Glossary
This glossary defines the vocabulary of AI Environment: the five environmental functions, the net-benefit principle with its three tests, the optimisation hierarchy, the records and the identifiers that the chapters use, the measures and boundaries on which every figure of the book depends, and the technical, regulatory and organisational terms that the chapters define on first use for a reader who is not a specialist in energy, data centres or machine learning. Each entry gives the book's own meaning in one to three sentences and names the chapter that owns the term, where it is defined or first used in that sense. A term that the chapters use without defining is left out. Terms that earlier volumes of the series defined are repeated briefly, with the volume named, so that the reader is not left with a gap.
How to use this glossary
Entries are in alphabetical order. Identifiers in code formatting (FPR-*, EBG-*) are sorted by their letters. An identifier such as FPR-* names a pattern: the wildcard stands for the number that the organisation assigns to each record, and the book's examples use NR for Nordvik Region and HM for Hansa Mutual (for example FPR-NR-0001); the prefix BK is used once, for the footprint record of the production of this book (FPR-BK-0001), which is real and not illustrative. AI Decision Records run from AIDR-1300 to AIDR-1399 in this book: Nordvik Region AIDR-1300 to AIDR-1349 and Hansa Mutual AIDR-1350 to AIDR-1399. The five environmental functions E1 to E5 are those of this book, not the operations functions O1 to O5 of AI Operations, the leadership functions L1 to L5 of AI Leader or the developer functions D1 to D5 of AI Developer. Every figure that the book attributes to the two organisations is illustrative. The legal entries attribute each statement to the instrument and article as the chapters cite it, and describe organisational implications; they are not legal advice. The labels observable, expected and speculation mark every statement about the future and are defined below.
A
Accounting life. The number of years over which the cost of hardware is depreciated. It is a financial estimate and is not the length of time the hardware is useful (Chapter 16).
Adapter fine-tuning. Adapting a model to a task by training a small set of added parameters while the base model stays fixed, so that training costs far less energy than changing every parameter. Also called parameter-efficient fine-tuning (Chapter 14).
Agent. A system in which a model directs its own sequence of tool calls towards a goal. In this book an agent measures, estimates, recommends, schedules or executes within a mandate; it does not care, want or decide (Chapters 4 and 15).
Agent budget. The environmental budget of an agent, stated per call, per run and per period and enforced at the gateway, so that an agent that can start work cannot spend more than a person has allowed (Chapter 15).
Agent without a budget or a stop (pattern 6). An anti-pattern of E3 Design that breaks the Cap test: an agent that can start work runs with no environmental budget set by someone else, or with a stop path that has never been drilled. The environmental budget EBG-* and the mandate manifest MND-* prevent it (Chapter 26).
AI Act. Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. The book uses its terms exactly (provider, deployer, high-risk AI system, general-purpose AI model, systemic risk) and reads the technical documentation of a general-purpose AI model (Article 53(1)(a) and Annex XI, which ask for the computational resources used) as Chapter 21 cites it (Chapters 5 and 21).
AI Decision Record (AIDR). The record of a decision about an AI capability or about the delegation of a class of work. In this book they are numbered AIDR-1300 to AIDR-1399, for example AIDR-1300 (adoption of the environmental register and the Count rule, Nordvik Region), AIDR-1309 (proceed with the citizen-service assistant, no environmental claim), AIDR-1350 (adoption at Hansa Mutual) and AIDR-1372 (approval of claim CLM-HM-0001, revision 2) (Chapter 4).
AI Energy Score. A public benchmark of Hugging Face and partners that measures the energy of a model on a fixed task at a fixed boundary. It compares models and does not give the footprint of a use (Chapters 6 and 12).
AI-use inventory. The one-page list of an organisation's AI uses with an owner, a supplier, a place and a note of whether anything is measured. It is the seed of the environmental register and shows where the evidence is thinnest (Chapter 1).
Allocation. The rule by which a shared resource, such as a pool of accelerators or the embodied impact of a server, is divided among the uses that share it. Every footprint record states its allocation rule and the share it produces, for example the share of tokens through the gateway. ISO 14044 asks that allocation be avoided by subdividing the process or expanding the system and, where several procedures seem applicable, that a sensitivity analysis be run (Chapters 5, 6 and 16).
Ambient AI. See Mode 0 Ambient AI.
Annex VII (Energy Efficiency Directive). The annex of Directive (EU) 2023/1791 that lists the information a data-centre owner or operator publishes each year under Article 12(1), for sites of at least 500 kW installed IT power. Installed IT power, PUE and water input in SDR-NR-0001 come from it (Chapters 18 and 21).
Annual matching. Accounting in which the renewable electricity a company buys over a year is set against the electricity it uses over the same year, whatever the grid delivered in a given hour. See hourly matching (Chapter 22).
Annual matching presented as hourly (pattern 9). An anti-pattern of E5 Account: a claim that annual purchases of renewable electricity match consumption is presented as if supply matched demand hour by hour. The claim record states both bases and the matching rule (Chapter 26; see also Annual matching and Hourly matching).
Answerability principle (Name, Explain, Reach). The principle taken from AI Leader, Chapter 4: AI may inform, draft, prepare and, within a mandate, execute, but it may not answer. A named person answers for an environmental decision by passing three tests: Name (the person is identifiable in the record), Explain (the person can give the reasons without the system that produced the figures) and Reach (the person can change or stop the use) (Chapters 4 and 28).
