The Register
Inspect a system’s class, test findings, conditions and reliance limits in one place.
Open the RegisterThis page is the app · A demonstration
See the limit.
Own the decision.
You are already in Plimsoll. It shows how far a bank may rely on one AI system for a chosen task.
There is nothing to log into or submit. Choose a task below. The dark panel then states the limit — and where a person still has to take over.
Start here. The four buttons are tasks. Click one. The dark panel is the answer for that task.
01 Choose a banking taskSample system: Northmere Investigate 4.2 A1 / In class
The Reliance LensEvidence → permitted use
Same system.A different boundary.
02 / Your reliance limit
L2
Small-business
accounts
accounts
Recommends.
You confirm.
The system may recommend. A person confirms before any action is taken.
Responsibility remains withYour institution.
Open this evidence record
03 / Follow the responsibility
The handover, made visible.
01SupportGather & organiseAI
02RecommendPropose a responseAI
03DecideConfirm the next stepPerson
04ActCarry out the decisionPerson
AI recommends. A person checks the recommendation before it becomes an action.
Illustrative workflow · not an accuracy scoreSpecimen / 0003 · AML alert dispositionAll systems, findings and outcomes in this demonstration are fictional.
01
Bank-funded. Built for those who rely on it.
02
Vendor-independent. Evidence before influence.
03
Continuously tested. Change must earn its place.
The Register / evidence you can inspect
Beyond a score.
A basis for a decision.
One AI system can be ready for one task and unsuitable for another. The Register makes that boundary explicit.
- 01See the test scope and findings.
- 02Read the conditions and vendor’s reply.
- 03Find the limit for the task in front of you.
A stronger foundation.
Above: the mark. Below: the proof.
Trust should be
traceable.
The proposed survey connects four things that belong together: access, performance, permitted reliance, and change.
01
Open the evidence.
State what could be examined and what was withheld.
02
Draw the boundary.
Assign the highest reliance the evidence supports for each condition.
03
Keep the mark earned.
Re-test releases and show when class or operating limits change.
One connected assurance model
Shared evidence.
Local accountability.
Your bank keeps the decision. Plimsoll proposes the common reference.
The Overlay
Review how the surveyed system fits your bank’s configuration, data and controls.
Explore the proposed serviceThe Watch
Re-test releases and identify changes that affect class or move the load line.
See how re-testing worksIndependence, by design
Funded by the institutions
that rely on the answer.
The proposed model excludes funding from the vendors being surveyed. One bank, one vote. Findings and the vendor’s reply appear together.
Read the independence principlesThe same argument, four ways
Watch, listen,
then look.
This is still the app. The media room holds the film, the podcast, four pictures, and the slide deck.
04 / DeckThirty platesFrom possibility to evidence
Find the line.
Then make the decision.
See the evidence records See the same limits in 3D →
A business concept by Richard Leclezio.
Illustrative systems. Proposed services. This page is the demonstration.
The media room
The line,
shown.
A film to sit with. A voice to follow. Four pictures to read whole. Thirty plates you can turn.
A film by Richard Leclezio. Open it on YouTube.
Listen.
The Plimsoll Line for Banking AI. Twenty-one minutes on a visible limit for reliance, not a score for intelligence.
The Plimsoll Line for Banking AI
21:380:00 · 21:38
Plays in this page. An explanatory recording, not a live survey.
Read them whole.
One picture at a time, uncropped. The drawings are concept illustrations. The words painted on them are not live findings.
01 / Load line
The maritime mark
A defined use, a visible boundary, and a register that more than one bank can read.
Turn the plates.
Thirty slides. Drag, use the arrows, or choose a plate from the strip. The left and right keys work when the stage is selected.
A concept deck. It is not a live assurance report.
01 / THE RELIANCE LANDSCAPE
One system.
Different limits.
Trust is specific to the task. Explore how the permitted level of reliance changes across four operating conditions.
OPERATING ENVELOPE / 3DSPECIMEN DATA · NOT A RISK SCORE
Drag to rotate · select a condition
Northmere Investigate 4.2
A1
In class
AML alert disposition
Reliance by operating condition
Selected condition
May act without review.
Domestic retail accounts, within the surveyed scope.
