Clear labeling for AML Models.

MRM has a question.

The model documentation says there were 1,250 below-the-line samples. The development file contains 1,237.

Why?

And off you go. Another time sink. A potential problem.

The problem

A familiar story.

You pull up the files from six months ago. There was the original sample file, the version sent to the first‑line reviewers, the file returned after their review, another sent for second‑level review, an adjudication file, and the final file used by the modeling team. Someone corrected labels along the way. A handful of records were removed. Another set was added. There may have been a sampling adjustment.

Somewhere in that chain is the answer. You have to reconstruct it.

What should have been a straightforward MRM question turns into hours—or days—of forensic reconciliation.

The GoldLabel solution

Seven components to reflect your process.

A model build is not the same job at every institution. You configure the states your programme runs on, and AML GoldLabel enforces the order, the allocation and the approvals from there.

Separation of duties is enforced by the system: the person who executes a state cannot approve it.

Population

A snapshot of the records to be included in your development dataset.
Sampling

Define your above-the-line and below-the-line populations, then select the sampling approach you want applied to each.
Label

Define your data labeling team, allocate samples for review, and approve the results.
Train

Initiate the training of your model through AML GoldLabel, on your own in‑house development platform.
Threshold selection

Choose the threshold that meets your risk and efficiency preferences by analysing the precision and recall trade‑offs.
Validation

See how your model performs on holdout data.
Sign‑off

Track sign‑offs throughout the process and on the final model.

Approval gate — the state cannot advance until it is signed

Labels

We specialize in labeling.

Custom review levels

Set as many levels of review as your programme requires. Two is common and the system does not assume it. Each level’s decision stands in its own right; none overwrites the one before it.

Disagreements are tracked, not resolved away

Where reviewers differ, the difference is recorded as a fact about the record: who held which position, when, and on what basis. Disagreement rates are visible across the whole label set rather than buried under a final answer.

Overturns are events, not edits

A changed decision carries its own comment and its own timestamp. Nothing is silently overwritten, so the history of a label reads back in the order it happened.

Every disposition carries its reason

Analysts write why, not only what. The rationale travels with the label into training, into validation, and into the model documentation.

Process integrity

We do more than labels.

Labeling is where the judgment sits, but a label set only earns its keep once a model is built on it. The four states after Label — train, threshold selection, validation and sign‑off — run inside the same workflow and under the same approvals, so the label set and the model that learns from it become one integrated audit trail.

Training runs on your own development platform, with the tools and the people you already have. AML GoldLabel initiates the run and records it — the code and its hash, the hyperparameters, a checksum of the dataset — rather than replacing a modeling stack your institution has already built and validated. The threshold is committed with author and timestamp before the model is measured against out‑of‑time data, so what gets reported is a measurement and not a selection.

None of it is assembled at the end. Each state writes its evidence as the work happens, so what an examiner reads is what the team actually did, in the order they did it. Six months on, the answer to an MRM question is already sitting in the record. That is where the peace of mind comes from: the work stands up because it was documented while it was being done.

AML GoldLabel

See a label set built end to end.

An interactive tour walks a full run, screen by screen, from population selection through officer sign‑off. The demo environment uses synthetic transaction data; the screens and the metrics come from real executions.

Watch the demo goldlabel@clarusforge.dev