# Pymetrics (Harver Games): HireAIScore fact sheet

URL: https://www.hireaiscore.com/vendors/pymetrics
Fact sheet: https://www.hireaiscore.com/vendors/pymetrics/facts.md
Last reviewed: 2026-11-14
Rubric: v1.0 (https://www.hireaiscore.com/methodology)

> HireAIScore rates Pymetrics (Harver Games) 50 out of 100 (grade F, Substantial gaps) under rubric v1.0, last reviewed November 14, 2026. Pymetrics (Harver Games) ranks 20 of 82 vendors rated in Assessment AI, against a category average of 46. That score is drawn from 26 evidence items across 7 rubric criteria - 19 with a cited source, 7 recording that nothing was located.

This file is the text of the profile at the URL above, for AI assistants and other software. Scores are an editorial assessment of public evidence against a published rubric. They are not legal advice, a certification, or a statement that a product is or is not compliant.

## Score

- Overall: 50/100, grade F (Substantial gaps)
- Rank: 20 of 82 in Assessment AI (category average 46)

## Criterion scores and evidence

### Article 11 Technical Documentation: 60/100 (weight 20%)

- Documentation (captured 2026-11-14): Wilson, Mislove et al., FAccT '21: case study auditing the pymetrics model-building and four-fifths de-biasing pipeline as it ran in 2020, co-authored with pymetrics staff. Source: https://www.ccs.neu.edu/home/amislove/publications/Pymetrics-FAccT.pdf
- Documentation (captured 2026-11-14): Companion Northeastern audit whitepaper: summer 2020 scope limited to four-fifths compliance; it did not test whether the games measure job-relevant capabilities. Source: https://cbw.sh/pymetrics/pymetrics_audit_result_whitepaper.pdf
- Documentation (captured 2026-11-14): Open-source audit-ai statistical testing library; last commit July 2020. Source: https://github.com/pymetrics/audit-ai
- Documentation (captured 2026-11-14): Current product page, 'pymetrics for Game-Based Assessments': 12+ games, about 25 minutes, trait areas such as effort, risk tolerance, attention and fairness; marketing level. Source: https://harver.com/gamified-assessments/
- Searched, not found (captured 2026-11-14): No current system card, model card or Annex IV-style document for the games. Searched: harver.com/gamified-assessments, the harver.com media library (pymetrics PDFs are 2023 re-uploads of pre-acquisition papers), pymetrics.zendesk.com candidate articles, trust.harver.com/resources (only a pymetrics CAIQ and shared security model, gated).

### Bias Audit Transparency: 62/100 (weight 18%)

- Audit report (captured 2026-11-14): BABL AI NYC LL 144 audit of the pymetrics Soft Skills Platform (v1.0 17 July 2025): data January to December 2024, 283,256 applicants with known gender (female impact ratio 0.992), lowest race impact ratio 0.914 (Asian); one pooled table across uses. Source: https://harver.com/wp-content/uploads/2025/11/pymetrics-Soft-Skills-Platform-2025-Bias-Audit.pdf
- Audit report (captured 2026-11-14): Standard Chartered-hosted copy of the same July 2025 audit, stating an annual BABL AI cadence. Source: https://www.sc.com/en/uploads/sites/66/content/docs/Harver_SSP_Public-Summary-2.pdf
- Public statement (captured 2026-11-14): 'Algorithmic Monocultures in Hiring' (Stanford, Chapman, Northeastern; May 2026, FAccT 2026) on 4.2 million pymetrics applications to 1,746 positions: 14.74% of Asian and 25.87% of Black applicants' applications went to positions showing adverse impact. Source: https://arxiv.org/abs/2605.27371
- Public statement (captured 2026-11-14): Harver's October 2026 rebuttal says the study's 2018 to 2022 dataset mixed test and production data and retired models, and that 2025 and 2026 LL 144 audits were passed. Source: https://harver.com/blog/what-backed-by-science-means-in-hiring/
- Audit report (captured 2026-11-14): 2020 academic audit of the four-fifths implementation (historical). Source: https://www.ccs.neu.edu/home/amislove/publications/Pymetrics-FAccT.pdf
- Searched, not found (captured 2026-11-14): No 2026 pymetrics LL 144 summary and no position-level or per-employer adverse-impact reporting. Searched: harver.com/security, the harver.com media library for audit, bias and pymetrics files, trust.harver.com/resources, web searches for 2026 pymetrics and Harver bias audits.

### FRIA Support: 25/100 (weight 15%)

- Documentation (captured 2026-11-14): Harver Trust Package covering pymetrics, available on request only. Source: https://harver.com/blog/harver-trust-package-security-privacy/
- Searched, not found (captured 2026-11-14): No FRIA template, intended-use data sheet or deployer guidance for the games. Searched: harver.com/resources, harver.com/gamified-assessments, pymetrics.zendesk.com, knowledge.harver.com (no pymetrics articles), trust.harver.com/resources.