Height represents the permitted reliance tier—not accuracy or confidence. Names and results are fictional.Read the evidence record ↗
THE REGISTER / SPECIMEN EDITION
Evidence, in plain sight.
Explore nine illustrative records. Select a system to see what was tested, what remains uncertain, and where reliance should stop.
09 specimen systems04 in class03 with conditions01 suspended01 awaiting survey
Select a row to open its evidence record. All names and results are fictional.
| No. | System and surveyed use | Class | Status |
|---|
A, B, C: what we could see. A, documentation, evaluation data, a staging endpoint and the change log. B, some components withheld by the vendor. C, surveyed from outside through a subscriber's licensed instance.
1, 2, 3: how it performed. 1, met every criterion in the Rules for the stated use. 2, met them with conditions of class open. 3, material shortfall; class is suspended until it is remedied.
L3 to L0: how far to rely on it. L3, may act without prior review. L2, recommends and a person confirms. L1, assists while a person does the task. L0, not for use under that condition.
Specimen entries. Vendor names, product names and results are invented. January 2027 is the proposed launch date.
THE SURVEY METHOD / PROPOSED
From a question
to a defensible limit.
Concept illustration.
Nominated
Three or more subscribing banks name a system and the use they have in mind. One bank, one vote, whatever its size.
Access agreed
The vendor provides documentation, a staging endpoint and release notes. If it declines, we survey from outside, and the entry says so.
Tested against the Rules
A harness built for that banking use: task corpora with ground truth set by practitioners, stress and injection tests, change-control checks.
Load line assigned
For each operating condition, the highest reliance the evidence supports. Where we did not test, the mark reads not for use.
Entered in the Register
Class, load line, findings, and the vendor's unedited reply of up to 300 words. Subscribers get the full survey file.
Kept in class
Re-tested on every release, reviewed each quarter against pooled outcomes, fully re-surveyed if the underlying model changes.
WHO PAYS / WHO DECIDES
Independence is
the starting condition.
The institutions relying on the work fund it. The companies being assessed do not. That separation is the foundation of the proposed model.
Pays
Banks and other regulated institutions that rely on the Register.
Cannot pay
Vendors, model developers and their controlling investors. No fees, no sponsorship, no equity.
Written into the charter
The proposed structure is a Delaware public benefit corporation. The founding charter would prohibit funding from surveyed parties, with the enforceability and amendment protections reviewed by counsel.
Public Rules
The Rules are free to read and change only after a public comment window. Vendors may comment. Subscribers and independent technical members vote.
Evidence on the record
Every class decision publishes what was tested and what was not, beside the vendor's unedited reply.
One bank, one vote
The survey queue follows subscriber nominations counted equally, so the largest fee cannot steer it.
No revolving door
Surveyors may not join a vendor they have surveyed for twelve months, and rotate off a system after three years.
Classification began this way. In 1760 the customers of Edward Lloyd's coffee house, the merchants and underwriters who needed to know which ships were sound, formed a society to survey them. Subscriptions to its Register Book paid the surveyors. Owner-pays came later. We are going back to the original arrangement.
MEMBERSHIP / PROPOSED COMMERCIAL TERMS
One subscription.
The whole bank.
A shared Register for model risk, third-party risk, compliance, audit and the business. Pricing follows total bank assets.
$10–30bn in assets
Regional
$95,000/ year
A common evidence base for a growing AI portfolio.
$30–100bn in assets
National
$190,000/ year
Shared assurance across teams, systems and use cases.
Over $100bn in assets
Global
$380,000/ year
Institution-wide access for complex AI estates.
Every membership includes
Every Register entry and its survey file. The Watch on systems you rely on. Two annual nominations. A vote at the Rules review. Unlimited readers.
The Overlay $22,000 / system
Your use, configuration, data and controls checked against the load line. Proposed delivery in three weeks, with an included refresh if the line moves within the year.
A founding cohort of eight banks.
The proposed founding programme gives early members a role in the first Rules and the first survey queue. Target opening: January 2027.
- Term. Twenty-four months, paid in advance.
- Price. 25% below list for the whole term.
- Nominations. Three systems each into the first survey queue.
- Rules Committee. A seat for the first three years.
