Methodology & AI Act Transparency

Allviss turns open-source material into AI-drafted, intelligence-style analysis for a human reader: a bottom line up front, competing hypotheses, indicators, and probability words mapped to fixed numeric bands (NATO/PHIA-style) so "likely" means the same thing in every report. This page says how that is produced and where it falls short, in the spirit of the EU AI Act's transparency obligation (Article 50).

An engineering self-assessment, not a legal compliance certification. Allviss self-classifies as limited-risk under the EU AI Act: it is not designed or licensed for automated decisions about a named individual (hiring, credit, immigration, insurance, law-enforcement scoring). That reading is our own — there is no lawyer-reviewed risk-tier determination on file.

How allviss works

Every analysis runs the four phases of the intelligence cycle in the Norwegian Armed Forces' unclassified Etterretningsdoktrine 2021: Direction (classify the question, plan what to collect), Collection (web search, news and RSS feeds, and our standing knowledge base, in parallel), Analysis (the four steps below), and Dissemination (the same analysis rendered to email, web and PDF).

The analytic steps follow the structured method taught at the Norwegian Intelligence School (Etterretningsskolen), published as Skjelderup, Haugestad, Pedersen & Stivang, Etterretningsanalyse (Fagbokforlaget, 2025). What follows is our own account of how allviss executes it:

Scope the problem. Is the question a secret (a hidden fact), a mystery (an undecided outcome), or a complexity (reflexive actors)? That sets how deep the analysis goes.
What do we know. An explicit knowledge base — findings, gaps, and a Key Assumptions Check where every assumption states what would prove it wrong.
What might happen. Competing, falsifiable hypotheses, never just one story: most-likely, most-dangerous, an alternative, and a null. An Analysis of Competing Hypotheses (ACH) matrix rates each piece of evidence against each hypothesis and favours the one with the fewest inconsistencies — the doctrinal counter to confirmation bias. A devil's-advocate pass can only downgrade an over-confident rating, never inflate one.
Test and monitor. Each hypothesis becomes concrete, time-bound indicators tied to named sources. The discriminating ones are searched for before delivery and carry their status as at publication. Read those as observations of the present, not as forecasts: an indicator that has already fired was not forward-looking, which is a weakness in the indicator rather than support for its hypothesis.

Resolving indicators in-run is a deliberate departure from doctrine, which places them at the end for a human to watch over weeks. A search costs seconds, so the run answers its own question rather than handing the reader homework. It does mean Analysis reaches back into Collection instead of each phase running strictly once. The epistemics are unchanged; only the order of work is.

What checks the output

Output is not trusted because it parsed. A two-tier evaluator checks each step with deterministic rules first, then a model review where those flag something; a failure triggers one retry. A grounding gate fails closed on evidence it cannot verify, so a fabricated citation never silently passes. Corroboration counts independent voices, not rows: references are clustered by domain, wire byline and near-duplicate headline, so ten reprints of one wire story count once — the defence against circular reporting. Every analysis keeps an audit trail of why each source and query was pursued, available on request.

Where it falls short

We would rather state these than paper over them.

The calibration record is thin. Allviss records dated forecasts, resolves them, and scores itself with Brier scores and reliability curves — but most probability bands do not yet hold enough resolved judgments to say what a past "likely" historically meant, and thin samples are suppressed rather than guessed at. The bands make the word consistent; they do not yet make it verified.
The checks are not independent of the analyst. The analysis is produced by a single model, and the checking layers are not currently independent of it — same vendor throughout, and the grounding gate runs on the analyst's own model. They catch slips and unsupported claims reliably; they cannot catch a blind spot the models hold in common.
A grounding pass is narrower than it looks. The gate grades a size-capped extract of the collection, so it defends against fabricated citations without certifying that every real one was located.
Human oversight is a commitment, not a gate. It is a product-design choice rather than a hard technical control — nothing stops a reader acting on a report without pausing at the disclosure.

How to use it

Allviss is a prioritization and flagging aid for a human analyst — never a substitute for one, and never an automated determination of fact. Every channel (PDF, web report, chat, email) carries an explicit AI-generated disclosure, built into the drafting step rather than added afterwards.

Data protection

Users can export or delete their data at any time (GDPR Articles 15 and 17). Retention and encryption-at-rest are enforced, not just documented, and questions and report content are never exposed outside the account that owns them — including in our own tooling and logs.