Anti-pattern. A practice that looks reasonable at the time and reliably leads to a failure, such as a figure without a boundary or an agent without a budget. Chapter 26 sets out the patterns and places each on the function whose test prevents it.
Architecture as Code. The companion discipline of the series, in which architecture, policy and tests are kept as versioned, executable artefacts, and the title of Book 1.
Aspect, environmental. An element of an organisation's activities, products or services that interacts or can interact with the environment. In an environmental management system the organisation determines its aspects, decides which are significant and controls them; for AI the unit of an aspect is a use, not a site (Chapter 24).
Aspect, indirect environmental. In EMAS, an environmental aspect that can result from the interaction of an organisation with third parties and can to a reasonable degree be influenced by it, such as the environmental performance of suppliers. A supplier's AI footprint is an indirect aspect of its customer (Chapter 24).
Aspect, significant environmental. An environmental aspect that the organisation, by criteria it has documented, judges to have or be able to have a significant impact. In this book the environmental register carries the criteria and the result per use (Chapter 24).
Asserted. The weakest of the four evidence grades for a benefit claim: a statement by the operator that is credible on its face and has no stated method or baseline (Chapters 2 and 3).
Average carbon intensity. The mean emissions per kilowatt-hour of everything the grid generated in a period. It describes the grid and not what one more load costs. See marginal carbon intensity (Chapter 17).
B
Baseline. The value of a measure before the change, or in the control group, against which an effect is stated. A benefit without a baseline is an intention (Chapters 3 and 9).
Basin stress. The water scarcity of the river basin from which a site draws. In this book it is a separate column in a placement policy, filled from a named source and never inferred from the country (Chapters 7 and 17).
BEN-* (benefit record). The record of what a use achieves, against which alternative and with what rebound: intended benefit, baseline, alternative, measured effect, comparison, rebound, decision and expiry. BEN-NR-0001 and BEN-HM-0002 are the examples; the number matches the footprint record of the use (Chapters 4 and 9).
Benefit ledger. One of the separate accounts in which a benefit is recorded (service, cost, environment). The ledgers are not netted: a saving of cost or an improvement of service never reduces the footprint in the register (Chapter 9).
Boundary. The bundle of choices that fixes what a figure counts: operational or lifecycle, accelerator only, whole server or whole facility, location-based or market-based, withdrawal or consumption, on site or total, mean or median, and the unit. A figure without its boundary cannot be compared with any other (Chapters 2 and 4).
Burst capacity. Capacity rented on demand to cover a peak on top of a base that the organisation owns or has reserved. At Hansa Mutual the burst route into Cloud A is limited to 10 per cent of the triage agent's calls by EBG-HM-0001 (Chapters 15 and 19).
C
Cache. A store of earlier results that lets a system avoid repeating work. See prompt cache and response cache (Chapter 13).
Cache read. A token of earlier text that a model call loads from a cache instead of processing it again. In the production of this book cache reads were 97 per cent of all tokens, and the energy of one cache read is not published for the models used (Chapter 8).
Cap. The third test of the net-benefit principle: the use runs under an environmental budget with a stop path outside the system it stops (Chapter 4).
Carbon intensity. The grams of carbon dioxide equivalent emitted per kilowatt-hour of electricity, stated as an average or marginal figure, for a place and a period. It is a property of the grid and not of the workload (Chapters 6 and 17).
Carbon-aware scheduling. Moving deferrable work to the hours or regions in which the carbon intensity of the grid is lower, within the floors of the placement policy. Its saving is shown by replay and not assumed (Chapter 17).
Cascade. A routing design in which a small model answers first and a larger model is used only when a rule evaluated outside the model says so, for example a low retrieval score. Its energy depends on the escalation rate (Chapter 12).
Claim kind. The class of an environmental claim: footprint, benefit or commitment. A claim record holds one kind, because each has its own evidence (Chapter 22).
Claim published without review (pattern 10). An anti-pattern of E5 Account: an environmental statement goes out without an approver other than its producer. The claim record names the approver, who is a person and never an agent (Chapter 26; see Environmental claim).
CLM-* (environmental claim record). The record that must be complete before an environmental statement is published: statement, claim kind, object, boundary, period, baseline, evidence, independent check, reviewers and a named approver. CLM-HM-0001 and CLM-NR-0001 are the examples (Chapters 4 and 22).
CMP-* (competence record). The record that a named person has shown a competence, by whom, when and with what result, and when it expires. CMP-NR-ENV-02 records that Sara Ekdahl re-performed a footprint record from raw meter data (Chapter 24).
Compare. The second test of the net-benefit principle: the benefit is measured against the non-AI alternative and against the rebound the use causes (Chapters 4 and 9).
Consumption (water). The part of the water withdrawn that is not returned to the source, mostly by evaporation in cooling. See withdrawal (Chapters 2 and 7).
Context length. The amount of text, counted in tokens, that a model processes for one request. A long context raises the work of the prefill stage and, in a conversation or an agent loop, is resent at every step (Chapters 6 and 13).
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. The Count rule forbids a figure in any record, table or sentence without its boundary and year (Chapter 4).