FOUNDER’S FILE / THE THESIS
A changing rulebook. A persistent need for evidence.
Three regulatory developments inform the proposed business. The opportunity is a founder’s hypothesis, not a regulatory endorsement.
17 Apr 2026
The agencies issued revised model-risk guidance that supersedes SR 11-7. Generative and agentic AI are outside its scope. That exclusion does not remove banks’ wider risk-management responsibilities.
OCC Bulletin 2026-13; Federal Reserve SR 26-21 May 2026Publication date
The Fed’s Vice Chair for Supervision discussed how existing risk frameworks apply to AI and the challenges of third-party tools. Plimsoll’s proposed response is shared, use-specific evidence for the relying banks.
Bowman, remarks to the FSOC roundtable on AI1 May 2027
OSFI’s revised E-23 guideline takes effect on 1 May 2027. Its enterprise-wide, risk-based approach includes AI and third-party models at federally regulated financial institutions. This date informs the proposed launch plan.
OSFI Guideline E-23 (2027)The gap
A bank can obtain a certificate that a vendor runs a tidy management system, and another that the vendor's agent passed a battery of adversarial tests. The vendor buys both, picks the assessor and sets the scope. Neither says how the system performs on the bank's actual task, or where reliance should stop. So each bank works that out alone, from outside, with whatever the vendor chooses to show one customer.
The bet
When supervisors stop prescribing, the burden of proof moves to the bank's own judgment, and judgment wants a reference point that boards, auditors and examiners all recognize. Shipping solved this in 1760 with a private body that surveys once for everyone who relies on the result. The bet is that banking will buy the same thing for AI, and that whoever writes credible Rules first will shape what sound practice comes to mean.
FOUNDER’S FILE / THE LANDSCAPE
Different evidence answers different questions.
| What a bank can get today | What it examines | Who pays | Who sets the scope | Sees results in production | Written for a banking use |
|---|---|---|---|---|---|
| SOC 2 report | The vendor's security and availability controls | Vendor | Vendor | No | No |
| ISO/IEC 42001 | The vendor's AI management system | Vendor | Vendor | No | No |
| AIUC-1 | An agent's behavior under adversarial and operational tests, re-run quarterly | Vendor | Vendor picks the agent and the auditor | No | No |
| Shared vendor-risk utilities | Control questionnaires and evidence, pooled across banks | Banks | A standard questionnaire | No | Financial services, controls only |
| The bank's own review | Whatever the vendor will show one customer | The bank | The bank | Its own only | Yes |
| Plimsoll | The system's performance and limits in a named banking use | Relying banks | Subscriber votes and public Rules | Pooled across subscribers | Yes |
A surveyor treats all of these as inputs, and reads a vendor's SOC 2 report, ISO/IEC 42001 certificate and AIUC-1 results the way a ship surveyor reads type approvals: gladly, before testing the thing itself. Independent commentary on the newest of them makes the same point about who chooses scope and auditor.
Banks have paid to share this kind of work before. TruSight, founded in 2017 by American Express, Bank of America, JPMorgan and Wells Fargo to pool vendor control assessments, was bought by S&P Global in 2023. Its focus on shared control assessments is a useful precedent for the proposed performance-survey model. The 2023 interagency third-party guidance already lets banks lean on shared utilities for due diligence while keeping responsibility for the outcome.
FOUNDER’S FILE / SHARED ECONOMICS
The whole company is one fixed cost, spread over everyone who relies on it.
A bank assessing one vendor AI system by itself spends about $79,000 of second-line time and sees only what the vendor shows a single customer. Plimsoll spends about $137,000 in the first year on the same system, sees far more, and divides the bill.
$6,842Plimsoll's cost per relying bank, first year
$79,200Each bank working alone
91%Less cost per bank, for deeper evidence
Cost per bank of assessing one vendor system, by number of banks sharing the survey
What the bank pays is a different number, and the gap between the two is the margin. A $50 billion bank relying on 25 entries pays $190,000 for the Register, about $7,600 an entry, plus $22,000 for each Overlay. That is roughly $30,000 per system in use against $79,000 alone.
