What we do not build, and why.
This category is full of products that ship features which are already unlawful in the European Union, or that no study supports. The list below was written before the first line of code.
- 01
No automated rejection
Not as an API route, not as a default value, not as a recommendation pre-filled in the interface.
Evidence
Article 22 of the GDPR. The lesson of SyRI in the Netherlands and Robodebt in Australia, where automated systems produced decisions nobody could explain afterwards.
- 02
No score, no risk level, no probability
There is no score column, no risk variable, and no probability anywhere in the product, including in internal data.
Evidence
A probability cannot be shown to a candidate who asks why, defended to a regulator, or audited. A fact with a source can be all three.
- 03
No emotion, stress or deception detection
Not in interviews, not in recordings, not as an add on.
Evidence
Prohibited under Article 5(1)(f) of the AI Act in the employment context.
- 04
No AI text detection
Candora never infers that a CV or a cover letter was written by a model, and writing style is never a signal.
Evidence
Independent testing found a high false positive rate on writers whose first language is not English. It systematically punishes the wrong candidates.
- 05
No professional network scraping
Only data the candidate or the recruiter provides, plus public registries and APIs whose terms permit the use.
Evidence
There is no defensible legal basis for scraping a professional network to inform a hiring decision in the European Union.
- 06
No voice or live interview analysis
Out of scope permanently.
Evidence
Decided as a permanent product boundary in July 2026.
- 07
No biometric template
The test is binary.
Evidence
For candidates there is no usable exemption. Consent fails on the imbalance between the parties, and Article 9(2)(g) requires a specific law that does not exist for private recruitment. The Spanish regulator confirmed this in June 2025, including that the AI Act provides no legal cover and that templates deleted within seconds are not a defence.
- 08
No software on the candidate's device
No reading of processes, windows, clipboard or screen.
Evidence
The French regulator has fined continuous webcam monitoring and screenshotting. In Germany, works council codetermination under section 87(1)(6) BetrVG attaches to the mere capability to monitor, regardless of intent. A second device defeats the whole approach in any case.
- 09
No gaze, behaviour or keystroke dynamics
None of it is read, stored or used as a signal.
Evidence
No construct validity study supports it. One proctoring tool flagged none of six instructed cheaters. Another flagged candidates with darker skin five times more often at zero difference in actual cheating.
- 10
No pre-crime prediction and no risk profiles
Candora never marks a candidate as risky in advance.
Evidence
The network layer is built so that it can never become this: typed links, profiles never merged, and a confirmed link only through human confirmation enforced by a database constraint.
- 11
No enrichment from third party data brokers
Candora does not buy or pull candidate data from enrichment vendors.
Evidence
Article 14 of the GDPR requires proactively telling the person what was collected and from where. The French regulator has fined a lead enrichment vendor partly for exactly this. A legitimate interest argument that survives for business email does not survive for an employment decision.
- 12
No special categories under Article 9
Health, religion, political opinions, union membership, sexuality and the rest are never extracted, stored or displayed, even when the CV contains them.
Evidence
Article 9 of the GDPR. Enforced by a test against a CV fixture that contains such data.
Specific signals we considered and do not use.
These are not the general principles above but specific entries from the current version of the product ruleset, generated the same way as the Signals page. Each one carries the date of the decision and a link to its source.
- 01
Estimates whether a human or a language model wrote the text in a CV or cover letter.
This detection misfires on writers whose first language is not English, exactly the candidates it would disadvantage the most.
Source
CLAUDE.md, cervena cara 4 · 2026-08-04
- 02
Tracks where a candidate looks on screen during an interview or test.
No study has shown it actually relates to cheating.
Source
Proctorio/Respondus studie, CLAUDE.md cervena cara 12 · 2026-08-04
- 03
Analyses typing rhythm and speed to infer fraud.
It sits in the same category as gaze tracking: an independent study found five times more flags for darker skin tones with zero difference in actual cheating.
Source
Respondus studie, CLAUDE.md cervena cara 12 · 2026-08-04
- 04
Infers emotion, stress or intent to deceive from voice, face or behaviour.
The AI Act lists it as a prohibited practice under Article 5(1)(f).
Source
AI Act cl. 5(1)(f), CLAUDE.md cervena cara 3 · 2026-08-04
- 05
Builds a biometric template from a face, voice or fingerprint and compares it against a reference.
It falls under GDPR Article 9, where no usable exemption exists for hiring because a candidate's consent fails on the power imbalance.
Source
AEPD 3. 6. 2025, CLAUDE.md cervena cara 10 · 2026-08-04
- 06
Buys or downloads additional data about a candidate from data brokers.
GDPR Article 14 requires proactively disclosing what was collected and from where, and a French regulator fined the vendor KASPR for exactly this.
Source
CNIL pokuta KASPR, GDPR cl. 14, CLAUDE.md cervena cara 14 · 2026-08-04
- 07
Compares the date a company was dissolved with the date the candidate claims they left it.
The public ARES API never returns a deregistration date, verified against real companies, so the rule could never have produced a finding against real data.
Source
ADR 0012, docs/OPEN-QUESTIONS.md Q024 · 2026-08-04
- 08
Verifies the reference as a person, not just the company and domain they speak for.
Profiling the reference as a person would violate GDPR Article 14 and the no-risk-profiles rule.
Source
TECH-VIZE-2026-08 sekce 2.3, cervene cary 13 a 14 · 2026-08-04
- 09
Has Candora contact the reference directly instead of the recruiter.
Talking to the reference is the recruiter's job; the system only documents whether the reference's company and domain hold together.
Source
TECH-VIZE-2026-08 sekce 2.3 · 2026-08-04
- 10
Retrieves the candidate's insurance record on their behalf, for example with a power of attorney or by signing into their account.
The record is obtained solely through an active step by the candidate, with no power of attorney and no signing in on their behalf.
Source
TECH-VIZE-2026-08 sekce 3.2, ADR draft D · 2026-08-05
- 11
Shows or stores nationality, ethnicity or country of origin, even when the right to work document carries them.
Those are discrimination grounds. The product reads the document only to state the date it expires, and the fields for anything else do not exist in the code, so there is nowhere to put them.
Source
T141, cervena cara 9 · 2026-08-11
- 12
Compares the photo in an identity document with a selfie of the candidate.
Banks may do this because anti money laundering law tells them to. An employer hiring has no such law behind them, so the exemption does not reach recruitment. In 2024 the ICO ordered one company to stop and delete the biometrics of two thousand employees for exactly this reason. Structural checks on the document, meaning checksums and expiry, are a different thing and Candora does those.
Source
T151, cervena cara 10 · 2026-08-11
If any of this changes, it will be visible.
Each limit above is covered by a test that fails when the limit is crossed.