Counted part. The share of a footprint that falls inside the stated boundary of a record. The excluded part is listed beside it and can only add to the total (Chapter 8).
Counterfactual. What would have happened without the use. A benefit is measured against a counterfactual, by a control group, a before-and-after comparison, a model or the non-AI method (Chapters 3 and 9).
Critical review. An independent evaluation of a life cycle assessment against the requirements of the standard. For a study that supports a public comparative assertion, ISO 14044 requires a review panel of at least three members chaired by an independent external expert (Chapter 5).
CSRD. Directive (EU) 2022/2464, the Corporate Sustainability Reporting Directive, with the ESRS as its reporting standards. It governs how Hansa Mutual reports energy and emissions, including AI inside Scope 2 and Scope 3 (Chapters 21 and 22).
CUE. Carbon usage effectiveness, the carbon emitted per unit of IT energy of a data centre. It is not defined in Delegated Regulation (EU) 2024/1364, and the book describes it in general terms only (Chapter 18).
Cut-off criteria. The rules that state which processes or flows may be left out of a life cycle study; ISO 14044 requires that the effect of leaving them out is assessed and described. Not the same as the cut-off allocation rule (Chapter 5).
Cut-off rule (allocation). The allocation rule under which a user carries none of the shared pre-training of a model, only its own fine-tuning and inference. It answers the cost of the organisation's own decisions and understates the cost of the model's existence (Chapter 5).
D
Decide (the verb). Used in this book only for a system to which a decision was formally delegated, and the record then says who delegated it. Otherwise a system measures, estimates, recommends, schedules or executes (Chapter 4).
Degrees of delegation. The series' five autonomy levels per class of action: Level 0 Advisory, Level 1 Assisted, Level 2 Supervised, Level 3 Bounded and Level 4 Autonomous. The action classes of this book are measure, recommend, schedule, provision, scale, stop and publish (Chapter 4).
Deployer. The AI Act's term for the organisation that uses an AI system under its own authority. Both organisations of the book are deployers (Chapters 5 and 21).
DLG-* (delegation profile). The record of the highest degree of delegation for each class of action of an agent. DLG-HM-ENV-0001 is the ceiling of the claims-triage agent's environmental mandate and DLG-NR-ENV-0001 fixes which breach actions are automatic (Chapters 15 and 20).
Dominating assumptions. The few inputs that account for most of the range of a footprint figure, such as the draw per accelerator, the grid factor and the pool share. A record names them so that the next measurement can start there (Chapter 6).
E
E1 Measure, E2 Justify, E3 Design, E4 Operate, E5 Account. The five environmental functions of the book. E1 counts the footprint (Chapters 5 to 8), E2 decides whether a use is worth it (Chapters 9 to 11), E3 chooses and shapes the solution for less (Chapters 12 to 16), E4 places, times, hosts and runs it (Chapters 17 to 20) and E5 complies, reports, claims, procures and governs (Chapters 21 to 25) (Chapter 4).
ECAL-* (evidence-age calendar). The file that turns the validity values of the register into dates on which each decision can be reopened. For each item the review date is the expiry minus 14 days, and the calendar is read by a script that runs outside any agent. ECAL-HM-0001 is Hansa Mutual's (Chapter 28).
Ecodesign Regulation. Regulation (EU) 2024/1781, the Ecodesign for Sustainable Products Regulation. Article 79 continues certain provisions of Directive 2009/125/EC for servers until 31 December 2026 (Chapter 16).
Economic life. The period over which keeping a machine is cheaper than replacing it. It differs from the accounting life and from the service life (Chapter 16).
Efficiency. Lowering the footprint of a unit while delivering the same unit. Contrast sufficiency (Chapter 25).
Egenkontroll (own-checking). The duty under Chapter 26, section 19 of the Swedish Environmental Code (Miljöbalken) on whoever conducts an activity that may harm health or the environment to plan and check it continuously and to keep informed of its impact (Chapter 24).
EMAS (Eco-Management and Audit Scheme). The EU's voluntary scheme under Regulation (EC) No 1221/2009. It requires the clauses of EN ISO 14001:2015 and adds an environmental review, an audit, core indicators and a public environmental statement verified by an accredited verifier (Chapters 21 and 24).
Embedding. A vector that represents a piece of text so that similar texts can be found by distance. Computing and storing embeddings is the energy cost of building a retrieval index (Chapter 14).
Embodied impact. The emissions and materials used before a device runs: mining, refining, fabrication, assembly and transport. In a footprint record it is carried as a separate, labelled field, with a stated allocation, or marked as unsourced (Chapters 2, 5 and 16).
EMS (environmental management system). The policy, roles, procedures, records and reviews by which an organisation plans, controls and improves its environmental performance; ISO 14001 and EMAS specify one. Adapted for AI, the aspects register becomes the environmental register and operational control becomes the budget and the stop path (Chapter 24).
Energy Efficiency Directive. Directive (EU) 2023/1791 (recast). Article 12(1) requires Member States to require owners and operators of data centres of at least 500 kW installed IT power to publish the information of Annex VII each year. The Swedish rule the register cites is Lag (2025:570) (Chapters 18 and 21).
Energy reuse factor (ERF). Heat reused outside the data-centre boundary divided by the facility's energy, as defined in Delegated Regulation (EU) 2024/1364. It is reported and is not a class on the rating label (Chapter 18).
