Computation-Only Collaboration
Two parties analysing their combined data without either receiving the other's records, by permitting only queries whose outputs are aggregate.
Two organisations want to know the overlap between their customer bases. Neither may hand over its customer list, and both would like the answer.
A clean room provides a controlled environment where both datasets are present but neither party can read the other's rows. Only approved query templates run, and outputs are constrained: aggregates only, with minimum group sizes so a result cannot identify an individual.
The constraints are what make it safe, and they are also what people try to negotiate away. Aggregation thresholds prevent a query returning a group of one. Query approval prevents a sequence of narrowly different queries reconstructing individual records by differencing — which is the attack that defeats naive implementations. Rate limiting bounds how much can be extracted overall.
Two cautions worth carrying. A clean room controls what can be computed; it does not remove the legal basis question of whether the joint processing is permitted at all. And the guarantees depend entirely on the operator's implementation of those constraints, which is a due diligence exercise rather than an assumption — "clean room" is a category name, not a certification.