Trust, plainly stated.

What happens in the first 72 hours: who declares it, how it gets contained, how the risk is judged, who gets told and when. The Latvian lead authority is named and both notification letters are already written.

The provider's own technical documentation. What ARK is, why it lands as limited-risk against Article 5 and Annex III, how transparency and human oversight work, and how the system is watched after release.

AI literacy is the product, so the evidence is the product. Every measure is listed with the table or file it can be read from, for learners, for organisations deploying ARK, and for the operator.

Scope, roles, access control, change management, logging, backup, vendors and the risk register, written against the Trust Services Criteria and mapped to the controls that already run.

Three severities with acknowledge and contain times, the seven-step response, the post-incident rule that every failure has to produce a migration, an eval or an alert, and the yearly test.

The AI management system policy: objectives with a measure each, the accountable person, the operational rules that constrain every model call, and how performance gets read quarterly.

All fourteen AI systems ARK runs, described by the capability they ask for rather than a model name, each with its input, its output, who oversees it and its deterministic fallback.

A table of what reaches a model and what never does, the identity strip, the private lane, retention, where processing happens, and why ingested material is data and never instruction.

Seven ways this could hurt somebody, from a level read wrongly to an aggregate used against a named person, each with its controls, the risk that stays, and a verdict with conditions attached.

Where every SOC 2 and ISO 42001 control's evidence lives, criterion by criterion, so a CPA firm or a certification body can scope without a discovery call. It also lists plainly what is not there.

A FORESIGHT run reads an occupation, a level band and a headcount. It never reads a person, a record, a capture or anything anybody wrote. The ability to look at one employee is absent from the path rather than switched off in a setting.

Any cell holding fewer than 3 people is withheld from every read, and the report prints how many cells were withheld. Small cells are dropped rather than blurred, because a blurred cell is still a cell about people.

Every figure is a projection over published occupational task data plus a headcount, moved by an epoch offset and by assumptions that are printed beside the number. Measured results live on the proof surfaces and carry the opposite label. The two never share a page.

A scenario is planning input. It supports a decision about a program, a sequence or a budget. It does not support a finding about a person, and it is not offered as one.

FORESIGHT is organizational planning analytics on modelled scenarios. It is never to be used for decisions about an individual employee or candidate, and no individual-level output exists.

FORESIGHT scenarios are shown as a labelled preview; sold results are gated on at least 5 closed calibration back-tests. Until that ledger fills, what you see is labelled as a preview wherever it appears.

We write evidence sets by hand, then render each one several ways: formal and casual phrasing, and the same content in English, Latvian and German. The level engine, the AI-exposure read and the voice gate run over every version. If a placement moves because somebody writes casually or writes in Latvian, that is a bug and the run fails on it. The tolerance is stated: a rung may not move at all, a share may wobble by one point.

Across the set we report outcomes by occupation group, level band and locale, and check that identical evidence lands on the identical rung whichever occupation group it sits in. AI-exposure shares do differ between occupations, because the published task data differs. That is data, not bias, so only the rung is held to equality.

ARK holds no data on race, gender, age, health or belief, and the fairness harness does not infer any of it. The groups it reports on are occupation group, level band and language. Nothing in the harness touches a member's rows: every piece of evidence in it is invented for the test.

One variant plants name-shaped tokens into the evidence on purpose, so the run can prove the identity strip removed them before any engine read the text. A single surviving token fails the run.

It is a synthetic internal check that we run and read ourselves. It is not a bias audit, it is not certified, and no outside party has signed it. The independent review stays owner-side and we will publish it here when it exists, with whoever ran it named. Until then, treat this note as our own testing, openly described.