Environmental benefit. What a use achieves for the environment, shown with a baseline and a comparison. A service or cost benefit is a benefit, but it is recorded in its own ledger and is not an environmental benefit (Chapters 3 and 9).
Environmental budget (EBG-*). A limit on one or more of energy, carbon, water, cost or compute for one use or one agent over a period, set by a person who is not the agent, measured outside the system it limits, enforced and paired with a stop path outside the system it stops. A budget with no recorded stop drill is not counted as a cap (Chapter 4).
Environmental claim. A statement about the environmental effect of an organisation, a use or a product that is published or sent to a customer. It needs a claim record (Chapter 22).
Environmental footprint. What AI costs the environment: energy, carbon, water, materials and land, over the life of models and hardware. It is counted per use and at a stated boundary (Chapters 1 and 4).
Environmental owner. The named person who answers for the environmental record of one AI use: the row is complete and dated, the footprint is current, the budget is respected and the stop path is known. ROLE-NR-ENV-01 is the charter (Chapter 24).
Environmental register. The register with one row per AI use: owner, footprint record, benefit record, budget, the date of each piece of evidence and the decision. It is evaluated by a script that reports what is valid, expired or missing (Chapters 4 and 24).
ESRS. The European Sustainability Reporting Standards under the CSRD. The revised standards entered into force on 10 November 2026 by Delegated Regulation (EU) 2026/1563 (Chapter 21).
Evaluation set. A fixed, labelled set of cases, never used for tuning, on which candidate models are measured for quality and energy at the same boundary (Chapters 12 and 13).
Evidence age. The time since a piece of evidence was produced, compared with the limit set for its kind. In this book a measured footprint of a running use is valid for 90 days, a benefit record for 180 days, a stop drill for 90 days, a supplier disclosure for 12 months and a claim for the reporting year at most; a change of model, region or volume ends validity earlier (Chapters 2, 4 and 24).
Expected. One of the three outlook labels, after AI Governance, Chapter 22: a statement about the future that rests on a decided or announced instrument, a published plan or a trend that a source documents. See observable and speculation (Chapter 2).
F
Figures compared across boundaries (pattern 1). An anti-pattern of E1 Measure that breaks the Count test: two figures with different boundaries, carbon bases or years are placed in one table or one sentence as if they were comparable. The boundary fields of the footprint record prevent it (Chapter 26).
Floor. A condition that comes before any saving from placement or timing: an interactive or latency-bound use, special-category data, a deadline closer than the run time, a region with no record. Floors come before intensity (Chapter 17).
FPR-BK-0001. The footprint record of the production of this book, measured on 16 November 2026 from recorded agent transcripts. The prefix BK stands for the book; it is the one real, non-illustrative footprint record in the book (Chapter 8 and Appendix A).
FPR-* (footprint record). The record of what one use, model, agent or pool costs the environment: functional unit, period, volume, boundary, method, multipliers, result with a range, dominating assumptions, independence and evidence age. FPR-NR-0001 and FPR-HM-0002 are the examples (Chapters 4 and 6).
Functional unit. The unit of work that cost, footprint and benefit are all stated per: a resolved enquiry, a claim triaged, a query answered, an item routed. It is chosen before counting begins. In a life cycle assessment the functional unit must be clearly defined and measurable; the per-use unit of Chapter 6 is an intensity metric of this kind and becomes a functional unit in the LCA sense only when allocated training, embodied and end-of-life shares are included (Chapters 4, 5 and 6).
G
Gate (scale-up). The decision point at which a proposal to extend a use is tested against Count, Compare and Cap. AIDR-1326 and AIDR-1376 are the Nordvik Region and Hansa Mutual decisions at the gate (Chapter 24).
Gateway. The component through which a use calls a model, and which holds the counters, the credential and the quota. It enforces budgets and meters volume outside the assistant or agent (Chapters 10, 15 and 20).
GreenOps. The practice of managing the environmental footprint of IT in operation, in the way FinOps manages cost: tagging, metering, budgets and review. In this book it is the work of the platform and finance functions under the register owner (Chapter 20).
Guarantee of origin. A certificate that one megawatt-hour of electricity was generated from a renewable source. Buying it can lower a market-based figure and does not change the energy consumed or the location-based figure (Chapters 18 and 22).
H
Hard limit. A limit at which the gateway stops a call, run or credential, as opposed to a soft limit, which warns. A limit that is not enforced is a request (Chapter 15).
Hardware bought for a peak (pattern 7). An anti-pattern of E3 Design: capacity is bought for the highest expected load and then runs at low utilisation, so that the embodied impact and the idle energy are spent for a load that seldom comes. The hardware lifecycle record HWL-* prevents it by requiring measured utilisation and a consolidation test before a purchase (Chapter 26).
Heat reuse. The share of the heat a facility produces that is delivered to someone who uses it, normally through a district heating network (Chapters 18 and 19).
Hold-out. A share of users kept on the old way of working for a fixed period, so that a service already rolled out to everyone still has a control group (Chapter 9).
Hourly matching. Accounting that sets renewable electricity against consumption hour by hour, also called 24/7 carbon-free energy. It is the stricter alternative to annual matching (Chapter 22).
