Model Control Framework, Universal Template

Insights

A model is a software artefact. That is the premise behind this template: a spreadsheet or calculation file that feeds a decision deserves the same disciplines as a CRM system, even though it sits outside the application portfolio and has no vendor. The reasoning is set out in the article Model Risk and EUC, from Blind Spot to Control. This is the instrument that goes with it.

The template does three things that make a control framework usable. First, it determines the risk class, so that the weight of scrutiny scales with what goes wrong if the outcome is incorrect. It then sets out, per discipline, the test questions that belong to that class, along with the evidence you should expect. And at the end it produces a coverage picture and a list of open points you can carry into a report or into follow-up.

Use it per artefact. Your assessment stays in your own browser; nothing is sent to a server. Export to Word if you want to keep or share the file.

Enter the artefact you are assessing

Nothing assessed yet
✓ Copied to clipboard
Model control framework
NAME OF THE MODEL
ORGANISATION · version VERSION · owner OWNER
Assessed by ASSESSOR on DATE
Use of the outcome: WHAT THE OUTCOME IS USED FOR · administration: ADMINISTRATOR

Step 1. Determine the risk class

The weight of the framework should scale with what goes wrong if the outcome is incorrect. Six questions determine the class; that class then determines which of the 35 test questions apply.

What happens if the outcome is wrong?
Where does the outcome go?
How many people use or reuse it?
How complex is the calculation logic?
How much manual work does a run involve?
How quickly does an error surface?
Class C

Step 2. The control framework

Seven disciplines, 35 test questions. Each question states the evidence you should expect and from which risk class the question applies. A question that does not apply to your class stays visible but does not count; mark it as met anyway if you want to be stricter than the class requires.

Discipline 1 of 7

Change management

NormEvery change to the calculation logic is requested, reviewed, tested, and approved by someone other than the builder, and can be traced to a version.

Without change management, nobody knows which version held the truth at the moment the decision was made. That is not primarily a control problem but a reconstruction problem: last quarter's outcome can no longer be recalculated.

ITGC and COBIT ISO/IEC/IEEE 12207:2017 ICAEW PRA SS1/23
1.1Is it designated which version is current, and is that visible on the file itself?From class C
Expected evidence: Version number in the file plus the entry in the model inventory.
Assessment: not assessed
1.2Are changes to formulas or structure recorded in advance: who, what, why?From class B
Expected evidence: A change log in the file or tickets in the management system.
Assessment: not assessed
1.3Is a change tested before it goes into use, with a recorded outcome?From class B
Expected evidence: A test report, or a comparison of the old and new outcome on the same input.
Assessment: not assessed
1.4Does someone other than the builder approve the change before it goes into use?From class B
Expected evidence: Approval recorded in the ticket, in an email, or on the change form.
Assessment: not assessed
1.5Are earlier versions retained, so that a past outcome is reproducible?Class A only
Expected evidence: A version archive with dates, or a document library with version history.
Assessment: not assessed
Discipline 2 of 7

Documentation

NormA peer who did not build the model can understand it, use it, check it, and take it over.

The test is not whether documentation exists but whether someone else can take over the model. Models that depend on one person collapse the moment that person leaves, and on average that happens sooner than the model gets replaced.

ICAEW FAST Standard 02c ISO/IEC/IEEE 12207:2017 ISO/IEC 25010:2023
2.1Is it documented which question the model answers and which decision the outcome serves?From class C
Expected evidence: An explanatory sheet or document header stating purpose and use.
Assessment: not assessed
2.2Are the assumptions and limitations explicit, including what the model is not intended for?From class B
Expected evidence: A list of assumptions with source and date, plus an explicit list of exclusions.
Assessment: not assessed
2.3Is the calculation logic described so that a peer can follow it?From class B
Expected evidence: A description of the structure per sheet or block, with the key formulas explained in words.
Assessment: not assessed
2.4Is it established who the owner is, who administers it, and who may make changes?From class C
Expected evidence: Roles named in the file and in the model inventory.
Assessment: not assessed
2.5Was the documentation updated with the latest change?From class B
Expected evidence: The documentation date is the same as, or later than, the date of the current version.
Assessment: not assessed
Discipline 3 of 7

Segregation of duties and access

NormBuilding, approving, using, and checking do not sit with one person, and access is limited to those who need it.

In a spreadsheet, segregation of duties almost always disappears: the same person builds, enters data, calculates, and presents. That is acceptable as long as it is a deliberate decision with a visible compensating measure, and unacceptable as long as nobody has noticed it.