Assumptions. Alone: two people for six weeks at a loaded $165 an hour. Plimsoll: a surveyor at $270,000 loaded completes 3.5 new surveys a year and keeps 12 in class; $30,000 of test infrastructure and practitioner time per new survey; $600 a month to re-test. These are illustrative operating assumptions.
FOUNDER’S FILE / FINANCIAL MODEL
Five years, month by month.
Explore the proposed business over sixty months. Change acquisition, usage or price to see revenue, cash and staffing respond. All figures are planning assumptions.
Case
Revenue by line, and EBITDA, $ million
Cash at each month end, $ million
Month 1 is January 2027. Founding members prepay 24 months at 25% below list; everyone else pays a year in advance; 6% of contracts are lost at each renewal in the base case. Surveyors and Overlay analysts are hired as volume requires; the other teams follow a plan. Tax and interest are left out. Funding is fixed at a $4.5 million seed and a $9 million Series A in month 15, so the cash line shows what each case does to the same money.
FOUNDER’S FILE / THE MARKET
A toll bridge, not a land grab.
About 160 US banking organizations hold more than $10 billion in assets. At list price and ten Overlays a year, they are worth some $64 million a year. Add Canadian, British and European banks of the same size, insurers and the largest credit unions, and the reachable market is about $146 million.
| US banking organizations by total assets | Count | Register, list | Register value a year |
|---|---|---|---|
| Over $100 billion | 35 | $380,000 | $13.3m |
| $30 to $100 billion | 40 | $190,000 | $7.6m |
| $10 to $30 billion (estimate) | 85 | $95,000 | $8.1m |
| Overlays, ten a year each at $22,000 | 160 | $35.2m | |
| US banks over $10 billion | 160 | $64.2m |
That is small for a venture fund and large for a business with 65% gross margins, contracts paid in advance and no competitor able to match its independence. The base case takes 17% of it by 2031. What grows is the number of vendor AI systems inside each bank, and every one of them needs an Overlay.
Counts over $30 billion are from FFIEC holding-company data at 31 March 2026, plus three large banks without a holding company. The $10 to $30 billion count and all adjacent pools are estimates and require validation before investment decisions. The FDIC reported 4,238 insured institutions at 30 June 2026.
FOUNDER’S FILE / GO TO MARKET
Start with a concrete decision and a date.
Begin with institutions preparing for OSFI E-23 on 1 May 2027. Validate demand, buying authority and applicable group requirements through discovery conversations.
Thirty conversations in six weeks
Heads of model risk and third-party risk at the six large Canadian banks and a dozen US regionals between $50 and $250 billion. One question: which five vendor AI systems are you being asked to approve this year? The overlap between their answers is the first survey queue, and the proof that the survey can be shared.
Sell the founding charter first
Eight founding seats, from $142,500 to $570,000 for 24 months depending on size, paid up front. Small against a second-line budget. The seat on the Rules Committee is what a head of model risk actually wants: a say in what good looks like before somebody else decides.
Survey what they named
Ten systems in year one, chosen by the founders' votes. Expect financial-crime investigation agents and contact-center agents to lead, because that is where banks are deploying first.
Put the Survey Clause in their contracts
Each founding member adds one paragraph to its next vendor agreement. Once twenty banks have it, access stops being a favor.
File the Rules where supervisors will read them
Submit Rules 1.0 in answer to the agencies' request for information when it is issued, and brief OSFI before E-23 takes effect. Ask for nothing. Be the reference that others cite.
The chief risk officer signs. The heads of model risk and third-party risk champion it because it takes work off their desks. The chief AI officer wants it because approvals stop being the bottleneck. Procurement resists the clause until the second vendor accepts it.
FOUNDER’S FILE / THE SURVEY CLAUSE
The moat is a paragraph in a contract.
Ships are classed because charter parties and hull insurance policies say they must be. The equivalent here is a clause in the bank's vendor agreement.
The Survey Clause
Supplier agrees that Customer may appoint an independent classification body to survey the System's performance and limits of use, and will give that body reasonable access to documentation, a staging endpoint and release notes for that purpose. Supplier will notify the body of any material change to the System, including any change of underlying model, before release. Customer may share the body's findings with other financial institutions that rely on the same survey. Supplier may not pay the body, and may publish a reply of up to 300 words alongside any finding.
