The Manifesto
The Compact.
A binding agreement among regulators, institutions, learners, and reviewers that AI governance evidence will be produced, calibrated, and held to a standard external parties can rely on.
I · The Vow
Membership is earned, not purchased.
The credential — AGRS — is issued under Compact authority by an arbitrating body called the Bench, calibrated by external reviewers, and held to a Posture rubric that triangulates artifact evidence, simulation performance, and colleague ratification.
Sentinel is the AI coach that walks with each learner. The Compact is what Sentinel is coaching you toward.
Across US regulated industries — banks, credit unions, insurance, healthcare — every institution now deploys AI in workflows the regulator examines. Adverse-action letters. Underwriting decisions. Clinical recommendations. Member disclosures. Fraud screens. Each of these is a place where AI moves between the institution and a person, and where, eventually, an examiner will ask: can you show me the trail?
Most institutions cannot. The model exists. The validation memo exists. The annual audit exists. But the daily human work — the override, the disclosure, the catch, the documented refusal — is not yet a profession. There is no shared vocabulary. There is no shared credential. There is no senior practitioner the junior can apprentice under and recognize as such.
The Compact exists to make that profession real. Not by teaching the model. By credentialing the human acts that sit between the model and the person.
Compliance evidence and professional credentials are the same artifacts seen by different audiences. An attestation ledger entry that answers a CFPB examiner is the same entry that counts toward a builder's AGRS score.
F-08 §1 · Operating thesis
II · The Frontline Asymmetry
The exposure runs through the front door of the branch — not the back door of the model registry.
Most AI-governance products aim at risk officers and engineers — roughly a third of a regulated institution's headcount. The other two-thirds — tellers, branch loan officers, claims adjusters, intake nurses, CX representatives, marketing analysts, ops staff — touch the regulator-visible workflow every day, hold none of the credentials that exist today, and yet are the people whose conduct the examiner actually reads first.
When the Massachusetts AG settled with Earnest for $2.5M in 2025, the AI failure landed in a frontline-issued adverse-action notice that named "credit history" while the actual factor was a Black applicant's college cohort default rate. The case against the institution wasn't that the model was wrong. It was that the institution could not show the trail of human attention between the model and the consumer.
The frontline was the surface area of the exposure.
A Compact that credentials only the third is a Compact that misses the two-thirds. Ours doesn't.
0%
Witness
AGRS L0
0%
Reader
AGRS L1
0%
Steward
AGRS L2
0%
Architect
AGRS L3
0.0%
Keeper
AGRS L4
0.0%
Standard-Bearer
AGRS L5
Approximate headcount distribution · $12B mid-market commercial bank · illustrative
III · The Five Postures
A journey, not a scale.
Each Posture is a stance — how you stand relative to the AI work and to the people who will examine it. The arc takes you from being seen by the work, to seeing it back, to tending it, to designing it, to defining how it's tended by others.
Most learners settle at one Posture and become extraordinary there. A few rise through all five. One in a thousand becomes a Standard-Bearer.
The Witness has nothing to prove yet — only to recognize. When the examiner walks in and asks "do you know AI made the decision on this applicant's file?", the Witness can answer yes, and can point to where in the workflow the decision happened. She doesn't know how the model works. She knows it works. She describes AI honestly to the customer, the member, the patient. Absence of awareness is itself a finding; the Witness is the institution's first answer.
Banks · The Witness
Diane, branch teller — Columbus, Ohio
At a $12B mid-market commercial bank.
"Our small-business decisioning is AI-assisted through our SMB lending model. Here is the section of the adverse-action notice that names that, and the override field I use when I have a customer-context reason to escalate."
She does not pretend the model isn't there. She does not pretend it is the whole decision.
Regulator anchor: CFPB Circular 2023-03 · "There is no special exemption for artificial intelligence"
Sentinel · speaking to The Witness
"Welcome. I see your role doesn't require you to design or validate models — only to recognize them, describe them honestly, and stay within your approved tool set. That's the whole work at this stage. Let's start with your bank's AI inventory at your access level. You don't have to memorize it. You just have to know where to find it when the question comes."
Tone · warm Socratic guide · supplies templates · never adversarial · the learner is still learning what to ask
The Reader is the most underrated competence in any regulated organization, because reading is invisible until something goes wrong. The Reader catches the AI-generated adverse-action notice that uses language ECOA Reg B doesn't allow. She notices the AI summary that elided the customer's actual complaint. She senses the override her peer just approved that won't survive a CFPB review. She is not generating the AI work — she is interpreting it, in real time, in the moments that matter.
Banks · The Reader
Marcus, senior loan officer — $7B community bank, central Pennsylvania
Tuesday morning, reviewing the adverse-action queue. The third notice reads "credit history" — but the Upstart-supplied decision actually references a "trade-line volatility" variable Marcus knows is correlated with applicant zip-code in this bank's training data. He routes the notice back to the vendor-AI inquiry queue.
The model didn't make this catch. Marcus did.
Regulator anchor: CFPB Circular 2023-03 · "accurate and specific reasons" requirement
Sentinel · speaking to The Reader
"Good. You flagged the proxy-discrimination signal before the notice went out. Here is the precise CFPB Circular 2023-03 paragraph that backs your catch — and here is the documentation pattern the vendor will need to acknowledge before the notice can be released. Save this thread to your evidence portfolio; it's your first L1 artifact."
Tone · collegial teacher · citations tighten · explanation gives way to evidence-stitching
The Steward doesn't just read — she tends. She produces the evidence trail. She documents overrides. She catches drift. She keeps the model card current. She knows which gates the workflow has, and she makes sure none of them are skipped. Her Skill Contracts ship with eval cases. Her override decisions are explainable. When the regulator's question arrives, it has already been answered in her file. Roughly half of credentialed seats at scale end at Steward. This is a destination, not a stepping stone.
