Requirements

Define the task, use case, risk, labels or scoring dimensions, and acceptance criteria.

Instruction design

Write examples, edge cases, evidence rules, and a clear path for uncertainty.

Reviewer calibration

Test reviewers on representative cases before production work begins.

Production review

Use project-specific checks, hidden quality items, and confidence signals.

Disagreement handling

Route uncertain or conflicting decisions to a more qualified review step.

Delivery approval

Check output format, quality findings, corrections, and documented limitations.

Qualification

Projects can require language proficiency, domain experience, prior role experience, or formal credentials. The actual requirement is agreed with the customer.

Calibration

Reviewers work through representative examples and receive feedback before production access is granted.

Ongoing measurement

Agreement, correction rate, hidden quality cases, confidence, and pattern-specific performance can be monitored during delivery.

Limited access

Reviewers should see only the data needed for their assigned task and only for the required period.

Evidence

What one reviewed item can record

The final schema is project-specific. These fields show the type of provenance Marka is designed to preserve.

Input and outputThe original item, model response, trace, or source record.
Instruction versionThe exact rubric or annotation guidance used for the decision.
Reviewer decisionLabel, score, correction, confidence, and written explanation.
Supporting evidencePolicy passage, tool result, source text, or expected outcome.
Quality historySecond review, correction, disagreement, and final approval status.
Dataset versionThe delivery batch and change history in which the item belongs.

Quality starts with a task we can define clearly.

Send a sample and the standard you currently use. Marka will identify what needs to be clarified before a pilot begins.

Start a pilot