Glossary

How to use this glossary

Terms are listed alphabetically. Terms defined by a volume of the series are used here as that volume defines them; terms of this dossier (such as established, illustrative, ledger and open thread) are defined by Chapter 1 and by DISPOSITION.md.

Terms

The terms below are those the dossier uses in a specific sense, with the sense it gives them. Chapter numbers point to where a term does most of its work. Identifier families are described by their prefix; the identifiers themselves are in Appendix A.

Across the usage modes. A subsection, normally a compact table, that shows how a chapter's subject differs between Mode 0 and Mode 5.

AI Decision Record (AIDR). The series' numbered record of a decision about an AI system: what was decided, by whom, under which conditions and until when. Each volume has its own range; Hansa Mutual's AIDR numbers are used from AIDR-0021 onward (Chapter 22).

Allocation function. The claims-triage agent's task of assigning claims to named handlers. It is the function reclassified as high-risk under the AI Act from AIDR-0651 (Chapters 5 and 13).

Architecture Decision Record (ADR). A short record of an architectural decision and its reasons, used in the series alongside the AIDR.

Autonomy level. The degree of delegation of an action class: Level 0 Advisory, Level 1 Assisted, Level 2 Supervised, Level 3 Bounded and Level 4 Autonomous. Hansa Mutual's ceiling is Level 3 for reversible actions (Chapter 5).

Baseline and overlay. The group's way of handling seven legal regimes: one baseline of controls, with market overlays that only tighten it (Chapter 4).

Benefit record (BEN-*). The record of the benefit of one use, measured against a counterfactual and carrying its own evidence age; its figure matches the footprint record. BEN-HM-0001 (claims assistant) and BEN-HM-0002 (triage agent) are Hansa Mutual's (Chapter 19A) (14, Nordqvist, 2026).

Builder. In Vibe to Dev, a person who is not a developer and builds software with an AI assistant. Hansa Mutual's builder works in the Polish service centre (Chapter 19).

Cap. The third test of the net-benefit principle of AI Environment: a use runs under an environmental budget with a stop path outside the system it stops (Chapter 19A) (14, Nordqvist, 2026).

Claim record (CLM-*). The record that every public statement about the environmental effect of AI must have before it is published, with a stated use, pool, boundary and expiry; approved by the group AI governance board. CLM-HM-0001 held the board's sentence that AI work "became more efficient" (AIDR-1361) (Chapter 19A) (14, Nordqvist, 2026).

Claims-triage agent. The Mode 5 agent that sorts and proposes on claims, bound to the market of each claim (UC-2026-014, AIDR-0024). Its predecessor was a Mode 4 assistant (Chapter 5).

Compare. The second test of the net-benefit principle: the benefit is measured against the non-AI alternative and against rebound (Chapter 19A) (14, Nordqvist, 2026).

Contour. The visual notation of Landscape: one mark per AI use, with seven dimensions (Chapter 20).

Count. The first test of the net-benefit principle: the footprint is counted at a stated boundary, in a stated year, with a stated method, and no figure appears without them (Chapter 19A) (14, Nordqvist, 2026).

Credibility check. The comparison of Part V of the fiction with real organisations of the same kind, giving each of sixteen dimensions a verdict of credible, generous or thin (Chapters 25 and 26).

Credible, generous, thin. The three verdicts of the credibility check. Credible: nothing found contradicts the series. Generous: the series goes further than the evidence found. Thin: the series gives a figure that the evidence neither supports nor contradicts (Chapter 26).

Decision seat. The place on a mark in Landscape that shows who holds the decision right, by tier T1 to T4 (Chapter 20).

Decision tier. One of four levels of decision structure at Hansa Mutual, numbered 4 to 1, from the boards above the group AI governance board down to the design authority and the pipeline gates (Chapter 3).

Delegation profile (DLG-*). The record of the degree of delegation per action class for a system or a body, with its ceilings and its clock (Chapter 22).

