The gate record template, G1 to G5.
Chapter 5 of OSFI E-23 for AI Systems sets out every field of this template in its text. The file is a convenience; the method is in the book. Every field is below, before anything is asked of you.
Every field, before the form.
Where a field cites E-23, the guideline states it. Where it says the discipline’s, the handbook’s method chose it, inside what E-23 permits. Open any part to see its fields.
Part 1. Gate holders
Name each holder in advance, with a deputy, so that the gate cannot be bypassed on the day the person is away (Chapter 5).
| Gate | E-23 lifecycle component or heading | E-23 role the discipline narrows | The discipline's accountable role | Named person | Deputy |
|---|---|---|---|---|---|
| G1 Data and Design | Model design, including rationale, data and development | Model Owner and Model Developer, each a unit or an individual | The named data owner | ||
| G2 Validation | Model review, independent of development | Model Reviewer | The named validator | ||
| G3 Approval | Model approval, a heading that occurs throughout the lifecycle | Model Approver, a unit, an individual or a committee | The named accountable executive | ||
| G4 Deployment | Model deployment | Responsibilities set in the institution's documented deployment procedures | The named platform owner | ||
| G5 Operation | Model monitoring | Escalation procedures the institution defines | The named head of the governance office | ||
| Retirement | Model decommission | Stakeholders alerted under a disciplined process | The accountable person at G3, or a successor in the role |
Separations. Two hold at any size; the rest are the discipline's placements (Chapter 5). Everything else may combine: the data owner at G1 may be the platform owner at G4; where there is no governance office, the executive who signs at G3 may hold the G5 decision.
| # | Check | Basis | Holds (yes / no) |
|---|---|---|---|
| S1 | Whoever developed the system does not review it | E-23 (review independent from model development) | |
| S2 | The person accountable at G2 is never the person accountable at G3 for this system | The discipline's | |
| S3 | G2 is not placed in the first line | The discipline's; E-23 assigns no model risk role to any line of defense | |
| S4 | G3 is not held by whoever built the system | The discipline's | |
| S5 | A Chief AI Officer who owns the platform does not hold G5 for systems that platform runs | The discipline's | |
| S6 | Any deputy drawn from another gate crosses no separation above | The discipline's |
Part 2. The record, one per gate
A gate requires six elements, and a checkpoint lacking any one is a review, not a gate: entry criteria, an accountable role, a decision, exit criteria, an evidence record and a rejection path (Five-Gate). Every gate writes one record with fixed minimum fields (Five-Gate).
| # | Field | Basis | Entry |
|---|---|---|---|
| R1 | System, and the version of it that passed | Five-Gate (Chapter 5) | |
| R2 | Gate | Five-Gate | |
| R3 | Decision: accept, reject, or accept with recorded conditions | Five-Gate | |
| R4 | Conditions, each dated and owned by a named person | Five-Gate | |
| R5 | Open conditions from earlier gates, shown to this gate's accountable person before deciding | Five-Gate | |
| R6 | Name of the person who took the decision | Five-Gate; one person is the discipline's choice inside | |
| R7 | Entry evidence relied on, identified so that it can be retrieved (Part 3, by criterion) | Five-Gate | |
| R8 | Policy version in force at the time | Five-Gate | |
| R9 | Date and time | Five-Gate | |
| R10 | Exit criteria: what the system may do once through | Five-Gate | |
| R11 | Rejection path: where a rejected system goes, and what must change before it returns | Five-Gate | |
| R12 | Boundary fixed | The declaration shape (Chapter 6) | |
| R13 | Optimizer freed | The declaration shape (Chapter 6) | |
| R14 | Evidence that the boundary held | The declaration shape (Chapter 6) |
Two rules for entry criteria (Five-Gate, Chapter 5). Entry criteria name artifacts that either exist or do not; policy, training, intent and scheduled work do not satisfy one. Where a criterion cannot be met and the institution intends to proceed anyway, the gate is recorded as passed with conditions, each condition dated and owned by a named person. An unrecorded exception is a skipped gate.
