Company-bound platform
Aurum at SThree
Controlled hiring simulations for fairness review, red-team testing, and audit records.
Recruitment AI safety
Recruitment AI Safety & Fairness
Recruitment AI is high-stakes because model outputs affect opportunity. Peter builds systems that test fairness before deployment and preserve human oversight where it matters.
Review surfaces
Evaluator diagnostics
Review surfaces make disagreement and failure modes visible before deployment.
Fairness signals
Bounded fairness signals
Fairness findings are framed as risk signals, not legal certification.
Red-team workflow
Adversarial harness
Adversarial testing is integrated with provenance and human oversight.
What the platform tests
Aurum creates controlled hiring simulations, then helps reviewers inspect how recruitment AI behaves under ordinary, fairness-sensitive, and adversarial conditions.
- Review surfaces make disagreement and failure modes visible rather than hiding them behind one confidence score.
- Fairness diagnostics are separated from compliance claims so weak signals do not become false certainty.
- Adversarial testing is treated as review material with provenance, not a standalone certification claim.
Why governance changes the architecture
Recruitment AI safety needs more than a model score: traceability, audit artifacts, small-sample caveats, and explicit points where a human stays accountable.
- Red-team mechanics remain private because Aurum is SThree employer work, but the public posture is clear: attack findings must be bounded and reviewable.
- Demo materials will use safe fixtures and avoid private prompts, workflows, attack recipes, or operational details.
- Fairness findings on synthetic data are framed as failure discovery and regression testing, not a guarantee of deployed-world fairness.
Self-improvement with claim discipline
The Aurum self-improvement research treats evaluator optimization as a governance problem as much as a modelling one. It works against a synthetic proxy oracle, not human hiring ground truth.
- A DSPy input-leakage audit keeps the proxy label and prior scores out of the evaluator-visible inputs, and the audit is mechanically checkable.
- MultiStep+GEPA led Phase 4 on fitness and separation, but the top-3 ensemble beat it on rank metrics in all five seeds — reported as a trade-off.
- A claim ledger and frozen manifest keep the work from claiming real-world hiring validity, legal sign-off, or universal GEPA superiority.
Continue exploring
Related work
Aurum article
The full enterprise agentic AI and recruitment evaluation narrative.
OpenAurum self-improvement research
Evaluator optimization studied against a synthetic proxy oracle, with negative results reported in full.
OpenMulti-agent systems
The agentic architecture behind audit workflows.
OpenAI engineer profile
Senior AI engineering context and the wider work.
Open