They’re going to sell you a score, and they’re not going to explain where it comes from.
That is, almost without exception, the experience of buying a psychometric assessment today. They hand you a PDF with a number, a radar chart, and three bullets of “recommendations.” Ask to understand the formula and you get marketing-speak: “our proprietary algorithm is validated with over X million data points.” Ask to adjust the weights for your specific role and they answer that “the model is scientifically calibrated,” which is the elegant way of saying you can’t.
This isn’t a technical detail. It’s bad scientific practice and, increasingly, a legal problem.
Regulation has already arrived
Two regulatory pieces every HR Tech buyer should know in 2026:
EU AI Act (Regulation 2024/1689)
In force since August 2024, with a staggered application schedule. It explicitly classifies AI systems used for personnel selection and candidate evaluation as high-risk systems (Annex III). Obligations for high-risk AI systems include:
- Detailed technical documentation of the system (Art. 11).
- Transparency and provision of information to users (Art. 13).
- Meaningful human oversight (Art. 14).
- Logging mechanisms for traceability (Art. 12).
- Risk management throughout the lifecycle (Art. 9).
Fines: up to 35 million euros or 7% of global annual turnover.
NYC Local Law 144 (Automated Employment Decision Tools)
In force since July 2023. Any company using AEDTs (Automated Employment Decision Tools) on candidates residing in NYC must:
- Perform an annual independent bias audit.
- Publish the audit results.
- Notify candidates that an AEDT is being used.
This is not a theoretical debate. It’s law.
Other jurisdictions are moving in the same direction: Illinois (Artificial Intelligence Video Interview Act, 2020), Colorado SB 21-169 (insurance, 2021, but set a regulatory precedent), California (multiple bills in 2024-2025).
Practical question: if your assessment vendor can’t deliver the technical documentation required by Art. 11 of the EU AI Act or an annual bias audit, how do you comply?
What “auditable” means concretely
Auditable doesn’t mean open source. It means you, as the client company, can inspect, adjust and record what the system does. Five minimum capabilities:
| Capability | Question it answers | Why it matters |
|---|---|---|
| Visible weights | How much does each dimension weigh in the final score? | Without this, you can’t justify a decision |
| Adjustable weights | Can I change the weights for my specific role? | A sales role isn’t evaluated with the same weights as a technical role |
| Transparent formula | How are inputs combined into the output? | Required to comply with EU AI Act Art. 13 |
| Versioning | When did the algorithm change and what changed? | Bias audits require history, not a snapshot |
| Decision logging | Why did this candidate get this score? | Traceability (Art. 12) and defense against claims |
If your vendor doesn’t check all five, it’s not auditable. It’s a black box with good UI.
What this looks like in Talen.to
I’ll be concrete about what we have today, not what we plan to have:
Adjustable OCEAN weights per company
In /settings/psychometric/ocean each company defines how much each dimension (Openness, Conscientiousness, Extraversion, Agreeableness, Emotional Stability, Structure & Rhythm) weighs for its roles. We don’t impose a generic default — we give you one and you edit it.
Configurable and versioned values
Each company defines its own values with weights and MIN_FIT per value (typically 75-85%). Values have versioning: if your company redefines “perseverance” in 2027, there’s a record and old assessments remain bound to the definition in effect at the time.
Documented and versioned scoring formula
FitCalculator combines oceanFit and valoresFit into overallFit, with a ScoringVersion field (v1 with equal weights, v2 with weighted). When you upgrade from v1 to v2 there’s a record; old assessments are not retroactively recomputed without your explicit action.
Editable AI prompts
The prompts that generate AI reports live in /companies/[id]/settings/psychometric/prompts. If you don’t like how the AI writes the narrative, you edit it. If your internal compliance requires specific language (e.g., “avoid references to age, gender, origin”), you add it to the prompt and it gets versioned.
Calibration with your own data
Endpoints /calibration/assessments and /calibration/recalculate-all let you correlate real performance with scores and adjust the engine. After 12 months your engine is calibrated against your company, with your data and your definition of success.
Auditable cultural factors
The scoring engine adjusts OCEAN by country. The factors are documented, editable, and adjustments are logged. It’s not “the algorithm tunes itself” — it’s an explicit, configurable, auditable adjustment.
The typical objection: “but then anyone can manipulate it”
I hear it often: “if the weights are adjustable, companies will tune the system to favor their picks.”
Two responses:
1. Auditability is precisely the defense against that. If weights are hidden, you can’t detect manipulation. If they’re versioned and logged, you can. Transparency doesn’t enable bias — it makes it visible.
2. The alternative is worse. A closed system hides the vendor’s bias instead of the client’s bias. There’s a long list of papers documenting systematic biases in proprietary assessments (see Raghavan et al., 2020, “Mitigating Bias in Algorithmic Hiring,” FAT* Conference). The black box is not neutral — it’s opaque.
The product objection: “but then it’s too complicated”
I hear that one too: “our clients don’t want to tune weights, they want it to just work.”
True. That’s why Talen.to has reasonable defaults for each combination of industry + size + role type. Most companies never touch a weight and the system works well. But when a mature company wants to adjust, they can. And when an audit shows up, there’s something to show.
That’s the difference between “optional configurability” and “impossible configurability.” Talen.to is the first. Most of the market is the second.
What this implies for your buying process
If you’re evaluating an assessment or a hiring platform in 2026, add these questions to your RFP:
- Can I see and adjust the scoring weights per role/company?
- Are AI prompts editable and versioned?
- Is there logging of individual decisions (why each candidate got each score)?
- Can you deliver a bias audit (e.g., NYC AEDT format)?
- Do you have the technical documentation required by EU AI Act Art. 11?
- Can I recalibrate the engine with my own performance data?
If three or more answers are “no” or “on the roadmap,” it’s a black box. Consider the regulatory and operational risk before signing.
Closing
Algorithmic transparency is not a 2030 nice-to-have. It’s a 2024-2025 legal obligation and basic scientific practice since forever. If your scoring engine isn’t auditable, you’re not doing science-based hiring — you’re doing astrology with radar charts.
If you want to see an auditable engine in practice, book a demo or create an account at app.talen.to/sign-up and go straight to settings/psychometric. It’s all there, without marketing.
And if you want the full panorama of what the platform does, What Talen.to actually does. On why generic assessments fail: Why generic assessments fail at culture fit.
— Fran
Related OCEAN+ profiles
Discover which personality dimensions to look for in each role.
Related Articles
Talen.to's 10 behavioral archetypes (and why we calibrated them with our own data)
These aren't archetypes borrowed from a management book. They're centroids computed in a 7-dimensional space from real field data. How the model works, what it predicts, and why in-house calibration changes everything.
Multi-evaluator competency assessment: when the psychometric test isn't enough
A psychometric assessment tells you who the person is. A multi-evaluator competency assessment tells you how good they are at what the role actually demands. When you need one, when you need the other, and how they combine.
Why generic assessments fail at culture fit (and what to do about it)
Traditional assessments evaluate the candidate against a general population. Talen.to evaluates them against your role, your company, and your cultural context. Extended OCEAN, cultural factors, 10 archetypes calibrated with real data.