R12 to R14 by gate. Chapter 6 reads the five records in the declaration's shape as follows.
| Gate | Boundary fixed | Optimizer freed | Evidence |
|---|---|---|---|
| G1 | Unclassified, unpermitted or unattested data may not enter the system | Choice of architecture, model and method over the permitted data | Classification, residency attestation, lineage and vendor clearance, cited in the record |
| G2 | A system whose behavior has not been independently examined may not be presented for approval | How the system achieves its result inside the validated envelope | The validation report, cited by identifier |
| G3 | A system nobody is answerable for may not enter production | Which systems the institution chooses to run at all | The approval record, naming a person |
| G4 | An approved configuration may not be altered on the way to production | Deployment mechanics, staging and infrastructure | The configuration cited in the record, the rollback test and the evidence layer's controls |
| G5 | A system that cannot be monitored against its approval conditions may not remain in production | Autonomy inside the conditions the approval set | The monitoring record and the armed triggers |
Part 3. Gate by gate
For each entry criterion record whether the artifact exists, and its identifier so that it can be retrieved.
G1 Data and Design
Purpose (Five-Gate): "To establish that the AI system may be built at all on the data and components proposed: that the data is classified and permitted, that its lineage is provable, and that any third-party component has been assessed."
| # | Entry criterion | Basis | Exists (yes / no) | Artifact identifier |
|---|---|---|---|---|
| G1.1 | A classification for every data set the system will use | Five-Gate | ||
| G1.2 | A residency determination for each jurisdiction | The discipline's addition, for a system that crosses borders | ||
| G1.3 | Lineage that reaches from an attested source to the point of use | Five-Gate; E-23: data traceable, "having documented lineage and provenance", rigor commensurate with the rating | ||
| G1.4 | Vendor clearance, where any component is third-party (the vendor assessment of Chapter 9) | Five-Gate; E-23: review may extend to third-party sub-components, including data and libraries | ||
| G1.5 | The deployment topology, where the system spans jurisdictions | The discipline's addition | ||
| G1.6 | A stated intended use, with the decisions the system will influence | Five-Gate; E-23: the inventory carries approved uses | ||
| G1.7 | Bias and privacy considered in the data and the rationale | E-23; placing them at G1 is the discipline's (Chapter 3) |
| # | Field | Basis | Entry |
|---|---|---|---|
| G1.8 | Domains consulted, where the system draws on more than one domain: one owner is named as accountable at Part 1, and the others are recorded here as consulted | The discipline's (Chapter 5) |
Exit, on passage. The system may be built, trained or configured against the classified data, and may not be exposed to any consumer.
Rejection. The system returns to design, with the record naming which criterion failed. A system rejected for an unclassified data set does not re-enter G1 until the classification exists rather than until it is scheduled.
Evidence. The record cites each entry artifact.
The G1 entry for a corpus (Chapter 8, for a retrieval-backed system). Chapter 8 adds no download of its own; its fields are carried here. Write one row for every data set the system retrieves from, by name.
| # | Data set, by name | Classification, including whether it holds personal information | The classes the retriever enforces at query time | Owner (one person) | Lineage from the document's authority to the index entry | Currency rule, and what removes a superseded document | Passages retrieved logged per call with the corpus version (yes / no) |
|---|---|---|---|---|---|---|---|
| G1.9 | |||||||
| G1.10 |
| # | Validation number | Value | Corpus version measured against | Date measured |
|---|---|---|---|---|
| G1.11 | Recall | |||
| G1.12 | Groundedness |
Where a line cannot be filled, the system has not entered G1, and that is the first finding of the readiness plan (Chapter 8). E-23's data principle asks that development data be accurate and fit-for-use, relevant and representative, traceable and timely; on the book's reading it reaches a corpus, because a corpus is data the system processes to generate results.
