Our Position
The quality of AI training data is
inseparable from how evaluators are treated.
This is not a moral argument — although we believe it on moral grounds too. It is an empirical claim about data quality. A confused, underpaid, contractually unprotected worker sitting in a feast-or-famine gig model does not produce the same quality evaluation signal as a retained, fairly compensated, well-calibrated domain expert. The research on evaluator motivation, task engagement, and inter-rater reliability supports this consistently.
The labor practices that Provenance AI commits to below are not charitable concessions to worker welfare. They are design specifications for a data production system that produces reliable output. We hold them as commitments because we believe them as design principles.
"You cannot produce expert-quality AI training data from a labor model built to minimize what experts are paid. Those are not the same business."
Compensation Commitments
What we commit to on
compensation — in writing.
Transparent per-task compensation disclosed at application
Every role at Provenance AI specifies per-task compensation and an estimated task duration range at the point of application — before any interview or onboarding. The advertised rate is the actual rate. There are no post-hire revisions to compensation terms downward.
Minimum effective hourly floor of $25
Every task type published on the Provenance AI platform must produce a minimum effective hourly rate of $25 for an evaluator working at the expected pace. Tasks that do not meet this floor are not published. This floor is calculated conservatively — using the upper bound of the expected task duration range, not the median.
Guaranteed minimum monthly hours on retainer contracts
All domain experts on retainer contracts are guaranteed a minimum number of monthly hours (minimum 10 hours/month for standard retainers, 20 hours/month for senior domain expert roles). Guaranteed hours are paid regardless of whether task volume requires them in a given month. This eliminates the feast-or-famine scheduling problem.
Quarterly quality bonuses
Experts whose inter-rater reliability scores and client satisfaction ratings exceed the target threshold receive a quarterly quality bonus of 5-10% of their quarterly earnings. Quality is financially rewarded — not just expected.
No retroactive pay reduction
Compensation rates set at contract signing cannot be reduced during the contract term. Pay rates can only be adjusted upward at contract renewal. No mid-engagement rate cuts, reclassifications, or restructurings that reduce effective compensation.
Contractual Commitments
Every commitment backed
by a contract term.
Zero pre-contract work
No expert touches a live task until their contract, NDA, IP assignment agreement, and background check are fully executed and current. Contract execution is a hard system gate — not a manual process that can be bypassed. This is not just an ethics commitment — it is also a legal IP ownership requirement.
30-day maximum background check validity
Background checks are valid for 30 days maximum. Any check older than 30 days at the planned deployment date triggers an automatic renewal requirement. Experts cannot be deployed on live tasks with stale background checks regardless of how their prior check resolved.
Written rubrics delivered before task assignment
Every expert receives the complete, version-controlled task rubric with explicit acceptance criteria before their first assignment. Rubrics are never communicated verbally or described in general terms. Experts know exactly what they are being asked to evaluate and exactly what the acceptance criteria are.
Review cycle compensation
Experts are compensated for time spent in structured review cycles when revisions are requested. Uncompensated wait time during quality review — where experts are required to remain available but are not performing tasks — is not part of our model. Retainer contracts cover availability.
Transparent performance feedback
Every expert has access to their own performance metrics — IRR scores, gold standard accuracy, client satisfaction scores — in real time. Performance improvement pathways are documented and available. Experts are never terminated for performance without prior written notice and a structured remediation period.
The Contrast
What we are
specifically not doing.
The following table documents eight documented practices at current market vendors and states our specific, contractually enforceable alternative for each. This table is included in our standard client briefing materials and is presented to every co-pilot lab before contract signing.
| Current Market Practice | Provenance AI Commitment |
|---|---|
| Per-task rate misrepresented as an hourly rate at application | Per-task compensation and duration range disclosed at application — no hourly misrepresentation |
| Workers deployed before contracts are executed | Contract, NDA, IP assignment required before any live task — hard system gate |
| Background checks 90+ days old at deployment | 30-day maximum background check validity — renewal is automated |
| Vague quality thresholds — no written rubrics pre-task | Complete version-controlled rubric with acceptance criteria before every assignment |
| Non-experts hired and role reclassified post-production | Credentials verified before assignment — role locked at contract signing |
| Unpaid wait time during review cycles | Retainer contracts cover availability — review cycle time is compensated |
| Feast-or-famine scheduling; top experts leave | Guaranteed minimum monthly hours prevent income gaps |
| AI resume screening excludes genuine domain experts | Human-reviewed credentialing — calibration session is the primary qualification gate |
Annual Worker Treatment Audit
Every co-pilot lab receives an annual report documenting Provenance AI's compliance with these commitments. The report covers:
- Expert compensation data: average effective hourly rate by domain, quality bonus payout rates, any compensation disputes and resolutions
- Contract compliance: pre-contract deployment rate (target: zero), background check validity at deployment (target: 100% within 30 days)
- Attrition data: annualized expert attrition rate by domain, exit interview themes, remediation actions taken
- Performance feedback: percentage of experts with documented performance feedback in the period, remediation completion rates
Labs may also trigger an unannounced worker treatment audit at any time under their co-pilot agreement audit rights. This right is permanent and unconditional.