ITGC and COBIT ICAEW PRA SS1/23 SR 11-7 and SR 26-2
3.1Is the builder someone other than the person who approves the outcome or uses it for the decision?From class B
Expected evidence: A roles overview, or a documented compensating measure if the separation is not feasible.
Assessment: not assessed
3.2Is write access limited to a named group and read access to those who need it?From class C
Expected evidence: A permissions overview of the folder or library.
Assessment: not assessed
3.3Are the parts that must not be changed technically protected?From class B
Expected evidence: Sheet protection or cell locking enabled, with the key held by the administrator.
Assessment: not assessed
3.4Is access adjusted on departure or role change?Class A only
Expected evidence: Alignment with the joiner and leaver process, with a periodic review of the rights.
Assessment: not assessed
3.5Is there a designated backup, so the model does not depend on one person?From class B
Expected evidence: The backup's name in the inventory, with evidence that they can run the model.
Assessment: not assessed
Discipline 4 of 7

Reference data and parameters

NormFixed data such as rates, percentages, and tables are kept separate from the logic, can be traced to a source, and are updated deliberately.

A rate typed directly into a formula is untraceable the moment it changes. The silent failure in this discipline is not a wrong calculation but an outdated figure that stays in place for years.

ICAEW FAST Standard 02c ITGC and COBIT EU AI Act
4.1Are parameters and tables kept separate from the calculation logic, so not typed into formulas?From class C
Expected evidence: A dedicated sheet or block of parameters that the formulas reference.
Assessment: not assessed
4.2Is the source and reference date recorded for every parameter?From class B
Expected evidence: A source column and a reference-date column next to every parameter.
Assessment: not assessed
4.3Is it documented who may change a parameter and when it is reviewed?From class B
Expected evidence: A management agreement naming the responsible person and the review moment.
Assessment: not assessed
4.4Is a parameter change treated the same as a change to the logic?Class A only
Expected evidence: The parameter change appears in the same change log.
Assessment: not assessed
4.5Can you determine which parameter set belonged to a published outcome?Class A only
Expected evidence: A recorded parameter set stored with the saved run.
Assessment: not assessed
Discipline 5 of 7

Input

NormThe data entering the model is complete and correct, and demonstrably comes from an identifiable source.

Almost every reconstruction of a model error ends at the input, not at the formula. A model that silently processes incorrect input is more dangerous than one that makes calculation errors, because the outcome still looks plausible.

ICAEW ITGC and COBIT EU AI Act ECB Guide to internal models
5.1Is the source and delivery method named for every input stream?From class C
Expected evidence: An overview of input streams with source system, frequency, and responsible person.
Assessment: not assessed
5.2Are there checks on completeness and reconciliation with the source system?From class B
Expected evidence: A reconciliation check on counts or totals, visible in the model.
Assessment: not assessed
5.3Are manual input fields visibly separated from calculated fields?From class C
Expected evidence: A fixed formatting convention for input, with a legend.
Assessment: not assessed
5.4Is deviating or missing input flagged instead of processed silently?From class B
Expected evidence: A built-in check that visibly triggers, with a documented follow-up.
Assessment: not assessed
5.5Is the input used for a published outcome retained?Class A only
Expected evidence: A saved input set with the run, or a reference to the frozen source data.
Assessment: not assessed
Discipline 6 of 7

Output

NormThe outcome is traceable, dated, and carries the context the user needs to read it correctly.

The outcome travels further than the model. The moment the figure lands in a presentation, the assumptions have fallen away. What you secure here determines whether the recipient can weigh the figure or can only take it on faith.

ICAEW ISO/IEC 25010:2023 PRA SS1/23 Solvency II
6.1Does every output carry a date, a version, and the model's name?From class C
Expected evidence: A footer or header on the output stating model, version, and date.
Assessment: not assessed
6.2Can the outcome be traced back to the input and parameters used?From class B
Expected evidence: A saved run or audit trail linking input, parameters, and outcome.
Assessment: not assessed
6.3Does the user receive the limitations and the uncertainty, or only the figure?From class B
Expected evidence: A standard explanatory note with the output, stating a range or sensitivity.
Assessment: not assessed
6.4Is it documented who releases the outcome before it goes out into the organisation?From class B
Expected evidence: A release recorded with name and date, separate from the builder.
Assessment: not assessed
6.5Is the outcome reconciled against an independent source or expectation?Class A only
Expected evidence: Reconciliation with the ledger, the prior period, or a second calculation.
Assessment: not assessed
Discipline 7 of 7

Validation

NormSomeone who did not build it periodically confirms that the model does what it should and still fits its purpose.