G2 Validation
Purpose (Five-Gate): "To establish that the system performs as claimed, under review by someone with no stake in it passing."
| # | Entry criterion | Basis | Exists (yes / no) | Artifact identifier |
|---|---|---|---|---|
| G2.1 | The system registered in the model inventory with an owner and a risk tier | Five-Gate; E-23: inventory fields | ||
| G2.2 | A record of data lineage, stated assumptions and stated limits, produced before validation begins | Five-Gate | ||
| G2.3 | An independent validation, performed by whoever the institution has placed outside development for the purpose | E-23: review "should be independent from model development", by internal reviewers or objective third parties; placing it in the second line is the discipline's |
| # | Field | Basis | Entry |
|---|---|---|---|
| G2.4 | The validation report, by identifier, with its overall recommendation on approval to the model approver | E-23 | |
| G2.5 | Where the validation finds the system fit for a narrower use than proposed: the narrower use, carried into the exit criteria | The discipline's (Chapter 5) | |
| G2.6 | Where rejected: returned to design (the failure was in the system) or to G1 (the failure was in the data) | Five-Gate | |
| G2.7 | The bias check, dated and scoped like every other finding | E-23; placing it at G2 is the discipline's (Chapter 3) |
E-23's reviewer recommends to the approver. Converting that recommendation into a gate decision, so that a system the validator will not accept never reaches G3, is a step the discipline takes and the guideline does not.
The four-part statement (Chapter 3): what the validation actually proved.
| # | Part | What it states | Entry |
|---|---|---|---|
| G2.8 | Object | Which build, which version, which model, which retrieval set, as of what date. If the version was not pinned, write that the object cannot be named | |
| G2.9 | Evidence | What was run, over what data, against what outcomes, by whom, independent of whom | |
| G2.10 | What was established | The criteria that were met, the population they were met for, the window | |
| G2.11 | What was not established | The properties the evidence did not cover, the changes it did not anticipate, the parts of the system that were not in the room |
The G2 evidence file for a system with no stable output (Chapter 3). The list belongs to the discipline's method; what E-23 asks for is the review, its extent set by the rating. Cite each by identifier.
| # | Item | What it holds | Identifier |
|---|---|---|---|
| G2.12 | The envelope and its pins | The model version, the prompt template and the retrieval index at the least, each pinned by version, with the rest of the pins as Chapter 13 lists them; a part left unpinned is named as such | |
| G2.13 | The versioned evaluation suite | Answers checked against references; groundedness and recall where it retrieves; inputs carrying an injected instruction through each of the doors the system has (the user's turn, retrieved content, tool output); pass thresholds written and dated before the first run; test sets, each versioned, among them a commitment set and a refusal set | |
| G2.14 | A miss rate with its interval for each statistical guardrail | The guardrail's version, the set it was measured on, and the interval that set's size supports | |
| G2.15 | The replay band | The fixed set replayed several times at validation, and the band its scores fall in | |
| G2.16 | The change table's G2 scope | Which pins may move on a passing run and which reopen G2, stated by the validator at the gate | |
| G2.17 | Grading by risk rating | How large the test sets are, and how wide an interval the validator may accept, set by the rating (the discipline's step; E-23) |
Exit, on passage. The system may be presented for approval, and its permitted use is the use the validation examined, and no wider.
Evidence. The record cites the validation report by identifier.
G3 Approval
Purpose (Five-Gate): "To establish that a person with institutional authority has accepted the system into production use, on the evidence, and is answerable for that acceptance."
