AN—002 ARCHITECTURE NOTE · PUBLISHED 2026-06-11

Residual publication in the digital twin fabric

Reconciliation method, how tolerance is defined and by whom, and why retuning a model is treated as an engineering change requiring a signature rather than a maintenance task that happens quietly.

READING TIME 3 MIN · IMPLEMENTATION DETAIL · SUPERSEDES NOTHING

1 — THE RESIDUAL IS THE PRODUCT

A digital twin that agrees with its asset tells you nothing. The engineering value is in the disagreement — the residual between what the model predicted and what the plant measured — because that is where a developing fault first becomes visible.

The fabric therefore publishes residuals as first-class values, not as an internal diagnostic. Every twin emits its residual on every reconciliation cycle, whether the residual is interesting or not, so that the absence of drift is as evidenced as its presence.

2 — RECONCILIATION METHOD

Reconciliation runs on a fixed cycle per asset class — minutes for rotating equipment, hours for thermal systems with long time constants. Each cycle compares a predicted state vector to a measured one and attributes the difference before reporting it.

STEP 1Measurement validation — reject readings outside instrument range or stale beyond their cycle

STEP 2Boundary reconciliation — close mass and energy balance across the asset boundary

STEP 3Attribution — separate instrument error, boundary-condition error and genuine model error

STEP 4Publication — residual, its attribution and its uncertainty, written to the graph

Step 3 is the one that matters and the one most systems omit. An unattributed residual sends an engineer to inspect a healthy machine when the real fault is a drifting transmitter.

3 — WHO DEFINES TOLERANCE

Tolerance is an engineering judgement, so it is set by the responsible engineer for the asset and recorded with a basis — not inferred from historical variance, which merely encodes whatever the asset was doing while the data was collected.

A tolerance may be tightened at any time. Widening one is an engineering change and follows §4. Where no tolerance has been set, the twin publishes its residual with the state UNBOUNDED and no exception is raised — silence is never mistaken for conformance.

4 — WHY RETUNING IS AN ENGINEERING CHANGE

When a model disagrees with an asset, there are two available actions: investigate the asset, or adjust the model until it agrees. The second is faster, and it destroys the instrument. A twin that is retuned whenever it disagrees can never again detect degradation, because degradation is exactly what disagreement looks like.

Retuning is therefore not a maintenance action. It requires a written basis, a named engineer, and a record of the residual history that preceded it — the same treatment as changing a design parameter, because that is what it is.

// model revision record · required fields

basis what changed physically, or what was wrong in the model

residual_before the drift history that motivated the change

authorised_by named responsible engineer, signed

prior_revision retained and replayable, never overwritten

Prior revisions are retained so a past decision can be re-examined against the model that actually produced it. A decision record that cites a model revision which no longer exists is not auditable.

5 — WHAT THE ENGINEER SEES

A residual crossing tolerance raises a reconciliation exception addressed to the responsible engineer. The exception states the magnitude, the attribution, the trend over the preceding window, and the two available dispositions — investigate, or open a model revision. Nothing is auto-tuned, and nothing times out.

AN—001 — the constraint gate EDTF — the twin fabric EP-WP-001 §6 — reference architecture

How to cite this note

ENGINPILOT (2026). Residual publication in the digital twin fabric. Architecture Note AN—002. Bengaluru: ENGINPILOT Intelligence Systems.

All architecture notes