A draft to take to counsel. It is not legal advice.
Pooled outcomes
Every subscriber returns production results for the systems it uses. Drift that no single bank could see shows up across twenty. The proposed advantage is pooled outcomes across relying banks.
A harness for each banking use
The second AML investigation agent costs a fraction of the first to survey. By year three the library of harnesses is the cost advantage.
Independence nobody can copy cheaply
Audit firms, certification bodies and ratings businesses all take money from the companies they assess. They would have to give that up to match this.
FOUNDER’S FILE / RISKS & STOP RULES
What could kill it, and when to stop.
| Risk | Why it might happen | What we do about it | Stop or reshape if |
|---|---|---|---|
| Banks will not fund a shared good | Everyone benefits from a public mark, so everyone waits for someone else to pay. | Class status is public. The evidence, the load line detail and the Watch are for subscribers only. | Fewer than four founding members by day 90, or fewer than eight subscribers by month 9. |
| Vendors refuse access | Fear for intellectual property, and contract terms that forbid benchmarking. | The Survey Clause, surveys from outside through members' licensed instances, and a printed right of reply. | Under half of nominated vendors give access by month 12 and outside surveys cannot get beyond class C. |
| Supervisors prescribe, or bless another scheme | The request for information ends in detailed rules or a named standard. | File the Rules in that record. Map to whatever is issued. Accept other schemes' evidence as inputs. | A mandated scheme makes private class redundant. Then become a surveyor under it. |
| A classed system fails badly at a member | It will happen. Every assurer meets this eventually. | Class states limits and is not a warranty. The load line is on the record. Indemnity cover, contractual caps, a published review after every incident. | Two failures inside a stated load line in 24 months that the tests should have caught. Halt new surveys and rebuild the method. |
| Systems change faster than surveys | Weekly releases, and silent swaps of the underlying model. | Automated re-tests on every release. Class lapses after 90 days without one. Swaps trigger a full re-survey. | Median time from a vendor release to a re-test exceeds ten working days for two quarters. |
| A bigger firm copies it | The audit networks, a ratings business or a certification body decides to enter. | They are paid by vendors elsewhere. The Survey Clause and pooled outcomes compound with every member. | A credible bank-funded rival signs five of the ten largest banks first. Then merge or sell. |
FOUNDER’S FILE / FIRST 90 DAYS
The first ninety days.
Week 1 begins Monday 28 September 2026. January 2027 is the proposed opening date, subject to the launch gates below.
1
2
3
4
5
6
7
8
9
10
11
12
13
Company
Incorporate; covenants in the charter
Bank regulatory counsel; indemnity cover quoted
Close the seed
Members
Thirty discovery conversations
Charter terms out; letters of intent
Sign founding members
Rules
Draft Rules 0.1: scope, performance, load line
Publish for comment
Comment window; brief supervisors in Washington and Ottawa
Harness
AML alert disposition: 2,000 cases, ground truth set by two former investigators
Contact-center servicing agent
Surveys
Vendor access for two pilot systems
Two pilot surveys
First entries
People
Chief Surveyor search
General Counsel search
Offers: three surveyors, three engineers
Day 30
Ten banks have named their systems and at least five systems appear on more than one list. If they do not overlap, the shared-survey premise is wrong for AI. Stop.
Day 60
Two vendors have agreed access and Rules 0.1 has drawn written comments from five banks. If vendors refuse, run both pilots from outside and price the difference.
Day 90
Four founding members signed, two entries in the Register, seed committed. With fewer than four signatures, do not raise. Reshape or stop.
FOUNDER’S FILE / TEAM & CAPITAL
Who builds it, and with whose money.
Founder and chief executive: Richard Leclezio. A career in financial-services change and governance across Deutsche Bank, Credit Suisse, Citi and Mizuho. Year one is the Rules, the founding members and the supervisors. The buyers are the people he has spent two decades working beside.
- 1
Chief Surveyor
Has validated machine-learning models inside a large bank and built evaluation harnesses for language models. Owns the method and every class decision.
- 2
General Counsel and Secretary to the Rules Committee
Bank regulatory counsel, ideally with time at an agency. Owns the charter covenants, the Survey Clause and liability.