| # | Entry criterion | Basis | Exists (yes / no) | Artifact identifier |
|---|---|---|---|---|
| G3.1 | The closed independent validation | Five-Gate | ||
| G3.2 | The approval package, stating the intended use, the validated limits, the residual risk and the open conditions | Five-Gate; E-23: approval affirms the risk rating and residual risk | ||
| G3.3 | The approval authority, assigned in advance: the decision-rights artifact of CADRE™, the AI Governance Operating Model, names the person who may approve before any system reaches the gate | Five-Gate; E-23: the rating drives the level of authority required to approve |
| # | Field | Basis | Entry |
|---|---|---|---|
| G3.4 | Suitability for production or continued use, for the intended purpose | E-23 | |
| G3.5 | Affirmation of the risk rating and residual risk | E-23 | |
| G3.6 | Known weaknesses or limitations, with the compensating mitigants in place or the stakeholder group's justification | E-23 (the guideline's form of accept with conditions) | |
| G3.7 | Approved use | Five-Gate | |
| G3.8 | Approved period, with an expiry date or a revalidation trigger | Five-Gate | |
| G3.9 | On the expiry date, by default: the system stops unless someone acts (a boundary), or continues unless someone acts (a reminder) | The discipline's test (Chapter 5) | |
| G3.10 | Who may extend the approval, and the authority the risk rating sets for an extension | The discipline's reading of E-23: an extension is an approval of continued use | |
| G3.11 | Where rejected: the reason; where it is a reason the validation did not examine, recorded as a finding against the validation's scope | Five-Gate |
OSFI's letter adds that approvals are obtained before a model change is implemented and after periodic reviews, with the institution choosing the stages (E-23).
Where a committee reviews. Nothing in the discipline forbids the committee E-23 permits; it cannot be recorded as the accountable role, because the question asked after a failure is which person is answerable, a narrower rule than the guideline's. The committee reviews, challenges and recommends; the executive named in the decision rights decides and signs (Chapter 5).
| # | Field | Basis | Entry |
|---|---|---|---|
| G3.12 | The committee's recommendation as voted | The discipline's (Chapter 5) | |
| G3.13 | Any dissent, with its reason | The discipline's | |
| G3.14 | The conditions the committee would attach | The discipline's | |
| G3.15 | The validation report, by identifier, as cited in the minute | The discipline's | |
| G3.16 | The accountable executive's record of the recommendation and any departure from it | The discipline's |
Who may loosen a boundary (Chapter 5). For each boundary the approval sets, write the two dates and the two names. A metric dated later than its boundary marks a review never held; a second name junior to the first marks a ceiling that can rise without the authority that fixed it.
| # | Boundary | Date it was set | Date the metric pressing on it was last retuned | Name of whoever set it | Name of whoever may extend or loosen it |
|---|---|---|---|---|---|
| G3.17 |
Exit, on passage. The system is approved for the stated use, for a stated period, subject to any recorded conditions, and the approval carries an expiry or a revalidation trigger.
Rejection. The system returns to G2 with the reason recorded.
Evidence. The record names the person. A record naming a committee is not a gate passage record (the discipline's rule).
G4 Deployment
Purpose (Five-Gate): "To establish that the approved system is being placed into production in the configuration that was approved, with its guardrails in force and its exposure bounded."
| # | Entry criterion | Basis | Exists (yes / no) | Artifact identifier |
|---|---|---|---|---|
| G4.1 | Release authority for the environment, which CADRE supplies as it supplies the approval authority at G3 | Five-Gate | ||
| G4.2 | Contractual controls in force, where any component is third-party (Chapter 9) | Five-Gate | ||
| G4.3 | The approved deployment topology, where the system spans jurisdictions | The discipline's addition | ||
| G4.4 | Guardrails configured and demonstrated in the target environment | Five-Gate | ||
| G4.5 | A tested rollback path | Five-Gate | ||
| G4.6 | The evidence layer's own controls in force (Chapter 13). An institution that only buys vendor APIs meets this with Chapter 9's minimum record kept under the same controls | The discipline's addition |
E-23 expects deployment to follow "clearly documented procedures that outline deployment steps, stakeholder responsibilities, approval hierarchies, change control, monitoring frameworks and exception handling including overlays", with production tests, consistent data and risk assessments before deployment.
| # | Field | Basis | Entry |
|---|---|---|---|
| G4.7 | The approved configuration, cited | Five-Gate | |
| G4.8 | The configuration matches what was approved (yes / no). If no, the system returns to G3. If yes and the environment is not ready, the system stays at G4 and the record says why | Five-Gate | |
| G4.9 | The rollback test: its identifier, the exposure it covered and its date | The discipline's (Chapter 5): each rollback test either covers today's exposure or visibly does not |
Staged exposure. Each stage carries its own entry condition, and the record states what triggers the next stage and what triggers a return (Chapter 5).
