Claude Desktop / Cursor / OpenCode when you want Claude to:
{
"final_verdict": "approve" | "escalate" | "block",
"confidence_weighted_verdict": "approve" | "escalate" | "block",
"aggregated_score": -1.0 to 1.0,
"voice_contributions": [{ voice, weight, confidence, score }],
"voices": [ { voice, verdict, confidence, rationale, adverse_action_codes, metrics } ],
"presentation_order": [3, 0, 1, 5, 2, 4],
"risk_packet": { var_95_10d, es_95_10d, concentration, sector_exposure, risk_budget_status },
"adverse_action_codes": [ { code: "AA01" | ... | "AA06", label, source } ],
"reason_code_dictionary_check": { ok, invalid, reason },
"protected_class_proxy_check": { ok, prohibited, reason },
"thresholds_applied": { fico_floor, dti_ceiling, ltv_ceiling, var_ceiling, ... },
"traceability": { ...maps every threshold to BRD vs Addendum vs Risk Appetite Note... },
"attestation": {
"version": "aex-attestation/v1",
"mode": "ed25519" | "hmac-sha256",
"request_commitment": "sha256 hex",
"output_commitment": "sha256 hex",
"model_id": "runLoanCouncil/pure-compute",
"completed_at_utc": "2026-07-03T...",
"signature": "base64 (Ed25519) or hex (HMAC)"
}
}
5 fixed voices; 6th (AML/KYC Investigator) attached when the loan carries aml_flags[] or kyc_status:
npx skills add alex-jb/shadow-mentor/skills/shadow-loan-council
Or install the full Shadow MCP server:
git clone https://github.com/alex-jb/shadow-mentor
cd shadow-mentor
node bin/install.mjs --host cursor # or claude / opencode / zed
This gives you shadow_loan_council as an MCP tool Claude can call directly.
Every response is signed. To let a bank auditor independently verify without holding your server secret:
SHADOW_ATTESTATION_MODE=ed25519 + private key env var on ShadowverifyAttestation(att, req, res, { publicKey }) on any past decisionRFC 8032 EdDSA. Bank can VERIFY, cannot FORGE.
lib/schemas/reason-code-dictionary.jsonlib/persona-schema.jsonlib/confidence-weighted-verdict.jslib/attestation.jslib/aml-kyc-voice.js