A record you can learn from
Context, plan, activity, response, decision and outcome sit in one dated record.
SAVRN solutions
How data becomes
The SAVRN data principle
How data becomes
intelligence.
Record the decision. Record the outcome. Then train.
Most organizations hold measurements. Few hold a record of what was decided and what happened next. SAVRN builds the governed record, the computing and the review steps that turn a dataset into AI people can rely on.
Decisions stay with people
A model supports qualified staff. A person approves the action, and the record keeps the reason.
Claims wait for proof
A model runs in shadow and passes a test on outside data before anyone relies on it.
The SAVRN data principle
A dataset is a record of
decisions and outcomes.
A sensor reading says what was measured. It does not say what anyone did about it, or whether that helped. A model trained on readings alone learns to describe. A model trained on decisions and outcomes can learn what works.
Data becomes intelligence when the record connects a decision to its outcome.
Context
Who or what is observed, under which permission, and what makes this case different from the next.
Plan
What was intended before the work began, kept as written even if it changes later.
Activity
What happened, with its source, time and units.
Response
How the person or system responded, including what they reported themselves.
Decision
What a qualified person decided, who approved it and why.
Outcome
What happened next over a stated period, including the cases lost to follow-up.
Most datasets stop at part three. Parts five and six are what a model needs to learn which actions work.
Five rules
The rules that make
a record trustworthy.
The rules apply before any model is trained. They decide what may be collected, how it is compared, who acts on it and what may be claimed.
Permission comes first.
Each purpose has its own permission: service, research, model training, publication and marketing. A withdrawn permission blocks the next job that needs it.
Compare each person with their own history.
Use the individual baseline first. Use a matched group only where the data are sufficient, and say so where they are not.
Show what is missing.
A skipped survey is not a good day, and a missing report is not a clean record. Gaps stay visible and are never filled with a score.
A qualified person decides.
The system lays out the facts and routes them to the right reviewer. It does not diagnose, select or exclude.
Prove it before you claim it.
A working pilot shows that the system runs. An effect on outcomes needs follow-up, a comparison group and a test on outside data.
Why the principle exists
Measurements are common.
Tested evidence is rare.
Our first design case is in sport, because the published research there shows the gap clearly. Every figure below links to the study that reported it.
sports injury prediction models had been tested on outside athletes
Bullock et al, Sports Medicine, 2022was the median sample size reported for one leading sports science journal
Mesquida et al, Royal Society Open Science, 2022published sports science studies held up when they were repeated
Murphy et al, Sports Medicine, 2025injured athletes are needed to detect a small to moderate risk factor
Bahr and Holme, British Journal of Sports Medicine, 2003The SAVRN method
Nine steps from a dataset
to a model you can rely on.
The order matters. A model is tested on outside data and run in shadow before it can affect a decision, and its effect on outcomes is tested apart from its accuracy.
- Step
Specify the task.
Separate reporting, forecasting, recommendation support and risk research. Define the prediction time, the population, the outcome and the action before choosing a model.
- Step
Approve the corpus.
Filter records by purpose and rights. Remove duplicates, keep provenance and exclude labels that no qualified person confirmed.
- Step
Build point-in-time examples.
Use only what was known before the decision. Later diagnoses, return dates and follow-up notes stay out of the inputs.
- Step
Separate the test population.
Hold out whole people and later time periods, and a separate site where one exists. One person's neighboring records never sit on both sides of the test.
- Step
Benchmark simple methods.
Compare against the current workflow, approved rules and a plain statistical baseline. A complex model has to beat all three.
- Step
Measure usefulness and harm.
Report calibration, precision and recall, uncertainty, results by subgroup, the effect of missing data and the workload that false alerts create.
- Step
Run in shadow.
Generate outputs without changing any decision. Check that the system is reliable and that staff find the output worth acting on.
- Step
Test the intervention.
Whether acting on the output improves outcomes is a separate question. It needs its own prospective comparison.
- Step
Control the release.
Lock versions, require scientific and professional approval, watch for drift and keep a way back. Nothing learns from unreviewed production conversations.
The full method.
See the five kinds of tools, the fields every record carries and the four mistakes the method is built to prevent.
Design cases
The principle,
applied to a field.
A design case applies the principle to one field. It sets out the record, the rules and the measures. It reports no results until results have been measured.
Design case
Athlete development
A coach-led record for youth and team sport. It connects training, recovery, video and outcomes, and keeps health data apart from recruiting and marketing.
Read the athlete development caseWhere SAVRN fits
The record, the compute
and the review.
SAVRN provides the governed environment behind the record. The organization owns its decisions, and qualified professionals keep their authority.
Permissions, approvals and audit
Governed records
Keep permissions, reviews and decisions attached to the work. The SAVRN platform connects projects, data boundaries, approved tools and deliverables.
Capacity placed by data class
Compute near the data
Place each workload where its data and operating requirements allow. Video, model training and serving run on capacity sized for the work.
SAVRN's own record
A model trained on your data
SAVRN trained an open model on its own documents and published what happened, including the mistakes. The same steps apply to a customer's record.
SAVRN Model Hub
Open datasets with their licenses
See which open datasets models are trained and tested on, with the license, size and structure of each one.
Research and education
University research
Connect investigators, permitted data, approved models and reviewed results in one university plan.
How SAVRN checks its work
Methods and corrections
Read how SAVRN sources its research, labels what is measured and what is designed, and corrects errors.
A practical first conversation
Start with the decision you want to improve.
Bring one decision your team makes often, the data you hold today and the outcome you would measure. We will map the record that connects them.
What is the SAVRN data principle?
Data becomes intelligence when the record connects a decision to its outcome. A useful dataset holds six things for each case: context, plan, activity, response, decision and outcome.
How is this different from collecting more data?
More readings describe what happened. They do not show what anyone did about it or whether it helped. The principle adds the decision and the outcome, because a model needs both to learn which actions work.
Can the system make decisions on its own?
No. It prepares the facts, routes them to the right reviewer and records the decision. A qualified person approves any action.
When can a model's output be used in practice?
After it has run in shadow without changing decisions, passed a test on outside data and been approved by the people responsible. Until then it is research.
Where does the data go when SAVRN runs a model?
It stays inside the customer's boundary. SAVRN runs its own model on its own hardware. The model is built on Ai2's OLMo 3, a fully open base model, and post-trained by SAVRN on SAVRN data. Nothing a customer works on is sent to an outside AI provider.
Which fields does the principle apply to?
Any field where people make repeated decisions and can observe what follows. The first design case is athlete development.
Sources for this page
- Just How Confident Can We Be in Predicting Sports Injuries?. Bullock et al, Sports Medicine, 2022. Systematic review of 30 studies and 204 injury prediction models.
- Replication concerns in sports and exercise science. Mesquida et al, Royal Society Open Science, 2022. Narrative review of sample size, power and replication in the field.
- Estimating the Replicability of Sports and Exercise Science Research. Murphy et al, Sports Medicine, 2025. Replication of 25 studies published in leading journals from 2016 to 2021.
- Risk factors for sports injuries: a methodological approach. Bahr and Holme, British Journal of Sports Medicine, 2003. Methods paper on how many injury cases a risk factor study needs.