AI Model Card
Version 1.0 · Effective July 13, 2026
What our scoring models do, what they don't do, and how we test them. Published in the spirit of EU AI Act Article 13 (transparency for high-risk AI systems) and NIST AI RMF 1.0.
Available in 8 languages via the language switcher (top-right). The English version is authoritative.
1. What Kyrovo's AI does
Kyrovo produces two personal scores from user self-reported inputs and public reference data:
- •Career Resilience Score (CRS) — a 300-850 composite score of the user's career stability. Purely informational.
- •AI Risk Score — a task-level Low/Medium/High rating of how exposed a user's weekly tasks are to current generative-model capability.
Kyrovo also uses LLMs (Lovable AI Gateway; foundation models from OpenAI, Anthropic, and Google) to draft plain-language explanations of these scores. No score itself is generated by an LLM — scores are deterministic computations on user inputs.
2. CRS inputs and weights
CRS = weighted sum of four sub-scores, mapped to a 300-850 scale:
| Sub-score | Weight | Source |
|---|---|---|
| Energy Pulse | 25% | 3 self-report questions |
| AI Resilience | 30% | Task-level self-report scored against capability benchmarks |
| Financial Runway | 30% | Savings ÷ monthly burn, vs. 6-month target |
| Skill Diversification | 15% | Breadth of task domains within user's industry |
3. Training data & reference sources
- •Task capability benchmarks: curated internally from public model evals (MMLU, HumanEval, MMMU, SWE-bench, GAIA) as of Q2 2026. Refreshed quarterly.
- •Rehire time benchmarks: BLS JOLTS, Indeed Hiring Lab, LinkedIn Workforce Report, Revelio Labs (all public).
- •Layoff signals: US DOL WARN Act filings, SEC 8-K item 2.05 filings, Challenger Gray monthly reports (all public).
- •LLM text generation: foundation models via Lovable AI Gateway. Kyrovo does not train foundation models on user data.
4. Known limitations
- •Not a prediction. The CRS is a snapshot of self-reported inputs, not a probability of layoff or burnout. It should never be used by an employer, insurer, or lender.
- •US-anchored benchmarks. Rehire time and severance data are strongest for US, UK, EU-5, CA, AU. Coverage for other jurisdictions is sparser.
- •Task capability drift. Generative models improve monthly. AI Risk scores can shift ±15 points within a quarter as we refresh benchmarks. This is documented in the score changelog.
- •Self-report bias. All four sub-scores depend on honest inputs. We surface but do not correct for over- or under-reporting.
- •Not medical or financial advice. Burnout is a clinical condition. CRS is not a diagnosis. Runway is not a financial plan.
5. Bias testing
We test CRS distributions across self-reported demographic segments (age band, gender, region) on a rolling 90-day cohort and publish drift >2σ in the score changelog. When drift is detected we investigate whether an input weight is over-representing a proxy variable. As of the current version:
- •No adverse impact detected across age bands.
- •Regional variance is within expected range from underlying rehire-time data.
- •Financial Runway sub-score reflects real income inequality — we do not adjust for it, but we surface the raw ratio so users can act on it.
6. Your rights
- •Right to explanation. Every score comes with a plain-language breakdown of which inputs drove it.
- •Right to opt out. You can disable AI-generated explanations in Settings → Privacy while keeping the deterministic scores.
- •Right to erasure. Delete your account from Settings → Account. All scores and inputs are purged within 30 days.
- •Human review. Contact privacy@getkyrovo.com to request a human review of any automated output about you.
7. Governance
Model changes are logged and versioned. Material changes to weights, inputs, or capability benchmarks trigger a version bump on this page and an in-app notice on next login. See also our DPIA, Privacy Policy, and Security pages.
Questions about this model card? Contact privacy@getkyrovo.com. Under the EU AI Act, our EU representative can be reached at the same address.