- 3
Three surveyors
Financial-crime operations, consumer compliance, and credit. They set the ground truth the harnesses are scored against.
- 4
Three engineers
Harness, Register and Watch. Simple for members to read, reproducible underneath.
- 5
Member lead
A former head of model risk or third-party risk who has bought this kind of thing and knows how it gets approved.
- 6
Overlay analyst, and a finance and operations lead
Twelve people at the end of 2027. Fifty-two at the end of 2031 in the base case, more than half of them surveyors and analysts.
Capital
- Seed, January 2027$4.5m
- Founding members' prepayments during 2027$1.5m
- Series A, by March 2028$9.0m
- Lowest month-end cash, February 2028$1.47m
- EBITDA-positive fromMarch 2030
- Equity needed after the Series A, base caseNone
Whose money it can be
No investment from any company Plimsoll may survey, from model developers, or from their controlling investors. Bank investors are welcome, capped at 9.9% each and 25% together, without board control. A members' trust holds a veto over any change to the independence covenants.
Where it ends: ratings and benchmark businesses buy assets like this, as S&P Global did with TruSight. Classification societies already sell assurance and may want a banking arm. Or nobody buys it. The base case throws off cash from 2030, and a utility owned partly by its members is a respectable place to stop.
FOUNDER’S FILE / THE NAME & MARK
Why it is called Plimsoll.
Samuel Plimsoll was the member of Parliament who fought for a mark on the side of every British merchant ship showing how deep it could safely be loaded. The Merchant Shipping Act of 1876 made the mark compulsory. His line did not make ships safe. It made overloading visible to anyone standing on the quay.
That is the product: a plain mark of how far a system may be relied on, that anyone can read. The disc and line are the historical mark. The letters either side name the body that assigned it. The ladder gives the limit for each condition, from tropical fresh water to winter in the North Atlantic. Ours reads from L3 to L0.
- Say what was tested, and what was not.
Enterprise-grade, battle-tested AI assurance.Tested on 2,000 retail alerts in English. Not tested on correspondent banking. - Numbers carry a unit and a date.Recall 94.1% on the Rules 1.0 test set, 14 March 2027.
- No adjectives about ourselves.The Register either has the entry a bank needs or it does not.
Two materials: the hull and the Register book
Hull
#EEF2F1
#EEF2F1
Ink
#0B2530
#0B2530
Oxide
#9E3522
#9E3522
Tide
#1C6E78
#1C6E78
Deep
#061A22
#061A22
Foam
#E6EEEC
#E6EEEC
Entered in the Register
Libre Caslon, for the book. Caslon's types set most of what London printed in the 1760s.
A1 L3 0017 WNA
Archivo at its narrowest width, for anything painted on steel: class, levels, figures.
Working name. No trademark clearance has been done.
FOUNDER’S FILE / SOURCES & ASSUMPTIONS
Sources, and what is assumed.
The research references below support the original business thesis; confirm their current application before relying on them. Every price, volume, cost and date about Plimsoll itself is an assumption, and should be validated before launch.
- OCC Bulletin 2026-13, Model Risk Management: Revised Guidance17 April 2026. Scope, the $30 billion relevance threshold, the promised request for information.
- Federal Reserve SR 26-2Supersedes SR 11-7 and SR 21-8.
- Sullivan & Cromwell memo on the revised guidance29 April 2026. The lighter treatment of vendor models; no fixed validation cadence.
- Vice Chair Bowman, Artificial Intelligence in the Financial SystemPublished 1 May 2026; delivered 27 April 2026.
- Interagency Guidance on Third-Party RelationshipsJune 2023. Utilities and consortiums may supplement a bank's due diligence.
- OSFI Guideline E-23, Model Risk Management (2027)Final 11 September 2025; effective 1 May 2027.
- FDIC Quarterly Banking Profile, second quarter 20264,238 insured institutions.
- Largest US bank holding companies, 31 March 2026FFIEC data as tabulated; used for the counts above $30 billion.
- S&P Global acquires TruSightJanuary 2023.
- Lloyd's Register Foundation, eighteenth-century Register BooksThe 1760 society, and subscriptions funding the surveyors.
- What to Make of AIUC-1April 2026. Who chooses the scope and the auditor.