Frequently asked questions

Harbor is human intelligence infrastructure for frontier AI. We run three solutions on one platform: the Data Network (real-world capture), the Expert Network (specialized judgment), and Managed Programs (fully operated delivery). Organizations use Harbor to source data, deploy experts, and receive eval-ready outcomes. The Expert Network is where vetted specialists contribute evaluation, validation, annotation, and review work.

The Expert Network supplies high-signal human judgment: rubric-based evaluation, domain validation, red teaming, model review, and quality assurance. The Data Network focuses on field and studio capture—contributors recording real-world image, video, audio, and multimodal data. Some contributors do both; invitations are always program-specific. Managed Programs may include dedicated expert cohorts operated by Harbor when a client needs full end-to-end delivery.

Practitioners with verifiable domain expertise—not generic labelers. Strong profiles include engineers, clinicians, lawyers, linguists, safety reviewers, robotics specialists, and multilingual evaluators who can apply published rubrics consistently. Harbor matches on demonstrated skill, not interest alone.

Submit your background through the Talk to Harbor contact form at harborml.com/book-a-demo. Include your domain areas, languages, credentials, and portfolio or work history where relevant. Harbor operations reviews applications for fit with active and upcoming programs. Approved contributors complete payout setup and workspace activation before accepting paid work.

No. Joining the Expert Network is free. You are never charged to apply, onboard, or receive invitations. Compensation flows from Harbor to you for accepted program work.

Vetting confirms identity, tax residency, domain claims, and program fit. Steps typically include profile review, credential or work-sample checks for specialized domains, and compliance screening. Robotics, clinical, legal, and safety programs may require additional verification. Harbor does not publish client names during vetting.

Most applications are reviewed within one to two weeks, depending on domain complexity and how quickly you respond to follow-up questions. Niche specialties with active client demand may move faster; new program areas may take longer while operations builds the initial cohort.

Be specific: name the domains you can judge (e.g., robotic manipulation evals, clinical note review, bilingual voice QA), languages and locales, years of practice, and tools you have used. Link portfolios, publications, or anonymized work samples where allowed. Generic “interested in AI” profiles rarely match frontier programs.

Yes, when you have demonstrable domain depth—research experience, internships, competition results, or published work in a relevant field. Student status alone is not a match signal; program rubrics require reliable judgment at production standards.

Harbor runs global programs where local law and client requirements allow. Tax residency, payout country, and export or data-handling rules determine eligibility for specific invitations. Your onboarding profile captures residency so matching respects geographic constraints.

You may receive a decline or a request to reapply when your profile strengthens—for example, after more domain experience or when Harbor opens programs in your specialty. Operations may also route strong capture profiles toward Data Network programs instead of expert evaluation work.

Your profile enters Harbor’s matching system. When a program needs your skill set, you receive an invitation with scope, rubric summary, time expectations, compensation structure, and confidentiality rules. You review and accept or decline before any client data is shared.

Matching is skill-based: verified expertise, language, timezone overlap, capacity, and historical quality on similar rubrics. Client demand drives volume—robotics evals, vision validation, red teaming, and multilingual review scale independently. Harbor does not assign work purely by availability; fit to the rubric comes first.

No program volume is guaranteed. Invitations depend on client pipelines, your domain, and quality history. Contributors with rare specialties often see fewer but higher-rate programs; contributors in high-demand eval domains may see steadier flow. Building a strong quality record improves re-invite priority.

Some contributors receive invitations within days of activation; others wait weeks until a matching program launches. Timing is driven by client demand, not onboarding order. You are notified by email and in the contributor workspace when relevant programs open.

Yes—always. Declining does not penalize your profile. Accept only work that fits your schedule, expertise, and confidentiality comfort. Repeated accepts followed by non-delivery without notice can affect matching priority.

Scope varies by program. Some invitations are a few hours of focused evaluation; others are multi-week part-time engagements with daily throughput targets. Each invitation states expected hours, deadlines, and whether work is synchronous or async before you accept.

Often yes, if schedules and conflict-of-interest rules allow. Some programs require exclusivity during sensitive eval windows—those constraints are disclosed upfront. Do not accept overlapping work when an NDA or program brief prohibits it.

Most programs run in the Harbor contributor workspace: task queues, rubrics, annotation viewers, and program messaging. Occasional programs integrate external tooling; requirements are listed in the invitation. You need a modern browser, stable internet, and sometimes a headset or second monitor for AV-heavy evals.

Expert Network contributors are paid as independent contractors through Harbor’s approved payout partners. Before your first paid task, you complete tax and banking profiles in the workspace. Rates and schedules are defined per program in the invitation—not retroactively changed after you accept.