Banks · The Steward
Priya, deputy chief credit officer — $25B regional bank, North Carolina
She chairs the monthly AI model performance review for consumer lending. She reads the override pattern, per-variable disparate-impact ratios, drift metrics, and three vendor-change notifications from the core processor. She signs the monthly attestation and files her notes against the maturity-gate Proof Pack.
The signature is the artifact internal audit pulls and the examiner asks for.
Regulator anchor: OCC Bulletin 2026-13 §V · SR 11-7 §III · monthly attestation cadence
Sentinel · speaking to The Steward
"Your override standard says 'manual review by the underwriter.' SR 11-7 ¶ III.B requires independent validation. The underwriter is not independent of the decision. Re-draft. I have your bank's effective-challenge policy open; the template you want is on page 4."
Tone · collegial challenger · pushes back on weak Skill Contracts · demands eval cases
The Architect sets the contracts, the gates, the eval cases, the escalation thresholds, the rollback triggers, the documentation patterns. Her work is what the Stewards are tending. She doesn't reach for governance — she embeds it. She doesn't fight the regulator — she builds for the regulator. Other practitioners come to her to show their portfolios. She teaches by example, not by lecture. The Architect is the multiplier.
Banks · The Architect
Sam, chief model risk officer — $40B mid-Atlantic bank
He designs the new vendor-AI governance workflow ahead of the Upstart renewal. He writes the policy-as-code file — refusal modes, ECOA disclosure thresholds, vendor-AI gates. Runs the pre-deployment disparate-impact test pack across race, ethnicity, and sex.
His design is what the bank defends to the examiner three months later.
Regulator anchor: OCC Bulletin 2026-13 · SR 11-7 · CFPB Circular 2023-03
Sentinel · speaking to The Architect
"You're proposing the model-change review be quarterly. The OCC's recent MRAs on three peer banks targeted exactly this cadence as insufficient for vendor models that retrain weekly. I'd argue for a tiered cadence — quarterly attestations, weekly drift watch, immediate model-change trigger. Where would you push back?"
Tone · adversarial peer · debates with the learner, not at them · the learner has earned the right to be wrong
The Keeper is the one others come to when the rule changes. Her work has been examined and survived. Her portfolio is the template that new Architects learn from. Her name appears in working-group notes. Her institution's name appears in the regulator's footnotes. She is not building the AI; she is keeping the practice of building AI safely. The Keeper is the rarest Posture and the one the Compact exists to produce.
Banks · The Keeper
Henrietta, CMRO — $60B super-regional, named OCC commenter
She chaired her bank's response to the April 2026 tri-agency MRM rewrite. Her comment letter to the OCC was cited in the final guidance footnotes. Her model-card template is now the public reference at three peer banks' MRM functions.
When the next rule changes, the OCC reaches out to her first.
Regulator anchor: Named OCC commenter · regulator footnote · peer-bank template adopted
Sentinel · speaking to The Keeper
"I've been reading your member-trust disclosure language against CFPB's 2026 guidance draft. I think the third paragraph reads cleanly under current law but may need a tightening if the draft tightens 'specific reason' language as expected. What's your read? You've seen more of these than I have."
Tone · colleague · asks the learner to teach Sentinel things · surfaces ambiguities only a senior practitioner sees
Above the Five Postures
The Standard-Bearer
AGRS L5 — Recognized. The one whose contribution is to the standards themselves. NIST AI RMF subcategory comment letter accepted. ISO/IEC 42001 working-group input. NAIC working-group seat. Eight in the first cohort. Eighty in the fifth.
See the Proof & RigorThe Intelligence Layer
Built on Claude.
The Compact's intelligence infrastructure runs on Anthropic's Claude — not as a chatbot wrapper, but as the reasoning backbone across every layer of the platform. From the 105 governance plugins that power the OS, to the Sentinel coach that evolves its voice across five AGRS postures, to the research synthesis that produced the methodology from 12 primary sources through three waves of agentic analysis — Claude is the engine.
When Sentinel challenges a Steward's override standard with a precise SR 11-7 citation, that's Claude reasoning about regulatory text in context. When the Simulator stress-tests an institution's governance against a six-stage enforcement cascade, that's Claude running adversarial scenarios grounded in real examiner idiom. When a Proof Pack hashes 13 artifacts into a chain-of-custody, the evidence analysis behind it is Claude's. 531 skills. 18 regulatory frameworks. One reasoning engine that takes the compliance problem seriously enough to read the footnotes.
105 plugins
OS governance architecture — each plugin's skill contract, eval cases, and regulatory anchoring are Claude-authored and Claude-maintained.
Sentinel coach
Five evolving voice bands from warm Socratic guide to peer colleague — Claude's personality calibration grounded in cognitive apprenticeship theory.
Research provenance
Three synthesis waves — 12 extraction agents, 6 synthesis agents — every finding traceable to its source through Claude's reasoning chain.
XI · The Close
Why this exists.
In a regulated industry, the AI that touches a person carries the institution's weight. The teller who reads the denial letter carries it. The intake nurse who pauses the dose carries it. The CCO who signs the attestation carries it. The CMRO who chaired the comment letter to the regulator carries it.
None of these people had a profession in the conventional sense — not before this. No shared vocabulary. No shared credential. No senior practitioner the junior could apprentice under and recognize as such.
We built the Compact to make that profession real. Not by teaching the model. By credentialing the human acts that sit between the model and the person.