Conversational Finance Governance Framework
An open-source, model-agnostic engineering governance standard defining runtime trust boundaries for conversational AI agents in Fintech. MIT licensed.
- Source repo
- evanslyhan/Fintech-CUI-Trust-Framework
- Stars
- ★ 175
- Last updated
- 2mo ago
- License
- MIT
- Primary language
- Python
- FA score
- 51/100 · Major gaps
At a glance
- How it runs
- Works with
- Universal · cross-platform
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- A payments product team validates, via the payments.yaml mapping, which customer utterances require authorization confirmation before letting an agent commit fees or transfers
- Not a fit if
- Teams needing a general-purpose assistant governance protocol; scope is Fintech CUI only
- Organizations seeking turnkey compliance software rather than a specification to implement
- Conversational AI projects outside payments, lending, insurance, collections, or onboarding
- Maintenance
- Commits in the last 6 monthslast commit 2026-09-09
- Source review
- 51/100 · Major gaps
What does this agent do, and when should you use it?
This repository (evanslyhan/Fintech-CUI-Trust-Framework) hosts an MIT-licensed, open engineering governance specification, currently at v0.4, defining the structural components a conversational AI agent must implement before deployment in Fintech contexts where its outputs could create financial obligations, regulatory exposure, or consumer harm. It is not a software product but an RFC-like public specification; its concrete artifact is a Python CLI sidecar under src/ that loads mapping files from mappings/ (payments, lending, insurance, collections) and classifies individual utterances through the control plane. The architecture is a five-layer control plane: an Intent Classification Matrix scoring Financial Commitment Risk (F0–F3) and Regulatory Sensitivity (R0–R3), an Authorization Trigger, a Generation Boundary, an Audit Trail Pipeline, and an Authority Circuit Breaker, extended in v0.4 by three runtime objects: Action Object, Scoped Credential, and Visible Authority State. Core and mapping are decoupled: the five layers stay invariant across Fintech products while product-line rules are configured via a Compliance Mapping Layer. It fits organizations with engineering resources whose conversational agents touch payments, credit, insurance, collections, KYC/onboarding, or other regulated consumer-finance decisions.
It reads a mapping file (e.g. mappings/payments.yaml, mappings/lending.yaml) and an utterance, then runs the control plane via the cf_gov.sidecar module in src/: the Intent Classification Matrix classifies each turn along Financial Commitment Risk (F0–F3) and Regulatory Sensitivity (R0–R3), defaulting to the stricter action on low confidence; the Authorization Trigger decision table determines whether generation must suspend for explicit user confirmation; transact-class actions must use a single-use, amount-capped, time-bounded Scoped Credential instead of durable account or card secrets; the Generation Boundary constrains permissible response space at pre- and post-generation stages; the Audit Trail Pipeline emits an immutable, hash-chained, machine-readable log of every generation, authorization, escalation, credential issue, and circuit-break decision; and the Authority Circuit Breaker severs generation privilege in real time and transfers control to a human operator or deterministic SOP. The Action Object binds each turn to one of inform, advise, commit, transact, represent, assigning authority to the action rather than the model. The same utterance can compile to different authority decisions under different mappings. The Lite demo requires no model at all.
- A payments product team validates, via the payments.yaml mapping, which customer utterances require authorization confirmation before letting an agent commit fees or transfers
- A lending platform's compliance engineering team configures generation boundaries and low-confidence classification fallbacks for advise-class responses
- An insurer's audit team requires that every authorization, escalation, and circuit-break decision lands in the hash-chained audit pipeline
- A collections business line uses collections.yaml to compile different authority decisions for the same utterance than the payments line
- A KYC/onboarding team exposes Visible Authority State showing current action class and live permissions where the agent must not make compliance representations
- A technical due-diligence reviewer runs the Lite sidecar on a single utterance to demo the control plane without hooking up any LLM
Key setup facts
Pulled from the install and usage notes below. The project README has the current values.
- Config files
../mappings/payments.yaml../mappings/lending.yaml
How do you install or deploy this agent?
No one-command installer is documented; clone the repository and install Python dependencies manually:
git clone https://github.com/evanslyhan/Fintech-CUI-Trust-Framework.git
cd Fintech-CUI-Trust-Framework/src
python3 -m pip install -r requirements.txtConcrete package names and versions live in src/requirements.txt; the README does not enumerate them. The full specification is in spec/protocol-spec-v0.4.md and mapping examples in mappings/.
How do you use this agent?
The Lite path needs no model; run the sidecar on a single utterance:
python3 -m cf_gov.sidecar --mapping ../mappings/payments.yaml --utterance "What's the fee for sending 200 dollars?"
python3 -m cf_gov.sidecar --mapping ../mappings/lending.yaml --utterance "Should I take the loan?"The Standard path adds periodic audit-trail review and loading a product-line mapping. The Full path means forking the spec and sidecar, integrating with existing compliance infrastructure, and customizing the mapping layer — without weakening core triggers, boundaries, or the breaker. The full specification is at spec/protocol-spec-v0.4.md.