Rates reflect rubric complexity, domain scarcity, turnaround time, and client program budget. Evaluation and red-team work typically pays above commodity labeling. Harbor publishes compensation structure in each invitation so you know the basis—per task, per hour, or milestone—before starting.

Payout schedules are program-specific and visible in your workspace. Most programs batch payments on regular cycles after work passes quality review. Incomplete tax or banking setup blocks payout regardless of delivery status.

Yes. As an independent contractor you are responsible for tax reporting in your jurisdiction. Harbor collects tax forms required for payout compliance (e.g., W-9, W-8BEN) but does not provide tax advice.

Payout currency depends on your tax residency and available Harbor payout rails in your country. Supported options are shown during payout setup. Currency for a given program is fixed in the invitation.

Use the in-program thread first for task-level questions. For payout blocks or quality disputes after delivery, follow the disputes path in contributor documentation or contact Harbor operations with program ID and delivery timestamps. Decisions reference published rubrics and reviewer consensus.

Active and emerging programs span Robotics, Computer Vision, Industrial Operations, Manufacturing, Safety, Healthcare, Legal, Finance, multilingual evaluation, agentic systems, and adjacent specialties where client rubrics require credentialed judgment. Demand shifts with frontier AI roadmaps—robotics and multimodal evals are common; niche legal or clinical programs open as clients need them.

Expert Network work includes evaluation, annotation, validation, red teaming, model review, quality assurance. Tasks range from rating model outputs against rubrics, validating capture quality, reviewing safety edge cases, and annotating multimodal training data where domain knowledge changes the label. You always receive instructions and examples before production work begins.

No. Harbor Expert Network programs prioritize judgment quality over throughput. Many tasks look like “labeling” but require domain reasoning—deciding whether a robotic grasp plan is feasible, whether a clinical summary is safe, or whether a model response violates policy. Data Network capture work is separate and listed as such in invitations.

Helpful but not always required. Client rubrics often need your subject-matter expertise applied to model outputs. ML familiarity helps on eval programs; clinical, legal, or robotics credentials help on domain programs. Your application should emphasize the expertise the rubric will test.

Yes. NDAs and program briefs apply to every engagement. You may not disclose client identity, model details, prompts, datasets, or outcomes publicly or on social media. Screenshots and exports are prohibited unless the program explicitly allows them.

You may describe the type of work at a high level—e.g., “model evaluation for robotics” or “multilingual voice QA”—when it does not reveal client identity or confidential program details. Follow each program’s publicity rules; some engagements require stricter silence.

Against published rubrics, gold-standard items, and reviewer consensus—not self-assessment. Programs may include calibration tasks, inter-annotator agreement checks, and spot audits. Consistent alignment with rubrics drives re-invites; drift or rushed work triggers review or removal from active programs.

You may receive additional training tasks, a reduced scope invitation, or removal from that program cohort. Failing calibration does not automatically ban you from Harbor, but repeated failures in the same domain reduce matching priority.

Yes, for confidentiality breaches, fraud, repeated quality failures after feedback, harassment in program channels, or attempting to circumvent payout or data-handling rules. Harbor may also pause profiles pending investigation of disputed deliveries.

Deliver on time, follow rubrics precisely, ask clarifying questions early, and maintain calibration scores. Contributors who reliably handle complex eval work are prioritized when clients expand programs or launch follow-on phases.

Profile, tax, and banking data are encrypted and used only for vetting, matching, payouts, and compliance. Harbor does not sell contributor data. Client-facing deliverables are de-identified—you are not listed in client exports without program-specific consent.

Only what a program requires to complete rubric tasks—prompts, media, model outputs, or capture samples inside the workspace. Data must stay inside approved tools. Downloading, forwarding, or reusing program data outside Harbor systems is prohibited.

Harbor follows applicable privacy regulations for contributor and client data processing. Regional restrictions may limit which programs you can access based on residency and data classification.

Unless a program brief explicitly allows it, no. Client rubrics often require unassisted human judgment. Using external LLMs to complete eval tasks without permission is treated as a quality and confidentiality violation.

Yes. Harbor’s referral program pays fixed bounties for net-new specialists who pass vetting and complete qualifying work. Amounts are visible per role before you refer. See harborml.com/refer-and-earn for current terms; existing Harbor profiles do not qualify.

Use the in-program message thread for rubric and task questions. Contributor documentation at harborml.com/docs covers payouts, matching, and standards. Contact operations when workspace access fails, payouts are blocked after setup is complete, or you need to report a security concern.

Yes. Harbor maintains contributor docs covering onboarding, the three solutions, program matching, payout setup, quality standards, disputes, and release notes. Start at harborml.com/docs for operational detail beyond this FAQ.