Privacy Policy update — the AI practical competency assessment

What we collect: the AI practical competency assessment is offered to accounts we allow to sit it, and we store the following in the account that sits it: (1) the work document (problem framing, initial judgement and check plan, deliverable, the reasons for your judgement and any revision, expected side effects and limits, validation plan, and which AI suggestions you adopted or set aside and why) and your code; (2) the requests and results of the check tools (viewing materials, calculations and code runs), the expectation you write before each check, and your interpretation of and decision on the result; (3) the requests and replies in the AI conversation inside the screen (the AI assistant), with the materials and draft sent with each request; (4) the drafts fixed at each stage and the final submission; (5) the judgement and reason you give in the AI proposal review; (6) your confirmation interview answers; (7) the writing history kept at each save, the times each stage started, was saved, submitted or ran out of time, request-processing records, the time and version of your consent, and your account identifier; (8) the analysis made after submission (the proposed score for each item and whether it was confirmed, quoted evidence, the reasons and narrative written by the AI, and the results against reference cases) and the AI usage and cost of the sitting. We do not collect input-origin signals such as paste or away-from-screen events, or video or audio.

Why we use it: to run the assessment (the AI assistant's replies and the check tools), to score it and write the result report, and to tell you the result and answer your questions. The AI usage of each sitting is also used for cost management and to check the monthly usage limit. We do not use these records for advertising or marketing, and we do not give them to AI providers for training their models.

How scores are set (automated decisions): no person enters a score in this assessment; scores are set in this order. (1) When you submit the confirmation interview or the time runs out, a scheduled job on our server starts the automatic analysis. We analyse sittings started after agreeing on the consent screen that states this (the October 8, 2026 version), and trial sittings started by our staff in their own accounts. If the analysis fails, the record is kept as it is and the analysis is retried; after a set number of attempts it stops and the assessment staff are notified. (2) For a coding task, your submitted code is run against non-public reference cases in an isolated environment on our server, and the share of cases it passes sets the maximum score for the deliverable item. For a calculation task, the values written in the deliverable are compared with the reference values. (3) The AI (Anthropic) proposes a score for each item in steps of 0.5 within that item's points, following the items and behavioural criteria set in the task, and quotes its evidence from records you wrote yourself. (4) Our server checks, for each item, that the quotes match the actual records, that the kind of evidence the item requires is there, and that the score stays within the points and the maximum. Only items that pass are confirmed; an item that fails stays undetermined, an item with no usable records stays unobserved, and if any item is undetermined or unobserved the total is withheld. If there are no usable records at all, the AI is not called and every item is unobserved. The assessment staff do not enter or change scores; they review the records and the analysis, release the result, and may run the analysis again when needed. We do not use these results to make decisions that affect your rights or obligations, such as whether you are hired.

How long we keep it: the record of a sitting, including the analysis and AI usage, is destroyed 90 days after you submitted the confirmation interview (or, if you did not, after the time ran out). Staff opening, analysing or releasing it does not extend this period. From the end of that period the record no longer opens on your screen or the staff screen, and a daily clean-up job deletes it. If you delete your account, the record is deleted with it even before then, and you can ask for earlier deletion as described in Section 7. For a sitting deleted because you closed your account, we keep only the sitting identifier and the deletion time as a deletion marker so the same record cannot be created again, and the AI usage of a deleted record is kept only as monthly token and cost totals, without account or sitting identifiers.

Processors and international transfers: · Anthropic, PBC (United States): generating the AI assistant's replies during the sitting, and analysing the records after submission and proposing a score for each item. When you send a request to the AI assistant, we send the request, the materials and current draft you chose to include (not the draft if you chose to leave it out), the check-tool records and interpretations up to that point, and the last 8 messages of the conversation. For the automatic analysis, we send the work document, code and fixed drafts up to your final submission, the check-tool records and interpretations, the requests and replies of the AI conversation, the proposal review, the confirmation interview answers and the results against reference cases. The records sent carry random strings such as your account identifier and the sitting identifier, and the time of each record; your name and email are not sent. What you write is sent as written, so we suggest leaving names and contact details out of your work and answers. Anthropic normally deletes API inputs and outputs within 30 days of receipt or generation. Content flagged for usage-policy violations may be retained for up to two years, and related safety classification scores for up to seven years; legally required retention also applies. · Turso (a United States company; servers in the Tokyo region of Japan): storing the record of the sitting, for the retention period stated above. The code-run tool runs in an isolated environment on our server with no outside connection, and its run records are sent with AI assistant requests and the automatic analysis as described above. Records of this assessment are not sent to OpenAI. Transfers happen over API calls: to Anthropic when you send a request to the AI assistant and when the automatic analysis runs, and to Turso when a record is saved; the recipients' contact details and how to decline are in Section 5. If you do not agree to the international transfer, you cannot start this assessment.

Who sees the results: only you and our assessment staff (operator accounts) see the record and the analysis. You can see your progress records and a submission summary in your account, and after the staff review the analysis and release it you can see each item's score and status (confirmed, undetermined or unobserved), the total or the fact that it was withheld, and the number of evidence records for each item. The AI's reasons and narrative, the quoted evidence, the reference cases and the reviewers' reference material appear only on the staff screen. This assessment has no feature that sends results to institutions, employers or anyone else, or that creates viewing links. We decide which accounts may sit it: the sign-in email of each allowed account, and an affiliation name if the staff added one, are kept in the allow list in our operating settings until the staff remove them from it.

Access, correction, deletion and explanation: write to help@preterview.com from the email address of the account that sat the assessment and we handle it as described in Section 7. If you ask for an explanation or a review of a result, the assessment staff check the record and the server's check results and answer you, and run the analysis again when needed. If an AI assistant reply during the sitting was inappropriate, tell us at the same address. If you do not want automatic analysis, you can choose not to sit the assessment; it does not start unless you agree.