Report onboarding

Goal

Three things are true before the new audience sees anything, and each is checked, not assumed:

1. Every report they were promised exists and runs on its schedule.
2. They see their own data and nothing else — verified in isolation, by querying or logging in as them, wherever a scoping rule applies.
3. Every headline number has been validated against a source they can check themselves, and which number was validated against what is recorded.

Procedure

1. Get what they need to see and who reads it. An operations lead and an executive want different reports, and building one and calling it both satisfies neither.
2. Confirm the data exists for them before promising anything. A new audience often has weeks rather than years of history, and a trend chart built on three weeks is misleading in a way that is hard to walk back once they have seen it.
3. Set up access and scoping first, and verify it in isolation. Log in as them, or query as them, and confirm they see their data and nothing else. One audience seeing another's data is the failure that costs most here, and it is cheapest to prevent at this step.
4. Add them to the pipeline. Which scheduled jobs have to include them, what the daily reshaping needs, and an entry in memory/conventions/pipelines.md with their expected output volume so the morning check can tell when their run is wrong.
5. Build from the standard report set before anything bespoke. A bespoke report built first becomes the thing that has to be maintained forever, and it usually turns out the standard one would have done.
6. Validate every number against a source the audience can check themselves, and say which ones you validated and against what. Their own count of sites, their own complaint log, their own invoice.
7. Run it for a full cycle before handing over, so the first real scheduled run happens while somebody is watching rather than on a Monday morning in front of the audience.

Boundaries

• Never enable a report before verifying data scoping in isolation, wherever a scoping rule applies.
• Never include another audience's data in any report or any aggregate, however useful the benchmark would be.
• Never ship a trend over a period too short to carry one. Say the history is thin instead.
• Never build bespoke before the standard set is working.
• Never hand over a report you have not watched run on a real scheduled cycle.

What to record

Keep one entry per audience in memory/conventions/audiences.md: their sites and business units, which reports they get and who reads each, their scoping rule, the pipeline jobs that include them, and which numbers were validated against what. Add their expected output volume to memory/conventions/pipelines.md at the same time, since an audience with no baseline cannot be checked.