The whole model life cycle, from a spreadsheet to a small language model
Model risk management is the discipline of making sure the models an organisation relies on are built correctly, validated independently, used within their limits and monitored over time - with an inventory of models, documented assumptions, approvals and an audit trail.
VionRisk starts with the risk models teams already run in Excel. The same discipline applies to every model a business depends on, including the AI models it now builds. This page separates what exists in VionRisk today from what is on the roadmap.
Build, validate, host, monitor, govern
Today in VionRisk 0.7.0. Roadmap planned, without a date; the design for small language models is still being agreed.
Build
Turn a spreadsheet into a model of uncertainty.
Today: 239 worksheet functions, distributions fitted to data, time series, correlation.
Roadmap: fine-tune small language models on your data.
Validate
Show the model computes what it claims.
Today: 1,700+ automated tests on every release; a conformance workbook compared with licensed @RISK; exact maths with documented assumptions.
Roadmap: versioned evaluation suites for your own models and SLMs.
Host
Run the model where it is needed.
Today: on your computers, in Excel; AI agents over a local MCP server.
Roadmap: hosted models on our servers, your cloud account or the edge, without Excel.
Monitor
Check the model against what really happens.
Today: the assumptions check tests observed data against a model's inputs and outputs: OK, Watch or Breach.
Roadmap: drift alerts on live use, for models and SLMs.
Govern
Know who changed what, and why it was approved.
Today: the software and environment stamped on every result; fixed seeds; the workbook never modified; what-if history kept.
Roadmap: a model inventory, approvals and an audit trail.
Validation you can show a reviewer, today
- Exact maths, written down. When two engines could differ, VionRisk follows the exact mathematical definition. Every convention and assumption is documented with its status: exact, documented by @RISK, observed in @RISK, or assumed.
- Conformance against @RISK. A workbook with every function is run in a licensed @RISK and in VionRisk; in the latest run all 121 functions compared agree - 110 exactly, the rest where VionRisk keeps the exact value and @RISK approximates, each documented. The comparison.
- Reproducible results. A fixed seed gives identical results, and every result records VionRisk, Excel, Python and package versions and the operating system.
- Assumptions check. Give last quarter's actuals and VionRisk tests each input against its distribution, checks values the model cannot produce, compares correlations, and counts outcomes outside the simulated P5-P95.
Small language models, on the roadmap
A small language model (SLM) is a language model with far fewer parameters than the largest general-purpose models, trained or fine-tuned for a narrower task: cheaper to run, able to stay on your own servers, and in need of the same validation as any other model.
The plan is to treat an SLM the way VionRisk treats a financial model: train or fine-tune it on your data, host it where you choose, run a versioned evaluation suite on every release, measure its uncertainty by simulation, watch for drift in live use, and keep approvals and an audit trail - usable by AI agents over MCP. None of this is available yet; if it matters to you, tell us what you need and help shape it.
Model risk management: questions
What is model risk management?
Model risk management is the discipline of making sure the models an organisation relies on are built correctly, validated independently, used within their limits and monitored over time, with an inventory of models, documented assumptions, approvals and an audit trail. Its aim is to avoid losses and bad decisions caused by models that are wrong or misused.
What is a small language model?
A small language model (SLM) is a language model with far fewer parameters than the largest general-purpose models, typically trained or fine-tuned for a narrower task. It is cheaper to run, can be hosted on an organisation's own servers or devices, and keeps data in house, but it needs the same validation and monitoring as any other model.
What does VionRisk do for model validation today?
Today VionRisk is tested on every release with an automated suite and a conformance workbook compared against a licensed copy of @RISK; it follows exact mathematical definitions with every convention and assumption documented; it stamps the software and environment on every result; and its assumptions check tests observed data against a model's assumptions, flagging each as OK, Watch or Breach.
Will VionRisk host and govern small language models?
It is on the roadmap, not available today: training or fine-tuning small language models on a customer's data, hosting them on our servers, the customer's cloud or at the edge, and validating and governing them like a financial model, with versioned evaluation suites, drift monitoring, approvals and an audit trail, usable by AI agents over MCP. The design is still being agreed.
Start with the models you run today
A pilot puts VionRisk's validation and AI agents on your existing Excel risk models, and gives you a say in what comes next.