Design case. A feature of the fiction that is more integrated than real examples, presented as a worked case and not as the typical organisation (Chapter 26).

DORA. Regulation (EU) 2022/2554, the Digital Operational Resilience Act. Not to be confused with the DevOps Research and Assessment programme (Chapter 4).

Dossier. A companion that gathers everything the series establishes about one organisation. This book and Nordvik Region are the two dossiers.

Energy and carbon budget (EBG-*). The environmental budget of a use, set by someone other than the use's owner, with limits in kWh, calls or accelerator hours, a stop path outside the system and a drill; a budget without a drill is proposed and is not a limit. EBG-HM-0001 to -0006 are Hansa Mutual's (Chapter 19A) (14, Nordqvist, 2026).

Established. A fact that the volumes state as part of the organisation's profile, as of 16 November 2026 (Chapter 1).

Evidence age. How long ago the test, drill, sample or review behind a claim was made, and when it expires. Each volume gives its own clock (Chapter 23).

Evidence clock. The period after which a piece of evidence is treated as expired; 90, 180 and 365 days occur in different volumes (Chapter 23).

Exception (EXC-*). A dated, approved departure from the baseline, such as EXC-0004 for a model provider outside the EU (Chapter 6).

Exit plan (EXIT-*). A plan, with rehearsals, for leaving a supplier or model; EXIT-004 is the main one (Chapter 6).

Extract. A working note in data/extracts/, one per volume, from which the chapters are written.

Fiction. Everything the dossier says about Hansa Mutual is invented; real organisations are named only in Part V as comparators.

Five functions of each volume. Each later volume divides its subject into five functions (for example B1 to B5 in AI Blue, O1 to O5 in AI Operations, L1 to L5 in AI Leader, D1 to D5 in AI Developer, E1 to E5 in AI Environment: E1 Measure, E2 Justify, E3 Design, E4 Operate, E5 Account); the dossier names them as the volume does.

Footprint record (FPR-*). The record of the energy and carbon of one use or one model: boundary, year, method, location-based and market-based values side by side, and the range. FPR-HM-0001 to -0007 are Hansa Mutual's (Chapters 5 and 19A) (14, Nordqvist, 2026).

Gate. In Vibe to Dev, a point at which a step is finished only when an artefact is shown; at Hansa Mutual also the pipeline gate that handles routine classes (Chapters 3 and 19).

Group AI governance board. The board chaired by the chief risk officer that governs AI across the group (Chapter 3).

Hardware lifecycle record (HWL-*). The record that compares keeping, consolidating, replacing or stepping the replacement of accelerators on energy, operational carbon and embodied carbon; HWL-HM-0001 led to consolidation to nine servers (AIDR-1366) (Chapter 19A) (14, Nordqvist, 2026).

Hosting comparison sheet (HCS-*). A view over supplier disclosure records that compares own, cloud and regional hosting on one index; HCS-HM-2032-11 (Chapter 19A) (14, Nordqvist, 2026).

Hub and spoke. The structure in which platform and enabling teams sit in group IT in Hamburg and use cases are owned by the national subsidiaries (Chapter 3).

Illustrative. A date, figure or event that a volume itself marks as an illustration. The volumes were written in November 2026, so events dated later are illustrative (Chapter 21).

Konzernbetriebsrat. The group works council in Germany, whose agreements and annexes govern how AI may be used on staff (Chapter 3).

Ledger. data/ledger.jsonl, the machine-readable record of the identifiers, events and figures of the dossier, from which Appendix A is generated.

Level 3 Bounded. The degree at which an agent may act without approval inside a defined scope on reversible actions, with a stop path outside the agent (Chapters 5 and 7).

Location-based and market-based. The two ways of stating the carbon of electricity: the first by the grid's average factor (350 g CO2e a kWh for Germany in the profile of AI Environment), the second by the contracts the buyer holds, such as guarantees of origin. Hansa Mutual carries both side by side (Chapters 19A and 23) (14, Nordqvist, 2026).

