Fairness assessment workpaper

What does the fairness assessment workpaper produce?

Prepare descriptive group metrics, limitations and source-linked fairness review points.

Use the exact local files and task-specific inputs listed below. Preparation and review stay in this browser. Inspect all findings before preparing the finalized pack.

What do you need before you start?

• Define the authorized source population and your review rules.
• Keep the unchanged original files and record the snapshot and exclusions.
• Review missing and invalid inputs before preparing the finalized pack.

What information do you enter?

Select the actual local sources and freeze the task-specific inputs below. Keep source files and complete findings private.

• Labeled model observations. Exact headings observation_id, group, actual_label, predicted_label, source_ref.
• Approved group comparison rules. Exact headings group, reference_group, selection_gap_limit_pp, tpr_gap_limit_pp, fpr_gap_limit_pp, minimum_group_size, model_version, approval_ref.
• Analysis reference. Customer-controlled review reference.
• Analysis owner. Responsible reviewer of method and evidence.
• Observation date. Real YYYY-MM-DD date. Future source observations are unresolved.
• Population and method scope. Name source population, selected period, approved model and known omissions.
• Population declaration. Select the explicit approved working method.

What you can download

Every result must identify its supplied facts, assumptions and unresolved items.

Fictional review report (PDF). Editable review (DOCX). Complete fictional pack (ZIP). Requires tailoring and responsible review. Samples use the finalized renderer.

How does the fairness assessment workpaper work?

1. 1
   Select the actual sources

   Labeled model observations. Exact headings observation_id, group, actual_label, predicted_label, source_ref.

2. 2
   Freeze the review rules

   Count complete supplied binary outcomes by your exact declared group. Calculate selection rate, true positive rate and false positive rate. Compare absolute percentage-point gaps with the explicit reference group and approved thresholds. Show denominators, small groups and undefined metrics. Missing or invalid records prevent complete group comparisons. No protected attribute, causation, statistical significance or legal standard is inferred.

3. 3
   Inspect complete findings

   Review all source-linked results and unresolved items. Free JSON and CSV preserve complete results.

4. 4
   Keep recoverable work

   Exact source files and input strings save in this browser. Backups verify tool identity, input schema, hashes and parsed bytes before replacing current work. Prior checkpoints and historical purchases stay separate.

5. 5
   Prepare the finalized pack

   Prepare nine complete files, including both exact originals, then download them free. Review the fictional sample below.

Method and sources

Apply the steps above to the supported inputs. Use the listed source sections to review the rule basis, assumptions and limits.

• Common Format and MIME Type for Comma-Separated Values FilesSupports source CSV mechanics only. Your factors, formulas, units, quotas, population and decision thresholds are not specified by this source. Section 2 describes records, exact field counts, quoted fields and escaped double quotes.RFC Editor. Reviewed 2026-10-09.
• Classification Accuracy, Recall, Precision and Related MetricsSupports descriptive binary classification arithmetic only. It supplies no fairness threshold, protected group, lending standard, causal conclusion or authorization to make decisions. Recall, false positive rate and precision sections define rates from true and false positive and negative counts. The denominator discussion explains undefined metrics.Google for Developers. Reviewed 2026-10-09.

Common questions

What still needs review?

Calculations use supplied observations and explicitly approved methods. Original files and prior work remain in this browser. Results do not establish source authenticity, population completeness, causation, regulatory compliance or approval to make a customer, supplier, lending or model decision.

Labels are exactly true or false. Thresholds use percentage points. Reference-group choice and contextual mitigations need responsible review.

Source limits are 10 MB, 5000 rows, 50 fields, 100000 cells and 5000000 parsed characters per file. Plain decimal inputs support 12 whole digits and 6 fractional digits.

Calculations use exact rational arithmetic and round displayed values half away from zero to six fractional places unless an explicit payment rounding rule applies. Decisions compare unrounded values.

Complete results support at most 20000 findings and 15000000 characters. Formal reports show the first 25 rows per calculation output, 1000 characters per calculation preview, and the first 100 named checks with 400 characters per check field. Excel and JSON retain all values.

Related working records

The related workspace contains working administrative records. Open it for the related administrative task.

Related tools

Browse all ai governance tools.

• Ground-truth answer-key managerVersion answers, rubrics and exceptions. Try the free example, then enter your own inputs and review the result before exporting.
• Evaluation run comparatorCompare results across models, prompts and versions. Try a free example, then review your inputs and the result. No account or payment required.
• Fairness metrics calculatorCalculate selected measures from datasets. Try a free example, then review your inputs and the result. No account or payment required.
• Regression acceptance checkerApply release thresholds and flag failures. Try a free example, then review your inputs and the result. No account or payment required.
• Model evaluation sampling plannerPrepare a defensible test design for approval. Try a free example, then review your inputs and the result. No account or payment required.