Differentiators
What sets Rosa apart.
Not another governance dashboard. The engine that fixes the data.
Eight differences that survive scrutiny
It removes bias. It does not just measure it.
Governance platforms and open-source toolkits flag bias and leave the fix to you. Rosa transforms the data, so the audit finding closes instead of being monitored forever.
Contrast: governance/GRC platforms and fairness toolkits stop at assessment.
See the COMPAS result →No proxy labelling required.
You name only the protected attribute. Rosa detects and removes every proxy-encoded correlation itself (a proxy is a feature that stands in for the protected attribute, like a postcode for race), even when the protected column was deleted years ago and the bias persists through proxies.
This is the single strongest technical differentiator.
How Diagnose works →Your data's statistics are preserved.
Rank-mapping keeps each column's distribution: bit-for-bit in training output (float64 precision), to roughly 1e-6 in inference. Your data stays usable and your model stays calibrated.
Methods that perturb or synthesise values cannot make this promise.
Data integrity figures →No fairness/accuracy trade-off.
Parity-constraint methods fight the accuracy objective and degrade it. Rosa treats fairness as a high-dimensional transformation of the data, not a constraint opposed to accuracy. In the controlled experiment: R-squared 0.747 with Rosa plus a model, vs 0.626 for the model alone, vs 0.569 with a parity constraint.
Phrased precisely: no trade-off, not "wins every metric".
The controlled experiment →Compatible with any model or stack.
Rosa outputs a clean CSV. No wrapper library in your training code, no model surgery, no vendor lock at the framework level.
The standard objection to toolkit approaches is the integration burden; Rosa removes it.
Pipeline integration →Data residency you choose.
The shared trial is processed in the UK (eu-west-2, London; EU-UK adequacy). A dedicated instance runs in your own region for PoC and production. Critical for regulated sectors.
Residency is explicit and chosen, never incidental.
Compliance posture →Works for human decisions, not just ML.
Rosa cleans data used by analysts making direct decisions (credit scoring, shortlisting) as well as data fed to models. The fairness problem is in the data, wherever the decision happens.
Model-monitoring tools cannot see the analyst's spreadsheet.
Use cases →Evidence is built in.
Immutable Run Manifest plus PDF report on every run, including failures. Audit-ready by construction, not by an end-of-quarter documentation sprint.
Proof is a byproduct of doing the work.
The Run Manifest →The comparison, kept fair
Capability by capability against the two categories Rosa is usually compared with. Where they genuinely do something, the table says so.
| Capability | Open-source fairness toolkits | AI governance / GRC platforms | Rosa |
|---|---|---|---|
| Detects bias | Yes | Yes | Yes |
| Removes bias at source | Partial: mitigation algorithms you assemble and tune | No | Yes |
| Preserves data statistics | Method-dependent | Not applicable | Yes, bit-for-bit in training output |
| No proxy labelling needed | No: fairness definitions and groups configured by hand | No | Yes |
| Works with any model or stack | Partial: integrates into your training code | Yes (monitoring layer) | Yes (clean CSV out) |
| Emits immutable audit evidence | No | Partial: documentation and attestations | Yes, hash-verifiable manifest per run |
| Runs as a live service (REST + MCP) | No: libraries | Yes (SaaS) | Yes |
| Free to regulators | Free to everyone, as code | No | Yes, Diagnose free with no end date |
The technology, briefly
Rosa is built on the Fair Adversarial Network (FAN): a discriminator repeatedly tries to recover the protected attribute from the data, and the network is trained until it cannot. In Rosa, that technology is productised as a dataset transformer: it takes your CSV in and outputs a debiased CSV.
The adversarial principle is what makes proxy detection automatic: the discriminator will exploit any feature that leaks the protected attribute, so the transformation has to neutralise all of them, named or not.
See the validation evidence →Remove the bias from your data. Keep the evidence.
Free to try. Free for regulators, with no end date. No credit card.