Diagnose / Remove / Prove

Rosa diagnoses bias, removes it, and hands you the evidence.

Bias hides in data, directly and through proxies. Rosa strips it out without changing your data's statistics, and every run leaves an immutable, verifiable record. Everyone else measures and manages bias. Rosa removes it.

Start free See how it works

Free to try · Run it live in the browser

Before
model race-disparity 0.54

column distribution

After
with Rosa 0.09

same distribution, preserved

−83% disparity on average Pre-conditioned COMPAS · downstream recidivism model
See the methodology

Regulators

See the bias in the data you supervise.

Free, independent, reproducible. Diagnose is free for supervisory authorities on the shared trial, with no end date.

Your path Auditors

Article 10 bias-audit evidence in one run.

A standard, hash-verifiable evidence artefact you can reuse across engagements.

Your path Enterprises

Comply before the deadline. Keep your model's performance.

Remove dataset bias and produce the evidence in one motion. UK processing, or a dedicated instance in your region.

Your path

Measured, sourced, reproducible

The numbers do the persuading.

~75-90% lower measured racial disparity in a recidivism model trained on Rosa-debiased COMPAS - a 6× reduction Pre-conditioned COMPAS, downstream model

0.747 vs 0.626 downstream R² with Rosa plus a model, vs the model alone (0.569 with a parity constraint) Controlled credit-risk experiment

95.2% bias reduction at production scale (0.8745 to 0.0421) At-scale capacity run

bit-for-bit distribution preservation in training output, to float64 precision Validation suite

R² (R-squared) measures how well a model's predictions fit reality; higher is better. Reproducible. See the methodology.

Why Rosa is different

The only tool in the conversation that changes the dataset and proves the change.

Open-source fairness toolkits

Assess and mitigate at the data scientist's desk. A toolbox of metrics and algorithms you assemble and operate yourself. No evidence artefact, no service.

Governance and GRC platforms

Measure, monitor, and document bias across the enterprise. Useful paperwork, but the dataset leaves the platform exactly as biased as it arrived.

Rosa

Removes the bias at source, preserves the data's statistics, and proves the change with an immutable, hash-verifiable record of every run.

The product in three beats

Diagnose. Remove. Prove.

01 / DIAGNOSE

Find the bias

Point Rosa at a dataset and one protected attribute. It finds where bias is encoded, directly or by proxy (a feature that stands in for the protected attribute, like a postcode for race), and scores it.

02 / REMOVE

Remove it at source

Rosa transforms the data so downstream models cannot recover the protected attribute, while preserving every column's statistics.

03 / PROVE

Keep the evidence

Every run emits an immutable Run Manifest and a PDF report. The audit evidence is generated by doing the work, not written up afterwards.

See how it works

EU AI Act · Article 10

High-risk AI must be examined for dataset bias, with mitigation in place. The deadline is 2 August 2026.

Article 10 of the European Union Artificial Intelligence Act requires the training, validation, and testing datasets behind high-risk AI systems to be examined for possible biases, with appropriate measures to detect, prevent, and mitigate them. Rosa produces exactly that examination, the mitigation, and the evidence, in one run.

EUR 35M / 7% maximum penalties under the Act for the most serious breaches; data-governance obligations for high-risk providers carry up to EUR 15M or 3% of worldwide turnover EU AI Act, Article 99 · how Rosa maps to Article 10

2 AUGUST 2026

Prepare for Article 10

Product as proof

Do not take our word for it. Run it.

The Customer Portal includes Test 1: the Apple Card gender-bias scenario, debiased in your browser in one click. Submit the dataset, watch the job run, download the debiased CSV, the PDF report, and the Run Manifest (an immutable, hash-verifiable record of exactly what Rosa did to the data).

Open the Customer Portal Read the Apple Card case

Run Manifest verified

job_id: 550e8400-e29b-41d4-a716-446655440000
mode: remove_bias_training
job_status: complete
timestamp_submitted: 2026-06-10T09:14:02Z
timestamp_completed: 2026-06-10T09:31:47Z
row_count: 2,000
bias_columns: ["race"]
input_hash: sha256:9f1c…e7a2
schema_hash: sha256:4b08…21cd
config_hash: sha256:d3aa…90f4
container_digest: sha256:71be…0c55
bias (pre): 0.21
residual_bias: 0.001
artifacts: remove_bias_report.pdf, compas_preconditioned_fair.csv

Immutable. Written once per job, kept indefinitely. Illustrative values; field set mirrors the live manifest schema.

Where bias hides

Use cases with numbers, not adjectives.

Financial services

What sets Rosa apart

No proxy labelling required. Name one protected attribute; Rosa finds the proxies itself.
Any model or stack. Rosa outputs a clean CSV. No rewrite of your training code.
Data residency you choose. UK region for the trial; a dedicated instance in your own region for a PoC.
Works for human decisions too. Cleans the data analysts use directly, not just model training data.
No fairness/accuracy trade-off. Fairness as a transformation of the data, not a constraint fighting your model.
Honest scope. Univariate: one protected attribute per run. Rosa says what it can and cannot detect.

All differentiators

Designed to support organisations implementing these frameworks at the data layer

EU AI Act ISO/IEC 42001 NIST AI RMF SOC 2 GDPR EN 304 223

Rosa is an evidence-producing instrument, not a certification. Standards and compliance

Remove the bias from your data. Keep the evidence.

Start free Talk to us

Free to try. Free for regulators, with no end date. Processed in the UK.