Training data and AI workflows with a human audit trail

Each process has a written rubric agreed with you before go-live. Batches are sampled at a rate we set together, reviewed against that rubric, and the measured error rate is delivered with the work.

Where two reviewers disagree, the record is escalated rather than averaged. Disagreement rate is itself a reported metric, because it usually points at an unclear rule rather than a careless reviewer.

Either. If you have guidelines, we train against them and flag the cases they do not cover. If you do not, we draft them from a pilot batch and hand them to you for sign-off.

Guidelines are living documents on our side — every escalated edge case is written back into them, so the same question is not decided twice.

Yes. Model output review, preference ranking and golden-set construction use the same reviewer pool and the same rubric discipline as annotation work.

We will tell you where a task needs subject-matter expertise we do not have in-house, rather than staffing it with generalists and hoping.

Steady-state teams run from two reviewers to a few dozen. Scaling an existing process by roughly half takes two to three weeks, because new reviewers train against a rubric that already exists.

For genuine spikes we agree a ceiling in advance rather than promising unlimited surge capacity we would have to staff with strangers.

Work happens from our own facility on managed machines, under an information security management system certified to ISO/IEC 27001:2022, with signed NDAs and a DPA in place before access is granted.

Access is scoped per project and reviewed on a schedule. We can work entirely inside your environment where that is a requirement.

Pricing is per reviewer, per month, based on the skill level the rubric demands — an attribute auditor and a multimodal annotator are not priced the same.

For well-defined, repeatable batches we can quote per record once a pilot has established the real throughput. We will not quote per record before that, because the estimate would be a guess.

Three to four weeks is typical, and annotation work front-loads more of that than most services: the guideline has to exist before anyone labels anything, and writing it with you is the first week.

A pilot batch follows, which is where the real throughput and the measured agreement rate come from. Starting sooner is possible where you already have a written rubric — but a team labelling to a guideline nobody agreed produces volume, not data.

The error rate ships with the batch rather than on request, so a problem is visible before it reaches your model. Where a cohort has drifted from the guideline, StaffCloud re-labels it at its own cost.

Contractually the remedy is re-performance or correction within a reasonable time, and a refund of the fees for the non-conforming part where correction is not practicable (Terms 20.2). For a training set that matters more than usual: a bad cohort is cheaper to redo than to inherit.

Sixty days either way for convenience (Terms 23.2), or immediately for a material breach left unremedied for thirty days.

For annotation programmes the practical question is usually the batch rather than the contract: work in progress at the termination date is payable, and the guideline, the rubric and the adjudication record are yours to take to whoever runs the work next.