Human oversight
High-risk AI outputs and actions can be reviewed, corrected or authorised by certified operators before they reach customers or regulated processes.
Controls
Assurance that can be inspected.
These controls apply to OmniHumanLoop Academy, the Task Laboratory, certification and production operations. They are specified for implementation and are not reduced for commercial speed.
Responsible AI
Safety, fairness, bias, disclosure and escalation rules sit inside the operating model, not only in policy documents.
Data security
Training uses synthetic, de-identified or specifically approved data. Live client data is separated from laboratory and calibration environments.
Privacy and consent
Privacy, consent and security questions are treated as critical controls. Missing a critical control fails the task regardless of overall score.
Protected gold answers
Gold decisions, gold-standard answers, error categories, severity, critical-control flags and reviewer notes are never visible to learners before submission. Protection is role-based, not cosmetic hiding.
Role-based permissions
Separate permissions for learner/operator, reviewer, adjudicator, coach/team leader, administrator, client owner and privacy/security owner.
Audit trails
Task IDs, decisions, evidence, corrections, escalations, reviews and adjudications are recorded for quality and client audit.
Critical-control rules
Automatic failure where a learner misses a critical privacy, consent, safety, security, mandatory-disclosure or authority control.
Reviewer independence
Reviewer and adjudicator workflows sit above the operator, with calibration tracking and protected gold standards.
Training-data separation
Training, calibration, shadow production and live client data remain separated, with version history for content, rules, gold answers and scoring.
Worker wellbeing and ethical data work
Human-in-the-Loop work is designed as paid, supervised production — not ungoverned crowd labelling. Coaching, quality and escalation are part of the operating model.
Incident and escalation management
Exception handling, escalation selection and adjudicator review are built into the standard task workflow.