HWL-* (hardware lifecycle record). The record of a hardware pool or a supply contract: inventory, utilisation, energy, embodied impact as a lower bound, options with their totals, decision and end-of-life route. HWL-HM-0001 is an inventory record and HWL-NR-0001 a clause set (Chapters 4 and 16).
I
Idle capacity. Capacity that is paid for, powered and unused. At Hansa Mutual the top-down figure for the triage agent includes idle capacity held for it, and is about thirteen times the bottom-up figure at one boundary (Chapters 6, 8 and 9).
Illustrative. The label on every figure of Nordvik Region and Hansa Mutual. These figures show the shape of a calculation and are never data (Chapter 1).
Impact category. A class of environmental effect for which an indicator is calculated in an impact assessment, such as climate change, water use or resource use of minerals and metals. ITU-T L.1801 makes climate change mandatory (Chapter 5).
Independent check. A check of a figure by a person or source that did not produce it and does not depend on the producer: a second meter, an invoice, a reviewer who did not make the change. Its result is recorded beside the figure, not inside it (Chapters 4, 6 and 22).
Inference. Running a trained model to produce an output. Whether inference or training is the larger share of a model's energy over its life depends on volume, utilisation and the grid and cannot be settled in general (Chapters 5 and 6).
ISO 14001. The international standard for environmental management systems. The 2026 edition was published on 15 April 2026, and certificates to the 2015 edition must transition by 30 April 2029, as reported by an accreditation body; the 2026 text was not read for this book (Chapter 24).
ISO 50001. The international standard for energy management systems. Its edition status was not checked for this book; Article 11 of the Energy Efficiency Directive requires certain enterprises to have a certified energy management system (Chapters 21 and 24).
ISO/IEC 42001. The standard for AI management systems. It shares the harmonised structure of ISO 14001; as read in a secondary guide, it lists environmental impact as a possible AI objective but has no environmental control in its Annex A and asks for no footprint or budget (Chapter 24).
ITU-T L.1801. The Recommendation of the International Telecommunication Union, approved on 6 February 2026, that gives guidelines for assessing the environmental impact of AI systems by life cycle assessment. It is voluntary and maps initial training to the production stage and inference to the use stage (Chapter 5).
J
Jevons effect. The observation that a gain in efficiency can lead to a rise in total use. It is consistent with the data on AI and is not the only explanation of it (Chapter 10).
K
Key-value cache. The memory in which a model keeps the intermediate results of the tokens it has already processed, so that it does not recompute them at each step. Its size grows with context length and limits how many requests a server can batch (Chapter 13).
L
Lagrådsremiss. In Sweden, a government bill sent to the Council on Legislation before the proposition to the Riksdag. It is a proposal, not law (Chapter 21).
Life cycle assessment (LCA). The standardised method (ISO 14040 and ISO 14044) for counting the environmental impacts of a product system from raw material extraction to disposal, in four phases: goal and scope, inventory, impact assessment and interpretation. A footprint record of this book is an inventory with a limited set of categories and is not an LCA until it has impact assessment, interpretation and, for a public comparison, review (Chapter 5).
Lifecycle (boundary). A boundary that adds the manufacture of the hardware and, in the widest form, its disposal to the operational energy. Contrast operational (Chapters 2 and 4).
Life cycle inventory (LCI). The phase, and the record, of the inputs and outputs of a product system. ISO 14044 says that an inventory alone cannot support a public comparative assertion (Chapter 5).
Location-based. A method that applies the average emissions of the grid where the electricity is used. It is the method that Scope 2 Guidance requires to be reported beside the market-based figure (Chapters 2, 6 and 22).
M
Mandate manifest (MND-*). The record of what an agent may do, with which resources, under whose decision and with what stop path. In this book MND-NR-ENV-0001 governs the platform team's agentic work tools and MND-HM-ENV-0001 the claims-triage agent (Chapter 15).
Marginal carbon intensity. The emissions of the plant that would change its output if demand changed by a little. It is the figure that matters for adding or moving load, and it is usually higher than the average (Chapter 17).
Market-based. A method that applies the emissions factor implied by the contracts the buyer holds, such as power purchase agreements and guarantees of origin. It can lower the reported figure whatever the grid delivered in an hour (Chapters 2, 6 and 22).
MCR-* (model-choice record). The record of a choice of model: the requirement written before any candidate was run, the candidates with quality and energy measured at one boundary, the choice, the date, and the triggers for review. MCR-NR-0001 and MCR-HM-0001 are the examples (Chapter 12).
Measured in deployment. The strongest of the four evidence grades for a benefit claim: an effect observed in a running use, against a stated baseline, at a stated boundary and for a stated period. See modelled, projected and asserted (Chapter 3).
Measurement that stops, or never starts (pattern 3). An anti-pattern of E1 Measure: a use is unmetered, or its last measurement has expired, and the register still shows a number or a blank that nobody reads. The register row with its evidence age prevents it (Chapter 26).
Memo line. A labelled line in a group statement that reports the part of an organisation's footprint that it can count for AI, with location-based and market-based figures side by side, where AI cannot be separated inside Scope 2 and Scope 3. It is not a restated inventory (Chapters 22 and 28).
Mixture-of-experts. A model design in which only part of the parameters is used for each token, so that the energy depends on the active size and not on the total size (Chapter 13).