| # | Stage | Exposure | Entry condition | What triggers the next stage | What triggers a return |
|---|---|---|---|---|---|
| G4.10 |
The evidence layer's controls, confirmed for this system (Chapter 13). Every gate record is generated from the evidence layer, so its controls are platform work, in force before any system's first gate record is written; G4 confirms them for this system.
| # | Control | What it requires (Chapter 13) | Record confirming it for this system |
|---|---|---|---|
| G4.11 | Access | Read access set by class; write access held only by the emitting services, so that no person, administrators included, can add to a record directly or alter one | |
| G4.12 | Change | A change to the record's schema, the pipeline that writes it or a retention period takes its own row in the change table | |
| G4.13 | One clock | One controlled time source for every stamp | |
| G4.14 | Completeness | A scheduled reconciliation of the store against an independent count, the gateway's request log, each gap an alert | |
| G4.15 | Immutability | Storage that refuses edits, and deletion before the retention in force ends, tested by an attempt to edit that fails |
Exit, on passage. The system may serve production traffic within the exposure the record states.
Evidence. The record cites the approved configuration, the rollback test and, for the evidence layer, the record of each control confirmed for this system.
G5 Operation
Purpose (Five-Gate): "To establish that the system in production remains within the conditions under which it was approved, and that the institution will know promptly when it does not." "G5 is the only gate that is not passed once."
| # | Entry criterion | Basis | Exists (yes / no) | Artifact identifier |
|---|---|---|---|---|
| G5.1 | Monitoring live and covering the conditions the approval was conditioned on | Five-Gate; E-23: monitoring "ensures models remain fit-for-purpose and aims to detect performance issues or breaches" | ||
| G5.2 | Detection triggers armed | Five-Gate; E-23: thresholds for breaches, contingency plans and escalation procedures | ||
| G5.3 | The revalidation trigger register populated (the trigger register template, Chapter 10) | Five-Gate | ||
| G5.4 | A named incident lead, appointed and reachable | Five-Gate. "Incident" is the discipline's word; E-23 never uses it |
The decision at every trigger. At G5 the decision is taken on entry and then again at every revalidation trigger. Its exits are retirement, rollback or loop-back to G2. The record is written on entry and dated again at each trigger.
| # | Date and time | Trigger (register row) | Decision: continue, rollback to G4, loop-back to G2, or retirement | Named person | Reason | Policy version in force |
|---|---|---|---|---|---|---|
| G5.5 |
Rejection. A system that cannot arm its triggers or cannot be monitored against its approval conditions returns to G4, rather than being left running while monitoring is arranged.
Part 4. The two ways back, and retirement
Rollback: G5 to G4
The system is withdrawn from production traffic by a decision of the person accountable at G5 or at G3. Rollback does not invalidate the G3 approval: the system remains approved, is not in production, and returns by passing G4 again (Five-Gate, Chapter 5).
| # | Field | Basis | Entry |
|---|---|---|---|
| B1 | System and version | Five-Gate | |
| B2 | Decided by (the person accountable at G5 or at G3), named | Five-Gate | |
| B3 | Stated reason | Five-Gate | |
| B4 | Date and time | Five-Gate | |
| B5 | Where the evidence of the reversal is preserved | Five-Gate | |
| B6 | The rollback test relied on, with the exposure it covered and its date | Five-Gate; scope and date are the discipline's (Chapter 5) |
A withdrawal recorded only in a deployment log is not a rollback.
