Say what you want. The model finds who, you decide where.

A model trained on raw transactions picks households far better than rules you write by hand — and it can only be trained here, because nowhere else is anyone allowed to see those transactions. Only identifiers ever leave, sent from inside the enclave with the retailer's own credentials.

Disclosure budget Every model run and every refinement spends budget, because each one tells you something new about a population you cannot see. Thresholds protect a single query; this protects the sequence. Meridian can watch this counter in their own console. To be precise about what this is: a counted allowance on access, not differential privacy — it bounds and attributes probing rather than carrying a formal mathematical guarantee, and the manifest records which regime applied.

The goal, in your own words

No SQL, no rules

Costs 2 of your quarterly budget

This demo answers these four

What this does not do

The model does not explain itself per household. The drivers shown are aggregate importances across the selected group; no reason code is ever produced for an individual.
It cannot invent demand. It ranks households by how likely they are to respond, which is useful when you are choosing who to reach and useless as a forecast of a product nobody has bought yet.
Predicted response is measured against a rule-built segment of similar size on historical campaigns, not against a randomised holdout of the campaign you are about to run. Measure that one afterwards on the lift screen.
Refining the audience repeatedly against a live counter is exactly how an attacker would probe a population. That is why the budget above is not decoration, and why it is charged to you by name.

What to do with this

Start in Tier A. The retailer's own media costs you nothing in disclosure and nothing in approval friction. Prove the audience works there before asking anyone to sign for Tier C.
Hold back a control group. Predicted response is not measured response. Exclude a random tenth of the audience from the campaign so the lift screen has something honest to compare against.
Spend the refinements deliberately. Each one costs allowance that the retailer can see. Three considered refinements read as diligence; twenty read as probing, and that is how they will be read.
Re-run the model each quarter, not the audience. A saved audience decays. The goal that produced it does not.

Proof of this run Model, datasets, policy, destinations, enclave quote

Demonstration screen built on synthetic data with the structure and distributions of a regional grocery loyalty file. Retailer, brands and figures are invented. What is real is the method: a model trained and scored inside the enclave, destinations constrained by a signed allowlist, egress with the data owner's own credentials, and a manifest that records where every audience went.