A digital twin can organize geometry, measurements and knowledge about an asset. A system that also promises to predict future conditions or compare interventions makes additional commitments. It needs a model of change, a way to update that model from evidence, and an account of uncertainty.
Calling the system a world model is useful only if those commitments become clearer rather than being hidden behind a broader name.
Research context
Towards Heritage World Models, by George Pavlidis, Vasileios Sevetlidis and Vasileios Arampatzakis, was published in Heritage 9(6), 233, on 13 June 2026, as recorded in the publisher’s issue listing. It develops a conceptual direction beyond a static digital representation toward prediction, memory, uncertainty and intervention evaluation. This tutorial works through a separate one-dimensional state-estimation example. It is not an implementation of a validated heritage world model, and its numerical values have no calibrated conservation interpretation.
Separate the asset record from the dynamic hypothesis
An asset record may say where an instrument is located or which material an element contains. A dynamic model adds a hypothesis about how an unobserved state changes. That state might describe an environmental exposure, a material condition or another quantity relevant to a question. It must be defined before its predictions can be assessed.
For a deliberately simple teaching abstraction, let x be a dimensionless latent exposure index. It is not a probability of damage and it is not constrained to lie between zero and one. Write:
The input u represents a hypothetical management influence. The terms w and v describe process and measurement uncertainty. The formula is an assumption to test, not a physical conservation law.
Prediction and observation are different pieces of evidence
Suppose our current state estimate has mean 1.0 and variance 0.04. Let a = 1.05 and process-noise variance Q = 0.01. With no action input, the next predicted mean is 1.05 and its variance is 1.05² × 0.04 + 0.01 = 0.0541.
Now a fictional instrument reports 1.2 with measurement-noise variance R = 0.09. A scalar Gaussian update combines the prediction and observation through:
The resulting mean lies between the prediction and the observation. The gain is about 0.3754, so the posterior mean is about 1.1063. The update does not simply overwrite the model with the latest sensor value.
Uncertainty belongs in the output
Under these Gaussian assumptions, a rough 95% interval is the mean plus or minus 1.96 standard deviations. The assumptions matter: if the model is systematically wrong, that interval can be narrow and still misleading.
I would therefore examine forecast errors on data not used to fit the model, check whether intervals have appropriate empirical coverage, and record missing or suspect observations. Uncertainty is not an optional shaded band added after a deterministic prediction has been chosen.
The companion code checks the scalar prediction and update algebra. It does not establish that Gaussian noise or linear dynamics describe a real site.
Comparing interventions adds a causal question
Running the equation with a different u produces a different simulated trajectory. That alone does not prove that a real intervention would cause the same change. The coefficient b would need evidence, and other changes might affect both the decision to intervene and the observed outcome.
A decision-support interface should therefore distinguish an observed history, a fitted forecast and a hypothetical intervention scenario. It should also state when an input lies outside the conditions used to fit or validate the model.
What a credible prototype would expose
I would expect an inspectable state definition, observation mapping, model version, data lineage, validation period and decision boundary. A user should be able to see which measurements updated the current estimate and which assumptions produced a scenario comparison.
The human decision process remains explicit. A model can help compare plausible consequences, while the responsibility for an intervention also involves material knowledge, conservation policy, resources and authorization.
The useful next step is a small prediction task with evidence of performance and limitations. Adding a world-model label before that work would not make an existing digital twin more predictive.
Complete reproducible example
This standalone example uses fictional data or prescribed mathematical objects; it is not a reproduction of the source paper’s experiments.
"""Uncalibrated scalar Gaussian teaching model; not a heritage forecast."""
import json
from math import sqrt
m,P=1.,.04
a,b,Q,R=1.05,.2,.01,.09
u,y=0.,1.2
predicted_m=a*m-b*u
predicted_P=a*a*P+Q
K=predicted_P/(predicted_P+R)
updated_m=predicted_m+K*(y-predicted_m)
updated_P=(1-K)*predicted_P
assert predicted_P>0 and 0<K<1
assert predicted_m<updated_m<y
assert 0<updated_P<predicted_P
assert abs(predicted_P-.0541)<1e-12
print(json.dumps({"predicted_mean":predicted_m,"predicted_variance":predicted_P,"gain":K,"updated_mean":updated_m,"updated_variance":updated_P,"gaussian_95_interval":[updated_m-1.96*sqrt(updated_P),updated_m+1.96*sqrt(updated_P)]},indent=2))