Loop-back: G5 to G2
An incident that calls the system's fitness into question reopens G2, and does not merely produce a fix. Which incidents trigger loop-back is defined in advance in the trigger register, not decided after the incident by the people whose work would be reopened. E-23's list of events that should prompt a review is where that register starts.
| # | Field | Basis | Entry |
|---|---|---|---|
| L1 | The incident | Five-Gate | |
| L2 | The trigger register row that defined it in advance | Five-Gate | |
| L3 | Date G2 reopened, and the named validator | Five-Gate | |
| L4 | What the validation is asked to do: revisit a judgment on a failure mode it examined, or widen its scope to one it did not | The discipline's (Chapter 5) |
Retirement
A recorded decision by the person accountable at G3 or a successor in the role, with the evidence records of every gate retained for as long as the institution's obligations require (Five-Gate). Decommissioned models stay on the inventory for a period the institution considers reasonable (E-23).
| # | Field | Basis | Entry |
|---|---|---|---|
| T1 | Who decided | Five-Gate (Chapter 5); Chapter 10 | |
| T2 | On what evidence | Chapter 10 | |
| T3 | On what date | Chapter 10 | |
| T4 | What replaces the system, or that nothing does | Chapter 10 | |
| T5 | What is retained, and for how long | Chapter 10; E-23: retaining the retired model and documentation for a set period | |
| T6 | What downstream depends on the retired output | Chapter 10; E-23: monitoring downstream effects | |
| T7 | For a vendor model, what the contract's exit terms actually delivered when exercised | Chapter 10; E-23: additional actions for third-party models | |
| T8 | Relevant stakeholders alerted of the planned decommission | E-23 |
Part 5. The ten-minute test of an approval record
Read the approval record for four things (Chapter 5). A record that fails the first test failed the gate, whatever the answers to the other three; one person is the discipline's narrower choice inside what E-23 permits.
| # | Test | Result |
|---|---|---|
| X1 | Does it name one person or a body? | |
| X2 | Does it cite the validation report by an identifier that would retrieve it? | |
| X3 | Does it state what the system is approved to do and until when, or under what trigger? | |
| X4 | Does it carry the version of the policy that was in force? |
E-23's inventory schema asks for a "model approver" field and a "next review date" field; a failed approval shows first in those two.
The claim, the test, the artifact (Chapter 5)
The claim. G3, approval, is the gate to check first: the one everyone believes was passed, because a minute says so.
The test. Read the approval record for one named person, the discipline's narrower choice inside what E-23 permits, then for the validation by identifier, the use and period, and the policy version.
The artifact. Your system on the five gates, with five gate records begun.
nabeelkhan.com/e-23/gate-record. Questions: nabeelkhan.com/contact.
Your download has started
Next in the book: The BOE Declaration template and the Coldbrook worked example (Chapter 6). Or take the E-23 check to see which template matters most for you.
Where this sits.
The Office of the Superintendent of Financial Institutions (OSFI) does not endorse, approve or recommend this book, its author or any framework in it. Conformance with any framework named here is self-declared, by the institution, on its own record. Coldbrook, Thornbury and Pellbrook are fictional institutions, invented for the book.
For one system, write the name of the one person who approved its deployment. E-23 allows a committee; the handbook asks for a name.
Ask your AI assistant instead.
This page is a snapshot, accurate at the release it cites. The same corpus is callable, publicly and without a key, so an assistant can query it live and return an answer carrying the source it came from. For this page that is get_framework, which returns the Defensible AI Framework Registry entry for any framework these templates are built on (the Five-Gate Deployment Model, the AVRF, PEVG, PARA), with its version and the concept DOI of its deposited specification. It does not yet hold the E-23 handbook or the guideline itself; for those, this page and the book are the source.
claude mcp add --transport http concylium https://mcp.nabeelkhan.com/api/mcp
Claude Desktop, ChatGPT, Cursor, VS Code and Gemini CLI take the endpoint on its own: https://mcp.nabeelkhan.com/api/mcp. No key, no account, nothing to sign. Setup for every client.
“Using Concylium, get the Five-Gate Deployment Model and the AVRF from the framework registry, with their versions and DOIs, and tell me which gate a vendor model decision belongs to.”
A framework quoted from memory drifts. One returned from its registry, with the DOI of the deposited specification, does not.