Guide · Trust
Trust tiers and medallion architecture: bronze, silver and gold
Medallion architecture is a good engineering convention that has quietly been promoted into a trust rating. Bronze, silver and gold describe how much processing a table has had. They say nothing about whether the number in it is right, and the gap between those two ideas is where confident mistakes come from.
The short answer
Medallion architecture is a storage layout: bronze holds data raw as it arrived, silver holds it cleaned and conformed, gold holds business-ready tables shaped for reporting. It describes processing, not correctness — a gold table built on a wrong definition is a confidently wrong number. A trust tier is the separate, business-facing promise about how much a decision-maker may rely on a dataset, and something has to check the two agree.
What the three layers actually are
Bronze
Data as it arrived, unaltered, with whatever mess the source system had. The point is fidelity: if a transformation turns out to be wrong six months later, bronze is what lets you rebuild without going back to the source.
Governance consequence people miss: bronze is where the identifiers live. Because nothing has been cleaned or dropped yet, the raw layer is routinely the most sensitive thing in the estate, full of IP addresses, device identifiers and free-text fields containing things nobody intended to collect. It is also the layer most likely to be left open to the whole analytics team.
Silver
Cleaned, de-duplicated, typed, conformed to shared keys. This is where a customer from three systems becomes one customer, and therefore where most of the interesting decisions get made — and hidden. The matching rules in the silver layer are business policy expressed as code, usually without anyone having agreed the policy.
Gold
Aggregated, shaped for consumption: the tables reports actually read. Gold is where the measures live, which is why an error at this layer propagates instantly to everything a director looks at.
Why gold does not mean correct
The layers describe distance from the source. Nothing in the definition of "gold" says a human checked it, that the measure matches the one the board agreed, or that the data underneath was fit for the purpose it is now being used for.
The characteristic failure: a gold revenue table, built correctly, joining correctly, aggregating correctly — on a definition of revenue that finance stopped using in the last restructure. Every technical check passes. The number is wrong in a way that no pipeline test can catch, and it carries the visual authority of the gold layer while being wrong.
That is the case for a second, separate label.
What a trust tier promises
A trust tier is a statement to a reader, not a description of a pipeline. Three levels work:
| Tier | What it promises | What it takes to earn it |
|---|---|---|
| Raw | Nothing. Explore it, do not decide on it. | Registered, owned, classified. That is all. |
| Conformed | Definitions and keys agreed; safe for operational reporting. | Quality rules running, a named steward, a documented definition. |
| Trusted | Safe to put in front of a customer, a regulator or the board. | All of the above plus an owner's sign-off, a reconciliation against an independent source, and a review date. |
The load-bearing part is the third column. A tier that can be self-declared is decoration. Promotion to the top tier should require a named person to sign, on a date, having checked something specific — which is exactly what the medallion layers do not require.
Why not just call them bronze, silver and gold?
Because your finance director is not going to. The medallion names are engineering vocabulary and they carry an implication of quality — gold is obviously better than bronze — that the definition does not support. A business-facing label should say what a reader may do with the data, in words that mean the same thing to an accountant and an engineer.
Keep the medallion names in the platform where they belong, and map them once, publicly, so nobody has to guess:
- Bronze tables are almost always Raw.
- Silver tables are usually Conformed, and sometimes still Raw if the conforming rules are unagreed.
- Gold tables are Conformed until someone signs them off, at which point they become Trusted. Gold does not confer the top tier on its own.
That last line is the whole point of running two ladders.
Making promotion mean something
Write a short promotion procedure and hold to it. Ours has five checks, and they are deliberately dull:
- The definition is written down in business terms and the owner agrees with it.
- The quality rules run on a schedule and their current results are visible.
- The figure has been reconciled against an independent source at least once, and the reconciliation is dated.
- An owner has signed the promotion, by name, on a date.
- A review date is set — annually, or on any change to the definition or the upstream schema.
Demotion needs to be possible too, and used. The first time a trusted dataset is publicly moved back to Conformed because its steward left and the checks stopped running, the tier stops being decorative for everyone watching.
If you have no lakehouse at all
Plenty of businesses reading this run Power BI on top of application databases with no medallion anything. The tiers still work, and arguably work better, because they attach to reports rather than tables: this dashboard is Trusted, that one is somebody's working analysis. Labelling the twenty most-viewed reports in your tenant is a week's work and removes a category of argument permanently.
Common questions
What is medallion architecture?
Does gold mean the data is correct?
What is the difference between a trust tier and a medallion layer?
Can you use trust tiers without a data lake?
Where the product comes in
Promotion to the trusted tier is a procedure with an approver
Lake On Rails carries the tier on the dataset record, publishes the criteria for each tier, and ships a written promotion procedure with the approval step in it. Moving a dataset between tiers is recorded in the audit trail with the previous value and the person who did it, so a demotion is as visible as a promotion.
The first step costs you nothing
Forty-five minutes with whoever runs your reporting
We tell you honestly whether this is worth doing at all, and roughly what it would take. If the answer is not yet, you will hear that. "Not for us" is a fine outcome, and a better one than a slow maybe.