MarieRosa
For AI labs

Executive judgment from the whole leadership population.

Marie&Rosa supplies identity- and title-verified women and non-binary executives, VP and above, for evaluations, rubrics, preference data and decision scenarios. Every item comes with a provenance record.

Request a conversation

What we deliver

Preference dataPairwise and ranked comparisons on leadership and strategy prompts, with written rationales.
EvaluationsScored model outputs against rubrics we write with you or that you supply.
RubricsCriteria for executive-function tasks, written by people who have made those calls.
Scenario banksDecision scenarios drawn from real operating situations, with reference answers.
Red-team sessionsExecutives probing the judgment calls where your model fails.
Bespoke benchmarksAn executive-judgment benchmark built for your model, including a gender-framing measure.
Why the evaluator pool changes the model

What evaluators reward is what the model learns to produce.

When those evaluators come from a narrow slice of leadership, the model's idea of good judgment narrows with them.

We will publish our own measurements as pilots produce them.

The evidence on the current record

  • Women wrote an estimated 26.5 percent of GPT-3's training text.1
  • Of more than 8,000 model outputs about CEOs, 18 percent were confirmed biased and 94 percent of those concerned women CEOs. The bias showed up as framing, such as "female CEO" and "breaking barriers."2
  • Women hold 23 percent of C-suite roles and 19.1 percent of CEO roles in the economies the World Economic Forum measured.3 In the US, women hold 29 percent of C-suite roles.4

How we vet

Every expert passes four checks before doing paid work:

  1. Identity verification
  2. Title verificationVP or above, or the equivalent described on our executive page.
  3. A 20-minute conversationWith a person.
  4. A written exerciseScored against a rubric.

We re-verify every year and track agreement scores on every expert.

How engagements work

  • PilotTwo weeks, one executive-judgment task, a fixed fee agreed up front. You score the output against your current vendor.
  • ProgramAfter the pilot we scope volume, task mix and quality targets in a statement of work.
Provenance and compliance

Every item arrives with its record.

Each item you receive comes with who produced it (by expert ID and verified tier), when, under which task instructions, and a consent and no-employer-data attestation. We provide the evaluator pool's composition for your documentation. Contractor agreements assign deliverables to you.

Provenance record
Expert
Pseudonymous expert ID
Verified tier
VP, C-suite or board, per our checks
Produced
Date and time of submission
Task instructions
Version of the instructions used
Attestations
Consent and no employer data, per task
The fields every delivered item carries. An illustration of the format, not a client record.
Request a conversation

Tell us what you need.

This goes directly to Tran Pham, the founder.

What you need

We use these details only to reply. See our privacy policy.

Sources

  1. TechPolicy.press, "Who authors the internet? Analyzing gender diversity in ChatGPT-3 training material," 2023. A published estimate. Link to source 1 (opens in a new tab)
  2. Russell Reynolds Associates with EqualVoice, "How the story gets told: do LLMs produce biased statements about CEOs?", April 13, 2026. Link to source 2 (opens in a new tab)
  3. World Economic Forum, Global Gender Gap Report 2026, September 16, 2026, using LinkedIn data. Link to source 3 (opens in a new tab)
  4. McKinsey & Company and LeanIn.Org, Women in the Workplace 2025. Link to source 4 (opens in a new tab)