Part Five Assessment: Model Answer and Rubric

Seventeen rows, and they do not all end in an answer. Seven systems are high risk through the list of uses. Two more arrive through product legislation and neither is settled. Six are not high risk. One row asks whether the thing is an AI system at all, and one is an entity with no systems recorded against it.

Not high risk. Every row here is settled except the last.

Classification and reach are two questions and the table answers only the first. On this evidence four systems have a European Union nexus.

Arbiter, because the Frankfurt pilot has run since 2024 and the German entity uses the output on applicants there.

Screen and Interview, because they are placed on the Union market in Ireland, Germany and the Netherlands.

Pulse, because it is configured for all employees of all nine entities, which takes in the 95 people in Frankfurt and the works council agreement concluded there in November 2024.

Wellcheck is a fifth if its Irish clients and the client workforces in Spain and Poland put its output in the Union, and nobody records where client employees are.

Everything else is a United Kingdom or United Arab Emirates deployment with its output landing in the same place, and saying so is part of the answer rather than a way of avoiding it.

It is out, and that is not a reason to relax. Handlers accept the fast-track recommendation about ninety four times in a hundred, and the Claims Operations Director has put in writing that she would not call that a human reviewing the decision.

Pricewise and Pricewise Life run on the same platform, are owned by the same person and were built by the same team.

Point 5(c) has two conditions and both must hold: The activity must be risk assessment or pricing, and the line of business must be life or health insurance.

Pricewise has the activity and the wrong line of business. It prices motor and home, and nothing on the list reaches it. Pricewise Life has both, so it is in.

Do not reach for the filter that takes a listed system back out. It shuts where the system profiles natural persons, and individual underwriting from a health questionnaire is profiling.

Over-classifying Pricewise is not the safe option. It buries the one that matters.

Screen and Interview are built by Talentflow Ltd and sold to other businesses, which makes Meridian their provider rather than their deployer. The two lists are not versions of each other.

The deployer list is about using a system properly once it has arrived. The provider list is about getting it lawfully to market and keeping it lawful there: Risk management, data governance, technical documentation, logging, instructions for use, human oversight, accuracy and robustness, a quality management system, the conformity assessment, the declaration, the marking, registration and post market monitoring.

Sortline does not reach the Act through the list of uses. It reaches it, if it reaches it, through the product route, and that route has two limbs. The AI system must be a safety component of a product covered by the listed product legislation, or such a product itself, and that product must undergo third party conformity assessment.

The machinery legislation is on the list. Neither limb is established here. The safety impact assessment on change record CHG0029877 states that adaptive pick optimisation does not touch the safety functions, which if right removes the first limb. Kessler's sales email of 14 April 2025 says the marking covers the software and nothing further is needed.

That is the claim to test, not to accept.

Proctor360 sits on point 3 for monitoring behaviour during tests, and that is the easy part. Flag F8 returns a per session attention and composure indicator derived from head position, facial movement, blink rate and voice characteristics, and flags the outer decile.

Inferring the emotions of a person in the workplace and in education institutions is prohibited, with an exception only for medical or safety reasons. If F8 infers an emotional state, no amount of classification work helps, because the answer is stop rather than comply.

Sekura's instructions for use give the intended purpose as one to one verification at a door, and verification of that kind is excluded from the biometrics point. The Dubai commissioning record shows one to many search enabled against the site template set, and the Leeds installation sheet records the configuration as copied from the Dubai site profile.

Interview is the more troubling system. It scores speech rate and pause distribution, eight of its twelve customers reject candidates below a score automatically, Talentflow has never told them so, and the published model card claims two hundred thousand interviews where the datasheet records four thousand one hundred and twelve, from three customers in one sector in one country.

But exposure is role, reach and population, not how uncomfortable a system makes you.

One document, drawn up before the system reaches the market, covering the design, the data, the metrics, the foreseeable risks, the oversight measures and the monitoring plan.

Screen has no model card at all. What exists is a one page FAQ maintained by Sales and a feature file in a repository.

The training, validation and test sets have to be examined for bias and the examination written down. The December 2025 run inferred sex from forename lists, reports ethnicity as not measurable, and carries no confidence interval and no breakdown by customer, sector or role family.

The minute reads "within four fifths, no action". Nobody has looked at which of the two hundred and fourteen remaining features move with the characteristics dropped at ingestion, and commute_distance_km is still in the set with an action to look at it and no owner.

The customer pack is what a customer configures the system from. Issue 7 says testing found no adverse impact on any protected group, with every selection ratio within the four fifths rule.

Point 4 goes to the assessment based on internal control, so no notified body is involved and this is work Talentflow can do itself.

None of it has started. Underneath sits a question nobody has asked: Talentflow Ltd is incorporated in England and Wales and operates from Manchester and Dublin, and whether the Dublin site is an establishment in the Union decides whether an authorised representative is needed.

Screen retrains quarterly through a manual promotion gate, which is more discipline than Interview has.

A model at 246 words.

Every item is a document reference, a date and a consequence, which is the register Marcus Whitfield acts in. It never uses the word responsible.

It puts the item he owns first, and asks for one decision rather than a mandate, because a decision is something a General Counsel can give you in an hour.

It explains the Act. It argues that Meridian should act because it is the right thing to do. It asks him to stop Talentflow selling, which is not his to stop and not yours to ask.

Score yourself on each criterion. Be strict.