Calibrate once. Deploy for years.

The setup

One actuator: a valve, an injector, a pump. The factory’s correction table says “to get 40%, command 47%,” measured on a reference unit and frozen at ship. This demo runs one imaginary valve and asks: what happens years later, when the hardware has drifted and the table hasn’t?

Reading the three rails

Each rail is the same valve at the same instant, run by a different strategy. They see identical targets.

• The black tick: what was asked for this instant. It jumps around because real equipment gets stepped through its range, not swept smoothly.
• Frozen calibration: runs the factory table exactly as shipped, forever. This is how nearly everything ships today.
• Trim head: the factory table plus two self-adjusting knobs, an overall gain and an offset, chosen in advance by an engineer. Roughly what “field recalibration” looks like in practice.
• Live map: the Synapse kernel. It learned the entire correction curve at commissioning and keeps re-learning it on the unit, from each command it issues and each delivery it measures.

Dot on the tick means the valve delivered what was asked. Dot away from the tick means it didn’t.

What to watch

1. Factory fresh. The valve matches the reference, so the shipped table is correct and all three dots sit on their ticks. This matters: frozen calibration works on day one, which is exactly why the industry ships it.
2. The unit ages (automatically after a few seconds, or press the button). Over about a minute the valve physically drifts: its dead zone grows and its response sags. The demo compresses years of wear into that minute.
3. The gray dot walks off the ticks and stays there. The table is still executed perfectly; it just describes a valve that no longer exists. Its error settles at several times the others.
4. Amber and purple stay near the ticks. Both adapt. This is adaptation after deployment, not a contest; the trim head is a real example of it.
5. Watch the readout throughout the run. The lesson count climbs every tick because the live map is rewriting itself continuously, and its largest weight stays near 1.0 the whole time. It is learning in the field, and by construction it cannot run away.

Why this costs money today

When a frozen calibration goes stale, the conventional fix is to pull the unit, put it on a bench, re-derive the table, and re-flash it per unit, in the field, forever. Most of the time nobody does that, so the fleet runs increasingly wrong and the error gets written off as tolerance.

The part that doesn’t show up as a number

The trim head does well, but its two knobs are a guess made in advance: someone chose them, set their limits, wrote the adaptation law and validated it, per actuator, per anticipated failure. That loop repeats for every product, and in the field it is a ceiling: the knobs only cover the drift someone predicted. The live map carries no guess. Even matching the hand-designed answer is the win, because it deletes the design loop and covers what the design would have missed.

Doesn’t this hide a failing part?

That’s why production trims are clamped and monitored: unlimited adaptation can hide a failing part until it is a warranty event. Wear is what this page shows, and the map should keep moving with it. Distinguishing that from a dying component is a separately demonstrated mechanism: on NASA’s turbofan benchmark, a consolidation gate on the same kernel keeps tracking benign drift and stops folding the fault in once degradation begins (Proof).

Scope

The valve is simulated; the learning is not, and it rests on one precondition: the delivered quantity has to be measurable. Actuators with position feedback have that signal today; a quantity only ever inferred from a model does not. The drift shown is wear, a change of shape, not a reversal of direction.