Market overlay. The set of tightening controls that a country adds to the group baseline, with its flags (Chapter 4).

Mandate manifest (MND-*). The record that sets what an agent may do, with which identity, and until when (Chapter 22).

Mode 0 Ambient AI. AI that arrives inside a supplier's product without a decision by the organisation.

Mode 1 Conversational assistants. Chat-style assistants, at Hansa Mutual approved enterprise tiers.

Mode 2 Embedded workplace assistants. Assistants inside the office suite.

Mode 3 Agentic work tools. Tools that act on the user's behalf in an engineering or office setting.

Mode 4 Built AI applications. Applications the group builds, such as the claims assistant and the underwriting-support model.

Mode 5 Autonomous production agents. Agents with their own identity and write access in production. The dossier's working count is six (Chapter 5; Open thread OT-HM-01).

Model choice record (MCR-*). The record of the choice among models by task, with measured energy and quality; MCR-HM-0001 holds the cascade of a small and a large model (Chapter 19A) (14, Nordqvist, 2026).

Observer. A check on an action that is independent of the system that performed it, so that a loop closed by machines alone does not hide its own failure (Chapter 7).

Open thread. A point on which the volumes disagree or stop short, recorded with its working value in Chapter 27 as OT-HM-nn.

Placement policy (PLC-*). The policy that places work in time and region by carbon and constraints such as water or data class; PLC-HM-0001 (Chapter 19A) (14, Nordqvist, 2026).

Policy point. A control outside the model that decides whether an action may proceed (Chapter 7).

Power usage effectiveness (PUE). The ratio of a data centre's total energy to the energy that reaches the IT equipment; about 1.38 at Frankfurt (Chapters 4 and 19A) (14, Nordqvist, 2026).

Profile. The description of an organisation in the volume's EXAMPLE_ORGANISATIONS.md, extended by each volume.

Reach. In Landscape, the size class of the number of people an AI use touches in a year, from S to XL.

Reversibility tier. The class of how easily an action can be undone, from R1 reversible at once up (Chapter 5).

Role charter (ROLE-*). The record that gives a role its duties, vetoes and limits (Chapter 24).

Seat. See Decision seat.

Series state. The statement on the title page of Chapter 1 of which volumes and which date a release reflects.

Stewardship principle. In AI Developer, the rule that AI may draft and execute within a mandate but may not own; a named person passes the tests Name, Understand and Revert (Chapter 18).

Stop path. The means of ending an agent's actions or credentials from outside the agent, with a time budget; tested by a stop drill (Chapter 7).

Stop right. The right to stop a system. At Hansa Mutual any named holder may stop without prior approval, while restart and a higher level need further approvals (Chapter 3).

Supplier disclosure record (SDR-*). The record of what a supplier has disclosed about energy, carbon and water, with the questions answered and their validity (12 months); SDR-HM-0001 to -0003 (Chapter 19A) (14, Nordqvist, 2026).

Takeover test. In Vibe to Dev, the test that a named colleague can run, change and ship the change, and switch off and restore the result (Chapter 19).

Tier 0. The criticality class of systems whose compromise would compromise everything else; the operator's own AI is classed Tier 0 (Chapter 7).

Usage mode. One of six classes, Mode 0 to Mode 5, ordered by depth of integration and autonomy (Chapter 5).

Volume. One of the fourteen numbered books of The Transformed Architecture; the two companions Landscape and Tools and the two dossiers are not numbered volumes.

Watch list (WCH-*). The list of markers that would change a decision on environmental claims, named and kept by the board; WCH-HM-0001 (Chapter 19A) (14, Nordqvist, 2026).

Working value. The value the dossier uses where two volumes disagree. It is a reading for continuity and not a decision between the volumes (Chapter 27).