Mode 0 Ambient AI to Mode 5 Autonomous production agents. The series' six usage modes, as named in AI Governance, Chapter 2: Mode 0 Ambient AI, Mode 1 Conversational assistants, Mode 2 Embedded workplace assistants, Mode 3 Agentic work tools, Mode 4 Built AI applications and Mode 5 Autonomous production agents. They set what the organisation can meter itself (Chapters 2 and 4).
Modelled. An evidence grade: an effect or footprint calculated from a model of the system and not observed. See measured in deployment, projected and asserted (Chapters 2 and 3).
Model-scope footprint. The footprint of a model across its life (training, tuning, hardware) as distinct from the footprint of a use. It is allocated to uses by a stated rule (Chapter 5).
N
Necessity. The first level of the optimisation hierarchy: do not run what is not needed (Chapter 4).
Neither side pays for the other. The rule that an unproven benefit does not offset a measured footprint and a purchased offset does not replace reduction. A proven benefit can justify a footprint and does not reduce it (Chapter 4).
Net-benefit principle. A use of AI is justified only when its benefit per unit of environmental cost is shown with evidence. It has three tests, applied in order: Count, Compare and Cap (Chapter 4).
Non-AI alternative. The best real alternative way of doing the same work without AI: the phone service, a rule, a smaller or older model, a process change, or no service. It is measured on the same functional unit (Chapters 4 and 9).
Nonconformity. In a management system, a requirement that is not met. In this book it is a failed stop drill, an exceeded budget, an expired footprint figure or a withdrawn claim; it is corrected, its cause is found and the effect of the correction is reviewed (Chapter 24).
O
Observable. One of the three outlook labels: a statement that rests on something documented now. See expected and speculation (Chapter 2).
Observer principle. The principle of AI Operations, Chapters 4 and 17, that what verifies an action must be independent of what performed it. The environmental budget applies it: the meter and the stop path lie outside the agent they limit (Chapter 4).
On site and off site (water). On-site water is used in the facility's cooling and humidification. Off-site water is used to generate the electricity the facility consumes. The two are never added without saying so (Chapter 7).
One supplier's figure taken as independent (pattern 2). An anti-pattern of E1 Measure: a figure reported by a supplier is used, or repeated, as if it had been checked by someone else. The kind field of the supplier disclosure record SDR-* and the independent check prevent it (Chapter 26).
Operational (boundary). A boundary that counts the electricity a use consumes while it runs and leaves out the manufacture of the hardware. Contrast lifecycle (Chapters 2 and 4).
Optimisation hierarchy. The five ordered levers for making a use cost less: Necessity, Sufficiency of model, Efficient serving, Place and time, Hardware and lifecycle. A lower level is no reason to skip a higher one, and each is tested for the saving it delivers (Chapter 4).
Output limit. A limit on the length of a model's answer, set at the gateway. It lowers the energy of generation and must be tested against quality (Chapter 13).
P
Paired outcome. An outcome measure kept beside a budget so that the budget cannot be met by handing work to people, for example the share of claims that go to a handler (Chapter 15).
Pattern check. The ten-question check of Chapter 26, run on one register row by the register owner and reviewed by the sustainability lead, which finds which of the ten anti-patterns are open by comparing each statement with its record. Its result is valid for 90 days, like the footprint it reads (Chapter 26).
Pay-back period. The time, in months, after which the energy saved by a trained model equals the energy spent on training it: the pay-back volume N* (the number of uses at which the two are equal) divided by the monthly volume. Training is justified only if the pay-back is reached with its range (Chapter 14).
Per-unit gain taken for a total reduction (pattern 5). An anti-pattern of E2 Justify: a fall in energy per unit is reported as a fall in total energy while volume grows. The rebound block of the register row shows unit and volume together (Chapter 26; see Rebound and Jevons effect).
PLC-* (placement policy). The record of where and when deferrable work may run: signal, regions, classes of work with windows and deadlines, floors, saving threshold, budget and stop path. PLC-NR-0001 and PLC-HM-0001 are the examples (Chapters 4 and 17).
Prefill. The first stage of generation, in which the model processes the whole input before the first output token. It grows with context length (Chapter 6).
Product Environmental Footprint (PEF). The European Commission's method (Recommendation (EU) 2021/2279) for the life cycle environmental performance of a product, with 16 impact categories, and the reference method named in the Ecodesign Regulation's definition of environmental footprint. Only the PEF method gives a formal definition of a screening step (Chapter 5).
Production-boundary reading. A measurement of a use as it runs in production, from tagged rack meters, as against a reading on a test pool. A claim that a test pool can no longer support needs one before it is renewed (Chapter 27).
Projected. An evidence grade: an effect forecast from modelled potentials, with no observation behind it (Chapters 2 and 3).
Projections R1 to R6. The author's six projections about the evidence and the law, written on 16 November 2026 and each labelled expected or speculation, with what would change it. They are, with the projections of Chapter 25, the only places where the book goes beyond the evidence, and they are reviewed on 16 November 2027 (Chapter 29).
Prompt cache. A cache of the processed form of a repeated prefix of a prompt, so that it is not computed again (Chapter 13).
Proxy energy per token. An energy per token taken from a published model of other systems and applied to tokens counted elsewhere, because the provider has not published its own (Chapter 8).
