Solutions · Regulators
See the bias in the data you supervise. Free, independent, reproducible.
Rosa gives supervisory authorities a standard, defensible instrument for measuring bias in the datasets and models of the firms they oversee, without taking on the firms' data risk or building tooling in-house.
The supervision problem
No standard instrument
Bias measurement across firms is ad hoc. Without a common, defensible yardstick, findings are hard to compare and harder to enforce.
Firms self-report
Supervisors mostly see the firm's own account of its data. Independent measurement closes the gap between assurance and reality.
Bias is invisible until harm
Encoded bias surfaces as unequal outcomes long after the data decisions were made. Detection has to happen at the dataset, before the harm.
What Rosa gives a supervisor
- A standing free Diagnose trial. Diagnose is free for supervisory authorities on the shared trial instance, with no end date: a standing commitment, not a time-limited offer. Point it at a dataset and one protected attribute; receive a measured bias score plus per-feature association and debiasing-adjustment scores, and a PDF Dataset Intake Report.
- A dedicated instance for sensitive supervisory work. The free shared trial is for evaluation and non-sensitive datasets. Diagnosing real data from the firms you supervise belongs on a dedicated Rosa instance in your own region, under your own controls. Dedicated instances are a paid engagement, scoped in conversation.
- An independent, reproducible measurement. The same data and configuration produce the same examination on anyone's desk. The methodology is open to inspection, not a black box you are asked to trust.
- A manifest a firm cannot dispute. Every run emits an immutable Run Manifest with the SHA-256 hash of the input file. The firm can recompute the hash and confirm exactly what was examined; so can you.
And the wedge logic, stated plainly: it is free to you because when the supervisor can measure bias, supervised firms adopt the same lens. Your use of Rosa is what makes the market take dataset bias seriously. There is no catch beyond that.
The questions you should ask
Is it scientifically defensible?
Rosa is built on the Fair Adversarial Network (FAN), an adversarial method with years of research and validation behind it: a discriminator repeatedly attempts to recover the protected attribute from the data, and its success rate is the bias measure. The published figures, the methodology, and the live product are all inspectable. See the validation page.
Where does the data go?
The shared trial processes data in the UK (AWS eu-west-2, London), under the EU-UK adequacy decision, and is intended for evaluation and non-sensitive datasets. For supervisory work on real firm data, a dedicated instance runs in your own region, on infrastructure you choose; dedicated instances are a paid engagement.
Are you neutral?
Rosa measures; it does not advise the firm how to pass. The instrument reports what the data encodes, identically for every party. Independence is structural: the same run, the same numbers, the same manifest, whoever submits it.
A defensible bias instrument, free to supervisors.
Diagnose is free for supervisory authorities on the shared trial instance, with no end date. Dedicated instances for sensitive data are a paid engagement.