This is the discipline that is almost always missing outside the financial sector, and it is the only one that can catch failures in the other six. Validation is not checking whether the sum is right but establishing whether the model still gives the right answer to the right question.

PRA SS1/23 SR 11-7 and SR 26-2 ECB Guide to internal models Solvency II ISO/IEC 42001:2023 NIST AI Risk Management Framework 1.0
7.1Was the model independently tested for arithmetic correctness before it went into use?From class B
Expected evidence: A report of an independent recalculation, with sample and outcome.
Assessment: not assessed
7.2Is the testing frequency documented and linked to the risk class?From class B
Expected evidence: A management agreement stating the interval per class, recorded in the inventory.
Assessment: not assessed
7.3Is it tested whether the model still fits its purpose and circumstances?Class A only
Expected evidence: A reassessment of assumptions and scope of application, not just the formulas.
Assessment: not assessed
7.4Have findings from earlier tests been followed up and closed?Class A only
Expected evidence: A findings list with status, owner, and closure date.
Assessment: not assessed
7.5Is the validation documented in a way a third party can follow?Class A only
Expected evidence: A file with design, work performed, findings, and conclusion.
Assessment: not assessed

Step 3. The result

Coverage counts only the questions that apply to your risk class. Questions marked N/A do not count in the denominator; unrated questions do, because a question that goes unanswered does not produce coverage.

DisciplineApplicableMetPartialNot metOpenCoverage
1. Change management00000
2. Documentation00000
3. Segregation of duties and access00000
4. Reference data and parameters00000
5. Input00000
6. Output00000
7. Validation00000
Total00000

Open points

Nothing assessed yet.

Appendix A. Four standards tracks and the coverage gap

Where do the standards this framework rests on come from, and why is there still a gap? Click a chip in the framework above for the explanation per standard.

TrackFrameworksCoversDoes not cover
Regulatory oversightECB Guide to internal models, Solvency II Articles 120 to 125, PRA SS1/23, SR 11-7 and SR 26-2All seven disciplines, including independent validation and a mandatory model inventory.Applies only to banks and insurers licensed for an internal model. Outside that, no equivalent exists.
Spreadsheet and EUCICAEW Twenty Principles (revised 2024), FAST Standard 02c (2019), EuSpRIGBuild quality down to cell level: separating input, processing and output, an explanatory sheet, no hardcoding, version control and backup, testing, built-in checks, protected sections.The organisational half: who approves, who validates, who owns it.
Generic software and ITISO/IEC/IEEE 12207:2017, ISO/IEC 25010:2023, COBIT, ITGCExactly that organisational half: lifecycle, configuration management, change management, logical access, IT operations.In practice, it is not applied to a spreadsheet. The file is not in the application portfolio, so it falls outside the IT auditor's scope.
Artificial intelligenceISO/IEC 42001:2023, NIST AI RMF 1.0, the EU AI Act (including Article 10)The same seven disciplines, but only for systems that learn or infer.Deterministic calculation models fall outside it by definition, no matter how much weight the outcome carries.

The conclusion is not a regulatory vacuum but a coverage problem. The strictest requirements sit with the sector that already saw the risk, the sharpest build standards lack an owner, the broadest frameworks look past the file, and the newest frameworks exclude precisely the most common model type. Outside the financial sector, you control a model not because you must, but because it is sensible. This template is the bridge: the seven disciplines of a software artefact, applied to a model, with a reference per discipline to the framework that already governs it.

Appendix B. The model inventory

This framework assesses one artefact. The question that comes before that is which artefacts you actually have. Without an inventory, you assess the models you happen to know about, and those are rarely the riskiest. Ten columns are enough to get started.

ColumnWhat you record
Model name and versionThe current version, matching what is stated in the file.
Purpose and decisionWhich question the model answers and which decision the outcome serves.
OwnerWho is responsible for the outcome. A role, not a department.
AdministratorWho may make changes and who is the designated backup.
Risk classA, B, or C, with the date of the latest classification.
Input sourcesWhich systems or reports feed the model.
Recipients of the outputWho receives the figure, and whether it leaves the organisation.
Last validationDate, performed by, and the status of the findings.
Next reviewDerived from the risk class.
LocationWhere the file is stored, and where the version archive is stored.

Start from the outcomes, not the files: go through the past year's reports and decisions and ask, for every figure, where it came from. What you find then is the real portfolio.

One framework per model is manageable. A hundred models is not.

This template assesses one artefact and gives you a file per model. If you want the inventory, the classification, the review schedule, and the follow-up of findings in one place, Modelio in the Audirium suite does that. Modelio is in the beta programme.

Back to Insights