blog

AI Models

Good on Its Best Day Is Not a Spec

Good on Its Best Day Is Not a Spec

Yehonatan Rubin

|

|

Reading Time:

5

min

Table of Contents

No airline picks an engine because it was brilliant on one test flight. They pick the one certified to behave the same way every time.

Airlines do not swap an engine's control software overnight because a newer version scored better in a lab. Every change goes through certification. The aircraft flies as configured, and every flight is logged against that exact configuration.

Nobody calls this slow. It is why you board without checking the release notes.

I think AI in serious environments needs the same discipline. Most of it does not have it yet.

The model underneath is moving

Build on a hosted frontier model and the weights are not yours. An alias like "latest" can point at new weights next month. The provider picks the date.

Here is what shifts when that happens:

  • Tool-calling and JSON adherence change. A parser that worked yesterday breaks today.

  • Verbosity and refusal behavior change. A flow that used to trigger does not.

  • Safety tuning changes. Prompts tuned for one version drift on the next.

  • Deprecations arrive on the provider's schedule, not your change window.

Even at temperature zero, hosted inference is not guaranteed to return the same output twice. Batching, hardware and kernel choices all leak in.

For a chatbot, that is noise. For a system that makes decisions, it is a defect you cannot reproduce.

Teams live in the tail, not the average

A leaderboard reports mean performance. A production system lives in the tail.

The question is not "how good is the model on its best day?" It is "what is the worst thing it does on a bad day, and how often?" One wrong decision in a thousand matters more than a point of benchmark gain.

I would take a narrow, measurable error distribution I can bound and test against. I would not take the highest peak.

Think about the supermarket. Tomatoes that are amazing some weeks and awful others do not earn loyalty. People go back to the store they can trust. Trust comes from the floor, not the ceiling.

Regulated industries do not get surprises

Banks, hospitals and critical infrastructure run on a different clock. Their data does not leave the building. Their auditors ask what ran, when, and why it decided what it decided.

"The provider updated the model" is not an answer an auditor accepts.

So the right question changes. Not "which model is the smartest this month?" Instead: "Which model will behave the same way next quarter as it does today?"

Sovereignty means weights you can name

AI sovereignty sounds like a policy word. Technically, it is simple. You can point to the artifact that made the decision and prove it has not changed.

The open-weight world has made this practical. Llama gave the industry a strong base. Nvidia post-trained on top of it and released Nemotron. OpenAI now ships open-weight GPT models. Any of them can run on GPUs inside an organization's own environment, so the data stays where it lives.

Real control means pinning the whole stack, not just the weights:

  • Model weights, by hash

  • Adapter version, if post-training ships as an adapter

  • Inference container, by digest

  • Precision, quantization and decoding settings

Deploy that stack inside the client's environment, and the answer to an auditor becomes a version number and a hash.

Post-training is where general becomes useful

A base model knows the world. It does not know your policy, your precedent or your way of deciding.

Post-training closes that gap. It shapes a general model around how one organization actually works. The signal comes from the organization itself: past cases, how people resolved them, what was approved and what was rejected.

Then you freeze the result. Fixed weights plus a fixed policy give you behavior you can test, replay and defend.

This is also why a smaller, pinned model can hold its own on a narrow job. When the model gets the right context attached to each decision, it reasons over evidence. It does not guess from memory.

Upgrades should be candidates, not surprises

Pinning is not the same as freezing forever. It means change is deliberate.

When a stronger model arrives, treat it like a release candidate:

  1. Replay it against historical cases.

  2. Run it in shadow next to the live model on real traffic.

  3. Compare its output side by side with what people actually decided.

  4. Let the owner of the risk decide whether it graduates.

Nothing acts until the owner says so. Every step stays reversible. The model you tested in week one is the model running in week four.

Every decision should carry a receipt

If a system decides something, it should be able to show its work. Which model. Which version. Which policy. What context it saw. Who approved it.

Replay it next year and you should get the same answer for the same reasons. That is what defensible looks like when AI is in the loop.

The tradeoff, said plainly

A pinned open model can lose to the newest frontier model on a broad reasoning benchmark. I accept that.

In regulated environments, the buyer's preference is clear. They would rather have a system that is a bit less brilliant on its best day and fully predictable on every other day. They can plan around known behavior. They cannot plan around surprises.

A spec is a promise about every day, not the best one. That is certification. Fewer thrills. More flights that land.

Questions? Knock twice.