A digital twin needs a mismatch register.

Video caption

A digital twin needs a mismatch register.

The model's sensor never goes stale.

Case, assumption, prediction, observation, use decision.

Turn mismatch into better tests, not hidden optimism.

#EricFieldNotes

Full written post / accessibility read

A twin can help test PLC changes before a plant window. But the useful record is not just pass or fail. It is where the model predicted one behavior and an isolated bench or authorized machine observation showed another.

An AI-generated stop routine passes because the simulated jam input is always fresh and the drive stops instantly. The actual acceptance requirement may include stale data and delayed feedback. The green run is honest about the model it had, not about the machine it omitted.

For each fault and recovery path, record the modeled input, physics assumption, predicted state, bench or authorized site observation, and decision about whether the twin is credible for that specific claim. Assign an owner to close or narrow each mismatch.

Update the model or add an independent bench and commissioning check where the twin is weak. Do this because a transparent limitation makes simulation valuable; an unrecorded mismatch makes a green run dangerous to overinterpret.

#EricFieldNotes

Four-beat scene transcript

1. A digital twin needs a mismatch register.

A twin can help test PLC changes before a plant window. But the useful record is not just pass or fail. It is where the model predicted one behavior and an isolated bench or authorized machine observation showed another.

Visual: Without one, every green run looks equally credible.

2. The missing fault creates false scope.

An AI-generated stop routine passes because the simulated jam input is always fresh and the drive stops instantly. The actual acceptance requirement may include stale data and delayed feedback. The green run is honest about the model it had, not about the machine it omitted.

Visual: The model's sensor never goes stale.

3. Make five fields mandatory.

For each fault and recovery path, record the modeled input, physics assumption, predicted state, bench or authorized site observation, and decision about whether the twin is credible for that specific claim. Assign an owner to close or narrow each mismatch.

Visual: Case, assumption, prediction, observation, use decision.

4. A useful twin knows its limits.

Update the model or add an independent bench and commissioning check where the twin is weak. Do this because a transparent limitation makes simulation valuable; an unrecorded mismatch makes a green run dangerous to overinterpret.

Visual: Turn mismatch into better tests, not hidden optimism.

Research and claim limits

• Siemens Automation Framework documentation (S140)
• NIST: Credibility Consideration for Digital Twins in Manufacturing (S141)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.