Proxy (measure). A quantity that stands in for one that is not metered, such as calls, tokens or billing hours in place of kilowatt-hours. It is labelled as a proxy and is used for trend and volume, not for the footprint (Chapters 8 and 15).
Publish (action class). The class of action in which an environmental claim is made public. It is never delegated to Level 4: an agent may draft a claim, a named person publishes it (Chapter 4).
PUE. Power usage effectiveness: the total energy of a data-centre facility divided by the energy of its IT equipment, as defined in Delegated Regulation (EU) 2024/1364. A PUE of 1.21 means 0.21 kWh of overhead for each kilowatt-hour of IT (Chapters 2 and 18).
Q
Quantisation. Storing and computing a model's parameters at lower numerical precision to reduce memory and energy. Its effect on quality is measured, not assumed (Chapter 13).
Quota. A limit in units of work, derived once at approval from an energy budget and the energy per unit, and enforced at the gateway. EBG-NR-0001 is 160,650 resolved enquiries a month (Chapters 10 and 20).
R
RBK-* (runbook manifest). The record of how a breach of a budget is handled: trigger ladder, delegation, steps, fallback, verification, stop and review. RBK-NR-ENV-001 is the example (Chapter 20).
Rebound. The gap between the saving that an efficiency gain promises and the saving that the total contains, because cheaper or easier service invites more use. It is measured at the use, in the organisation and in the economy, and not assumed away (Chapters 4 and 10).
Reference flow. The amount of product needed to fulfil the functional unit, for example the tokens that produce one answer or one case decided (Chapter 5).
Refresh list. The table of the book's dated snapshots (regulatory status, demand and grid figures, provider disclosures) that will age first, with the chapters that depend on each, the status on 16 November 2026 and the label of the outlook. It is re-read at each review date and a stance changes by decision, not by drift (Chapter 29).
Register owner. The person who keeps the environmental register, runs its age check and reports on it: Henrik Dahl at Nordvik Region and Anja Weiß for the platform rows at Hansa Mutual (ROLE-HM-ENV-02) (Chapter 24).
Renewable energy factor (REF). Renewable electricity divided by the facility's energy, as defined in Delegated Regulation (EU) 2024/1364. It does not say where or when the renewable electricity was generated (Chapters 18 and 19).
Requirement block. The part of a model-choice record, written and signed before any candidate is run, that fixes the task, the functional unit, the quality threshold, the latency and the data conditions (Chapter 12).
Response cache. A cache of whole answers to repeated requests. An exact-match FAQ cache saves energy only if repeats occur (Chapter 13).
Re-tiering. Moving a workload to older or smaller hardware that still meets its requirement, so that newer hardware is kept for the work that needs it (Chapter 16).
Retraining cadence. How often a model is retrained, set by a measured trigger and a training budget and not by the calendar (Chapter 14).
Retrieval. Supplying a model with relevant text found in an index at the time of the request, in place of training the knowledge into it (Chapter 14).
ROLE-* (role charter). The record of what a role answers for, what it does not hold and the ceiling for AI in it. ROLE-NR-ENV-01 is the environmental owner of an AI use and ROLE-HM-ENV-02 the register owner (Chapter 24).
Router. The component that decides which model a request reaches, by rules evaluated outside the model. Its allow-list limits the models the route can reach (Chapter 12).
Rung. A position on the ladder of supplier evidence, from 1 (metered by the supplier on a stated boundary, with an independent check) to 5 (not measured). A record names the rung it stands on (Chapter 8).
S
SCI (Software Carbon Intensity). The specification of the Green Software Foundation, adopted as ISO/IEC 21031:2024, in which the rate is ((E × I) + M) per R: energy, carbon intensity of that energy, embodied emissions of the allocated hardware, and the functional unit (Chapters 4 and 8).
Scope 2 and Scope 3. The categories of the Greenhouse Gas Protocol: Scope 2 is the emissions from the electricity, steam, heat and cooling the organisation buys, and Scope 3 is everything else in its value chain, including what suppliers emit on its behalf. Scope 2 can be calculated by the location-based and the market-based method, and both are reported (Chapters 2 and 22).
Scorecard (thesis). The record that states, for each of the five parts of the thesis, what is claimed, on what evidence, how old the evidence is and when it is next read. Table 29.1 gives the book's own, and AIDR-1329 is the decision with which Nordvik Region adopted its own as a file (Chapter 29).
Screening LCA. An LCA that uses readily available data and a fixed goal, unit, boundary and set of categories. It may inform an internal decision and cannot support a public comparison; the working rule is the book's own, built on the screening step of the PEF method (Chapter 5).
SDR-* (supplier disclosure record). The record of what a supplier says, at what boundary and year, on which rung, and what it did not give, with the date from which each item is stale. SDR-NR-0001 and SDR-HM-0002 are the examples (Chapters 4, 8 and 23).
Sentinel set. A fixed set of labelled queries, never used for tuning, run through a test pool before and after a change so that the energy and the quality are compared on the same work. SENT-HM-007 has 400 queries (Chapter 13).
Service owner. The named person who owns a service and its volume. In the register the owner of a use is the owner of its environmental row (Chapters 6 and 24).
Shifted set. A test set that differs from the training data, for example letters on a form introduced after training. Quality on it is a separate requirement in a model-choice record (Chapter 12).
