Solutions · Auditors

Article 10 bias-audit evidence, generated from the data, in one run.

From August 2026, your clients must show their high-risk AI datasets were examined for bias, with mitigation in place. Rosa turns that examination into a standard, repeatable audit artefact you can stand behind.

Why bias audits do not scale today

Bespoke every time

Each engagement reinvents the measurement: which metrics, which thresholds, which caveats. Slow, expensive, hard to defend consistently.

No reusable artefact

The output is a one-off report tied to one team's judgement, not a standard evidence object your methodology can cite engagement after engagement.

Clients resist

Bias remediation looks like a research project to the client. An audit that only finds problems, without a path to fixing them, is a hard sell.

The repeatable evidence motion

  • One run, one artefact. Run a client dataset through Diagnose and receive the immutable Run Manifest plus the PDF Dataset Intake Report: a measured bias score plus per-feature association and debiasing-adjustment scores, in a fixed, versioned format.
  • Independently verifiable. Every artefact can be verified from the original file: the manifest's input hash is the SHA-256 of the bytes your client gave you. Your risk function can check it; so can the regulator.
  • The audit stays clean. Diagnose examines; it does not alter the client's data. When remediation is needed, Remove preserves every column's distribution, so before/after comparisons remain meaningful.
  • Embeddable. REST API and an eight-tool MCP (Model Context Protocol) server, so the run can sit inside your own audit workflow rather than alongside it.
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.

Fit with your methodology

Will this hold up inside our audit methodology?

The artefact is stable, versioned, and hash-verifiable: manifest fields, model format versions, and report structure are fixed per version and documented. The measurement itself is reproducible and the validation evidence is published. The right next step is a working session with your methodology owners; that is the conversation we ask for.

Can we embed or white-label it?

The REST and MCP surfaces are built for embedding into an audit workflow. Channel terms, branding, and engagement structure are a partnership conversation, not a self-serve checkout; talk to us.

Does it scale across clients?

The same run produces the same artefact for every client: that is the point. Per-client variation lives in the configuration (which protected attribute, which columns), and the manifest records the configuration hash so each engagement's evidence is self-describing.

Make bias audit a product, not a project.

Every artefact is independently verifiable from the original file.