### Data Governance Disclosure: 56/100 (weight 15%)

- Documentation (captured 2026-11-14): Describes training on client incumbents' gameplay against a baseline sample, with models retrained until they meet the four-fifths rule (2020 process). Source: https://www.ccs.neu.edu/home/amislove/publications/Pymetrics-FAccT.pdf
- Documentation (captured 2026-11-14): Candidate help (September 2026): existing gameplay data is reused for 330 days and re-assessed against other employers' role profiles. Source: https://pymetrics.zendesk.com/hc/en-us/articles/24084100889489-I-ve-previously-completed-the-games-can-I-play-again
- Public statement (captured 2026-11-14): Claims the games are validated for fairness across gender, ethnicity and socioeconomic status and matched against data from over 100 million assessments; no metrics are published. Source: https://harver.com/gamified-assessments/
- Searched, not found (captured 2026-11-14): No current statement of retention periods or of whether candidate gameplay trains models. Searched: harver.com/privacy-policy (Harver acts as processor for pymetrics), the harver.com media library (2021 pymetrics privacy policy translations only). pymetrics.com/privacy/notice is a client-rendered page and was not read.

### Human Oversight Design: 50/100 (weight 12%)

- Audit report (captured 2026-11-14): Outputs are Do Not Recommend, Recommend and Highly Recommend tiers set by percentile bands that customers can configure. Source: https://harver.com/wp-content/uploads/2025/11/pymetrics-Soft-Skills-Platform-2025-Bias-Audit.pdf
- Documentation (captured 2026-11-14): Group-level decision-support, not decision-making, framing. Source: https://harver.com/why-harver/responsible-ai-in-hiring/
- Searched, not found (captured 2026-11-14): No public pymetrics documentation of overrides, threshold configuration, per-candidate explanations or audit logs. Searched: pymetrics.zendesk.com (21 candidate-facing articles), knowledge.harver.com and support.harver.com (no pymetrics articles), harver.com/gamified-assessments.

### Post-Market Monitoring: 40/100 (weight 12%)

- Public statement (captured 2026-11-14): Group-level 'Report an AI Issue' address (ai-reporting@harver.com), added in mid-2026, beside GlobalSecurity@harver.com. Source: https://harver.com/security/
- Searched, not found (captured 2026-11-14): No pymetrics status page, release notes or model-change record. Probes: status.pymetrics.com does not respond, pymetrics.statuspage.io and status.harver.com are inactive Statuspages; Harver Platform release notes on knowledge.harver.com do not mention pymetrics; pymetrics.zendesk.com has no release-note articles.

### Customer Documentation: 52/100 (weight 8%)

- Documentation (captured 2026-11-14): Current product page with an FAQ on candidate experience and intended uses such as early careers and internal mobility. Source: https://harver.com/gamified-assessments/
- Documentation (captured 2026-11-14): Public candidate help centre (game length, devices, languages, accommodations, the 330-day replay rule). Source: https://pymetrics.zendesk.com/hc/en-us
- Documentation (captured 2026-11-14): Open-source audit-ai library, a 2020 implementation reference. Source: https://github.com/pymetrics/audit-ai
- Searched, not found (captured 2026-11-14): No employer-facing instructions for use, limitations statement or intended-use boundaries for the games. Searched: harver.com/gamified-assessments, pymetrics.zendesk.com, knowledge.harver.com, the harver.com media library. harver.com/products/games-based-assessments now returns 404.

## Company facts

- Website: https://harver.com/gamified-assessments/
- Founded: 2013
- Headquarters: United States (US)
- Pricing tier: Enterprise
- Market side: Employer-side (AEDT)
- Categories: Assessment AI
- Description: Behavioral assessment using browser-based "games" that infer cognitive and emotional traits. Acquired by Harver in 2022 and folded into the Harver Games product line.

These are company details, not findings, and carry no weight in the score.

## Head-to-head comparisons

- [Pymetrics (Harver Games) vs BrightHire](https://www.hireaiscore.com/compare/brighthire-vs-pymetrics): 50 vs 51
- [Pymetrics (Harver Games) vs SHL](https://www.hireaiscore.com/compare/pymetrics-vs-shl): 50 vs 51
- [Pymetrics (Harver Games) vs Xobin](https://www.hireaiscore.com/compare/pymetrics-vs-xobin): 50 vs 50
- [Pymetrics (Harver Games) vs Test Partnership](https://www.hireaiscore.com/compare/pymetrics-vs-test-partnership): 50 vs 50
- [Pymetrics (Harver Games) vs Versant by Pearson](https://www.hireaiscore.com/compare/pymetrics-vs-versant-by-pearson): 50 vs 49
- [Pymetrics (Harver Games) vs Jobma](https://www.hireaiscore.com/compare/jobma-vs-pymetrics): 50 vs 49

## Conflicts of interest

No vendor pays for placement, scoring, or removal. Casework - the consulting firm that operates this directory - provides paid services to some vendors. Any active or recent (within 24 months) commercial relationship is disclosed on the affected vendor profile and the review is reassigned to an independent reviewer.

Casework has no commercial relationship with this vendor.

Full policy: https://www.hireaiscore.com/about

## How to cite

HireAIScore, "Pymetrics (Harver Games)" profile, rubric v1.0, last reviewed 2026-11-14. https://www.hireaiscore.com/vendors/pymetrics

Please link the profile URL rather than this file, quote the score with its rubric version and review date, and include the vendor's reply where one is published.