Soft limit. A limit at which the gateway warns and does not stop. See hard limit (Chapter 15).
Sovereign or regional cloud. A cloud offer in which the provider, the data location and the control are European or national. In this book it is one of the hosting options that are compared by the same ten questions, with no option excluded on environmental grounds before the answers are in (Chapter 19).
Speculation. One of the three outlook labels: a statement about the future that rests on neither a documented fact nor a decided instrument. The author's own projections are kept to Chapters 25 and 29 and labelled as such (Chapter 2).
Stale (quota). A quota that no longer rests on a valid conversion factor, because the factor was re-measured and drifted beyond the tolerance. At Nordvik Region a drift beyond 0.05 makes the quota stale at once and it is re-derived (Chapters 20 and 27; AIDR-1328).
Stop drill. A rehearsal in which the stop path is used and timed, with the date, the result and the minutes recorded. A stop never drilled is not counted, and a drill is valid for 90 days (Chapter 15).
Stop path. The way in which a named person halts or reduces a use without the use's cooperation, outside the system it stops: revoking a credential, closing a route, setting a queue quota to zero (Chapters 4 and 15).
Sufficiency. Lowering the absolute level of use, as distinct from the footprint of a unit. Contrast efficiency (Chapter 25). As a level of the optimisation hierarchy, Sufficiency of model means using the smallest model that meets a stated requirement (Chapter 4).
Supplier disclosure. What a supplier reports about energy, water, grid, hardware and method for the capacity it sells. It is vendor-reported unless a check says otherwise (Chapters 8 and 23).
System boundary. The statement of which processes belong to the product system studied in a life cycle assessment. It is stated with every figure; see Boundary (Chapter 5).
T
TA-* (tender annex). The annex to a procurement that makes environmental disclosure a minimum specification and an award criterion. TA-NR-ENV-01 carries 12 of 100 quality points (Chapter 23).
Time-shifting. Running deferrable work in a different hour. A placement policy enables it only after a replay record shows a saving above the threshold (Chapter 17).
Token. The unit in which a language model reads and writes text, roughly a short word or part of one. Tokens through the gateway are the usual basis for allocating a pool to a use (Chapters 6 and 8).
Total water. Water withdrawn on site plus the water used to generate the electricity consumed. See withdrawal and on site and off site (Chapters 2 and 7).
Training. Producing a model by optimising its parameters on data. It includes development runs and failed runs, so that development is a multiple of the final run (Chapters 5 and 14).
U
Unit footprint. The footprint per functional unit. Total footprint is the unit footprint multiplied by the volume, and both factors belong to a named owner (Chapters 4 and 10).
Unproven benefit set against a footprint (pattern 4). An anti-pattern of E2 Justify that breaks the Compare test: a benefit that has not been measured against a baseline and a non-AI alternative is offset against a measured footprint. The benefit record BEN-* prevents it (Chapter 26; see Neither side pays for the other).
Usage modes. See Mode 0 Ambient AI to Mode 5 Autonomous production agents.
Use (the unit of management). One service, route or agent doing one kind of work for one purpose. The method manages the use and not the model, because the owners of the uses decide how often, with how much context and where a shared model is called (Chapter 4).
Utilisation. The share of the time or capacity of hardware that does useful work. Low utilisation raises the footprint of each unit, and the figure carries its period and its definition (Chapters 13 and 16).
V
Vendor-reported. A figure published by the party that sells the thing measured. It is labelled as such and is not treated as independent (Chapters 2 and 3).
Volume trigger. A level of volume above which the register owner is told and the budget is reopened, because cheaper answers have drawn more queries. EBG-HM-0003 sets it at 1,210,000 queries a month (Chapters 10 and 13).
W
Watch list. A dated list of five markers with a trigger for each direction, reviewed every 90 days, that says what would change the verdict of the debate of Chapter 11. WCH-NR-0001 is the Region's (Chapter 11).
Water stress. The scarcity of fresh water in a region or river basin. A placement policy excludes regions above a stated level, and treats an unknown level as stressed for a move into a region (Chapters 7 and 17).
Water usage effectiveness (WUE). The total water input of a data centre divided by the energy of its IT equipment, as defined in Delegated Regulation (EU) 2024/1364 (Chapters 2, 7 and 18).
What would change it. The last element of a conclusion in this book: the dated evidence whose arrival would reverse the decision or the verdict. It is kept in the evidence-age calendar and, for the thesis, in Table 29.3 (Chapters 28 and 29).
Withdrawal (water). Water taken from a river, aquifer or mains supply. It is not consumption, which is the part not returned (Chapters 2 and 7).
Work moved to a clean region that is short of water (pattern 8). An anti-pattern of E4 Operate: work is placed where the grid is cleaner without reading the water stress of the destination. The placement policy PLC-* prevents it by requiring a filled and tested water field for every destination (Chapter 26).
Works agreements and employee representatives. Environmental telemetry can become a measure of a person. It is therefore reported by team, never per person, and negotiated with employee representatives before switch-on (MBL-NR-IT-2027-01 at Nordvik Region; section 87(1) no. 6 of the Works Constitution Act at Hansa Mutual) (Chapters 4 and 15).
Y
Year plan. The plan for the first year of a register, in which each environmental function has a gate and a dated artefact and every expiry is on the calendar from the first day (Chapter 27).