There is a kind of project I have seen succeed many times, and I have learned to watch it with mixed feelings. Two organisations need to exchange data, their models describe the same things differently, and a small team builds a mapping between them. Within a few weeks the exchange works. The team demonstrates it, the programme report records one more connection, and everyone involved is right to be pleased. A year later the report records a dozen connections, and the programme counts them as a dozen steps towards interoperability.
Then one of the models at the end of those bridges publishes a new version, and somebody has to check every mapping that leads to it. Some are checked in time. Others keep running and quietly stop meaning what they used to, and nobody notices until an exchange produces a result that nobody can explain.
The cobra bounty, again
In the first part of this series I told the story of the cobra bounty in colonial Delhi. The government paid for every dead cobra, people began breeding cobras to collect the reward, and when the bounty was withdrawn the breeders released their stock, leaving the city with more cobras than before. Goodhart’s law, in the phrasing the anthropologist Marilyn Strathern gave it, explains the mechanism in one line: “When a measure becomes a target, it ceases to be a good measure.”
In our research on semantic convergence we found that each of the four main approaches our field uses to pursue interoperability rewards something countable in place of the outcome itself. For the mandate, the measure was compliance with the named standard. For bilateral alignment, the measure is the connection: the number of exchanges made to work.
A mapping is one of the most honest repairs in our field. Counting mappings as progress is how an ecosystem learns to breed them.
A mapping repairs the exchange and leaves the cause in place
A mapping works at the boundary, after both models exist. It says that this concept here corresponds to that concept there, and the exchange between the two parties starts to flow. The divergence that made the mapping necessary stays exactly where it was, in the two models, and nothing in the repair encourages either side to move closer to the other.
The reward follows the repair. A connection can be demonstrated, reported and celebrated. The decision that would have made it unnecessary, two modellers choosing the same design when they first faced the concept, leaves nothing to report, because an exchange that works from the first day looks like no work at all. So the programme rewards the people who connect divergent models, and has no way to reward the people who avoid divergence in the first place.
Every bridge is a new artifact to keep
A mapping is a new artifact, and it needs everything an artifact needs: an owner, a version, tests and a place to live. Its cost grows faster than the ecosystem. Five parties that each model a concept differently can need up to ten mappings to understand one another; ten parties can need up to forty-five.
Mappings also age. Each one depends on two models that evolve on their own schedules, and every new version at either end reopens the question of whether the bridge still holds. Studies of the linked open data landscape have found that most published alignments drift as the vocabularies at either end evolve.
Health classification shows the shape of the obligation at scale. When the World Health Organization’s ICD-11 came into effect on 1 January 2022, WHO published mapping tables to ICD-10 to help health systems move across. Those tables are serious, necessary work. They also show how far one change travels: every system that codes diagnoses, and every national adaptation of the classification, has its own bridges to rebuild, and each of them will need revisiting with the next release.
When the funding stops, the snakes get out
The cobra story has a second act, and it applies here with uncomfortable precision. Mappings are usually built by projects, and projects end. When the budget closes and the team moves on, the mappings keep running without an owner. Each new version at either end widens the gap between what the bridge says and what the two models now mean.
At that point the ecosystem holds more divergence than before the project began: the two original models, still apart, and a bridge that encodes a correspondence neither side would sign today. The count of connections in the programme report stays the same. The quality of the meaning flowing through them has quietly declined.
Mappings as evidence
Mappings carry useful information when they are read differently. Each one shows that a party has reached towards another party’s model, which makes it a signal of where convergence is beginning to form. When several parties in an ecosystem map their own terms to the same artifact, they are telling the next modeller which design their peers already lean towards.
Read that way, a growing number of mappings inside a governed ecosystem becomes an early warning of a convergence gap, and the mappings themselves point to where the gap could close. A modeller who sees, before building, that most of the peers already map to one artifact can choose that artifact directly, and the next bridge never needs to exist.
Five questions to ask of your mappings
If your programme counts connections, these questions will show what the count leaves out. Most of them can be answered from the mappings you already maintain.
- How many mappings does our ecosystem maintain, and who owns each one today? An owner who has left the programme counts as no owner.
- Which mappings exist because two parties modelled the same thing differently, and which bridge needs that are genuinely different?
- When was each mapping last checked against the current versions at both ends?
- Which artifact do most of our mappings point towards? That is where convergence is already forming.
- What could the modeller see before making the choice that later needed a bridge? If the answer is “nothing”, the next mapping is already on its way.
Keeping the bridges, changing what we count
Alignments will always have a place at the edges, between parties that will never share a model and have good reasons not to. Inside a governed ecosystem, each new mapping deserves a second look: as a repair that works today, and as a record of a convergence gap that someone could still close at the source. Counting convergence at the source, and reading mappings as evidence of where it is forming, is part of the framework we are developing at METALEVELS.
Next in this series: when mapping becomes too costly, many programmes let popularity decide. Part three looks at critical mass, and at what an adoption count actually measures.
If you would like to look at the mappings in your own ecosystem through these five questions, start a conversation with us.