VII · Metron ariston · Measure is best
Built to improve as models improve.
The best engineering system improves as models improve.
New models adopted only after technical, security, licensing, and economic checks.
In practice
What it means for your system.
Better models arrive often. A well-built system can adopt one without a rebuild, once it passes the same checks as everything else.
- Model choice sits behind a clear boundary, so it can change without a rebuild.
- Candidates pass technical, security, licensing, and economic checks first.
- Rollback is preserved, and a person approves every switch.
The word for it
Epidosis
ἐπίδοσις
Epidosis is Greek for “growth, steady progress over time.” The system is built so that each better model adds to it.
Why it matters
A better model is an option, not an obligation.
New models and methods arrive often, and some are a real gain for your work. Each one also brings new behavior, new terms, and a new cost per run.
A switch made on reputation can change outputs your team relies on, or move your data under terms nobody reviewed. The cost of that surfaces later, in rework and risk.
So a new model is treated as a candidate, checked on your own work before it is adopted. Because rollback is kept, each switch stays a reversible business decision.
Read it accurately
Scope of the rule.
- The system is complete at acceptance.
- The system is complete and accepted on its own terms. Improvement is a later option, taken when a candidate measures better on your work.
- New techniques pass the same checks as models.
- New techniques, such as retrieval or prompting methods, pass the same checks. Any change that alters behavior is treated as a candidate.
- Cost, security, and licensing count alongside quality.
- Technical, security, licensing, and economic checks each count. A candidate that scores higher but costs more per run may still be declined.
How it is measured
Every rule is something you can check.
New techniques adopted only after technical, security, licensing, and economic checks, with rollback preserved.
In an engagement
Where the rule is applied.
The rule is checked at each stage of the work, from the first design to the system in operation.
In the design
Model choice is placed behind a clear boundary. A held-out set of your own work is reserved for testing candidates.
In the build
The system records which model and version produced each run. The previous configuration is kept ready, so rollback is a tested step.
At acceptance
The accepted system becomes the baseline. Any later candidate is measured against that baseline on the same held-out cases.
In operation
AI Stewardship assesses relevant candidates against the agreed checks. Your owner approves each switch, and the rollback route stays in place.
Check your own system
Five questions to ask this week.
Each one has a yes or no answer. A no marks where to start.
- Could you change the model behind one task without rebuilding the system?
- Is there a held-out set of your own work to test a new model against?
- Does someone review a new model’s license and data terms before it is used?
- Do you compare running cost, not only quality, before a switch?
- Could you return to the previous model today if a switch went wrong?
Delivered through
The services that apply this rule.
Metron ariston
Measure is best.
Nothing ships on faith.
Apply the rule
Check it against your own system.
Questions
Do we need to adopt every new model that is released?
No. Each candidate is logged and assessed against your work. A switch happens only when it measures better and your owner approves it.
Who watches for new models and methods?
Under AI Stewardship, we track what is relevant to your workloads. Research-to-Production adds a dedicated pipeline for reproducing and testing new techniques.
What if a model’s license or terms change after adoption?
License changes are reviewed like a new candidate. If the terms no longer fit, the rollback route lets your owner approve a return to an earlier model.
How do we check this rule on our own system?
Look at the last model or technique change. It should have passed technical, security, licensing, and economic checks, with rollback preserved.
Does checking a new model add cost?
Each assessment takes engineering time, so candidates are screened for relevance to your workloads first. A full test runs only when a candidate looks likely to pay.
How is rollback kept ready?
The previous model and its configuration stay available after a switch. The run records make a return measurable, and your owner approves it.