What are this agent's strengths and limitations?
- Model-agnostic: governance mechanisms are embedded in the interaction architecture rather than tied to any LLM vendor, so they survive model swaps
- Core/mapping decoupling: the five layers stay invariant while product-line rules are configured through the Compliance Mapping Layer, letting the same utterance compile to different authority decisions per line
- Strong auditability: audit logs are immutable, hash-chained, machine-readable, and cover authorization, escalation, credential issuance, and circuit-break decisions
- A runnable Lite sidecar plus classification gold-standard tests allow verifying control-plane behavior without any model
- It is a specification, not finished software: adopters must build their own engineering implementation of the five layers and three runtime objects
- Reference mapping files are explicitly stated not to be legal advice; regulatory conclusions remain the adopting organization's responsibility
- Full-level adoption requires integrating with existing compliance infrastructure and customizing the mapping layer — a significant integration cost
- Scope is strictly Fintech CUI and the repository offers no hosted service or deployment packaging
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Conversational Finance Governance Framework This agent | 51 · Major gaps | CLIFree | ★ 175 | 2mo ago | Python | — |
| Preloop | 81 · Good | — | ★ 95 | today | Python | Codex · Claude Code |
| Trigger.dev | 78 · Good | Library / SDKFreemium | ★ 17k | today | TypeScript | — |
| Auto Browser | 76 · Some gaps | Self-hosted serviceFree + model costs | ★ 929 | 5d ago | Python | OpenAI API · Claude API |
How does FollowAgents rate this agent?
Why each dimension lost points
Evidence shows least privilege (single-use, amount-capped, time-bounded scoped credentials), authorization gating, hash-chained audit trail and circuit breaker defined at spec level, with tests asserting secrets never enter model context and high-risk utterances must escalate. Deductions: these are design claims and test assertions without a full implementation in the packet; requirements.txt is unseen so dependency security is unassessable; the spec describes immutable logging but no transaction rollback mechanism; the copyright holder (madguyevans-creator) does not match the repo owner (evanslyhan), leaving source attribution unclear.
Tests demonstrate internal consistency requirements (payments/collections mappings must diverge, classifications may never be weaker than the gold standard). Deductions: requirements.txt is not in the sample so dependency availability is unknown; tests cover only happy paths; no evidence of error-message design or exception handling.
Audience and scenario scoping is explicit (Fintech CUI only, explicitly not a general-purpose protocol); trigger precision is backed by the dual-axis intent matrix with a stricter-default rule for low-confidence classifications; capability boundaries include a 'not legal advice' disclaimer. Deductions: environment fit is evidenced only by a Python sidecar demo; the model-agnostic claim has no cross-environment evidence.
Repository structure, documentation hierarchy, and the MIT LICENSE file are all present; license merits full marks. Deductions: install notes cover only the Lite sidecar; no FAQ; no known-limitations section beyond scattered disclaimers; version numbers exist but no changelog; no maintainer statement and the copyright holder name mismatches the repo owner, weakening naming/attribution stability.
The sidecar emits structured decisions and a visible authority state, and the Lite entry is cheap (no model needed for the control-plane demo). Deductions: marginal value depends on the full spec and mapping files not included in the packet, so real governance effectiveness cannot be confirmed from three sample files.
Fact/inference separation is handled well (research framing documents are explicitly separated from the normative spec; mappings are marked 'not legal advice'). Deductions: key claims (five-layer architecture, whitepaper, concept paper) point to PDFs and spec files not present in the packet; a static review cannot cross-corroborate them, and there is no evidence of alignment with external standards.
- This is a specification plus a demo sidecar, not a deployable governance product; read the full spec and implement/validate it yourself before adoption.
- The copyright holder (madguyevans-creator) does not match the repository owner (evanslyhan); the publisher is unverified, so confirm provenance independently.
- The dependency manifest was not in the reviewed sample; audit requirements.txt and the supply chain yourself before integration.
- Tests cover only a few paths and the coverage of the 'high-cost false negative' gold standard is unknown; do not treat a passing 'ok' as compliance certification.
- Mapping files are explicitly 'not legal advice'; any regulated deployment requires independent review by a compliance team.
FAQ
Which LLM or API does it require?
Can I deploy this as compliance software out of the box?
Which business domains does it cover?
How does an agent get transaction authority?
What happens when classification is uncertain?
Is Conversational Finance Governance Framework free?
Is Conversational Finance Governance Framework open source?
Is Conversational Finance Governance Framework safe to use?
Is Conversational Finance Governance Framework still maintained?
Who makes Conversational Finance Governance Framework?
Thanks — we use this to improve the page.