Autonomy Ladder™ — Reference Libraries

Six open, DOI-archived governance libraries for autonomous AI in regulated operations. One framework (A0→A4), six co-equal regulated verticals, a common evidence discipline. This is the single landing page for the family — start here.

Framework + whitepaper: autonomy-ladder.io

What this family is: reference implementations of deployment-authority governance for AI agents — the A0→A4 rungs, a non-overridable sovereign veto, a hash-chain audit ledger, hard envelopes, and mechanical demotion — encoded per regulated decision class as runnable, tested Python. What it is not: deployed production controls. These are reference IP for adoption and inspection — never represented as a control running in any institution's production.

---

The six libraries

Library	Vertical	Decision classes encoded	Primary regulator / rule	Tests · coverage	DOI	Start here
finserv-agent-audit	Cross-vertical financial services	Lending, core FS decisioning	EU AI Act Art. 14 · SR-letters · FCRA	630 · 93%	10.5281/zenodo.20434570	DEFCON + Sovereign-Veto demo
banking-agent-audit	Banking	Model risk · ECOA/Reg B adverse action · BSA/AML/OFAC	ECOA 12 CFR §1002.9 · FINRA 3110	182 · ~99%	10.5281/zenodo.20564584	tests/ ECOA adverse-action
payments-agent-audit	Payments	OFAC screening · BSA/AML · Reg E · rail finality	OFAC · Reg E · rail-finality (AL-PROBE-06)	183 · 98.97%	10.5281/zenodo.20592773	AL-PROBE-06 irreversibility gate
payer-agent-audit	Health-insurance payer	Utilization mgmt · prior auth · claims/appeals	NAIC Model Bulletin (no medical-necessity call)	156 · 100%	10.5281/zenodo.20564377	Funding-type obligation routing
private-capital-agent-audit	SEC-registered investment advisers	Best execution · MNPI · custody · marketing · allocation	Advisers Act §206	181 · 98.7%	10.5281/zenodo.20564496	Golden corpus of real SEC matters
cre-agent-audit	Commercial real estate	Tenant screening · lease provenance · PII residency	Fair-housing · FCRA §607(b)	336 · 86%	10.5281/zenodo.20437081	Fair-Housing Pre-Flight

All six: zero runtime dependencies · mypy --strict · SHA-pinned CI · golden corpora of real, primary-sourced enforcement actions. Released as a coordinated family.

---

The framework they all implement (A0 → A4)

Every library encodes the same five rungs of deployment authority. The decision class changes by vertical; the rungs, the climbing rule, and the controls do not.

A4  Production Autonomous   ── writes across coordinating agents; monitor-led promotion + validated escalation
A3  Supervised Autonomous   ── writes one decision class autonomously; non-overridable sovereign veto + live ledger
A2  Delegated               ── writes inside a hard, mechanically-enforced envelope; sampled review
A1  Assisted                ── drafts; a human approves every write
A0  Informational           ── reads & recommends; no write authority
        ▲ climb one rung at a time, only on evidence ·  ▼ demote the moment assurance degrades (routine, not crisis)

Three controls make each rung a control and not a slide: a sovereign veto the agent cannot switch off, a hash-chain audit ledger (every action reconstructable for a safety case), and a mechanical demotion trigger.

---

Which library should I read first?

• A bank, lender, or fair-lending / model-risk team → banking-agent-audit (ECOA/Reg B adverse action) or finserv-agent-audit for cross-vertical.
• A payments company / fintech moving money on irreversible rails → payments-agent-audit (OFAC screening + the rail-finality gate).
• A health plan / payer (UM, prior auth, appeals) → payer-agent-audit (governs the audit record, not the medical-necessity call).
• An asset manager, allocator, or RIA → private-capital-agent-audit (Advisers Act §206).
• A CRE owner / proptech / screening platform → cre-agent-audit (fair-housing pre-flight, lease provenance).
• A frontier lab or cloud platform deploying agents into regulated customers → start with finserv-agent-audit; the A0→A4 control surfaces are domain-agnostic (see each repo's "Why this exists for frontier autonomy stacks").

---

How to read these (the 30-second version)

1. Open the library matching your decision class (above).
2. Read its "Read me first" pointer — the single most concrete regulator-mapped test + the worked example.
3. The test suite is the falsifiability proof — each qualitative rule is a runnable check.
4. The framework behind all six: autonomy-ladder.io.

---

Authored by Kunjar Bhaduri. Each library is MIT-licensed (one dual MIT OR Apache-2.0), DOI-archived for citation, and maintained as living research. Comments and critique welcome — open an issue on the relevant library.