Book Blueprint and Chapter Briefs

This file records the editorial brief for each chapter of AI Environment. The briefs below are generated from scripts/outline_data.py by scripts/generate_outlines.py; edit the data, not the generated section. The design of the whole book is fixed in DISPOSITION.md. Figures named in the briefs come from the research dossiers in research/, several of which were seen only as search snippets: the writer verifies every figure against the primary source, states its boundary and year, and drops what cannot be verified. Items marked [verify] concern law or numbers that must be checked before they are cited.

Part I: The Question

Chapter 1: Introduction: AI and the Environment, Seen from Europe

  • File: 01_introduction.md
  • Status: draft
  • Purpose: Introduce the book to a practitioner or decision-maker who has heard both that AI is an environmental catastrophe and that it will solve the climate. State the promise and its limits: the book does not take a side in advance; it gives a method for deciding, use by use, whether a benefit is worth its footprint, and what evidence a claim needs. Introduce the five environmental functions, the net-benefit principle (Count, Compare, Cap) and the optimisation hierarchy on one page; the two organisations; the rule that every figure carries its boundary and year; and reading paths for sustainability leads, IT and cloud heads, platform and FinOps practitioners, procurement officers, board members and auditors.
  • Builds on: AI Operations, Chapter 11; AI Platform and Security, Chapter 22; AI Governance, Chapter 22; AI Leader, Chapter 12
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: Two headlines, one ledger; What this book promises; Footprint and benefit; The five functions, the hierarchy and the principle on one page; The standard of evidence; Two organisations; A currency warning; How to read this book
  • Practical guides: A one-page inventory of the AI uses in your organisation, with owner, supplier, region and whether anything is measured
  • Worked examples: The two organisations' opening questions, stated in their own words
  • Nordvik Region and Hansa Mutual: Nordvik Region: buys AI mostly as services through frameworks and a regional cloud and measures little; the opening question is how to put environmental evidence into contracts and how to handle a citizen-facing assistant whose volume grows. Hansa Mutual: runs its own and cloud GPU capacity and reports under the CSRD regime in the form in force in 2026; the opening question is how to count AI inside Scope 2 and 3 and make the board's claims defensible. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Introduces the usage modes as the series uses them and says that footprint and delegation differ across them; the degrees of delegation are introduced with the Level 4 prohibitions of this book.
  • Observable, expected, speculation: Observable: growing AI energy use and growing disclosure by providers. Expected: more regulation of disclosure. Speculation: any statement about the net effect of AI on the climate. The thesis is the book's own claim, judged in Chapters 25 and 29.
  • Figure: The five environmental functions and the net-benefit principle on one page, with the optimisation hierarchy beneath
  • Planned length: 4,500-6,000 words of prose
  • Evidence to verify:
  • Public positions to contrast, each with a primary source: the IEA Energy and AI work (2025, with its 2026 updates) and the 'AI Climate Hoax' report of Beyond Fossil Fuels et al. (Feb 2026) [verify both against the primary text; the dossier saw them as summaries]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 2: What the Evidence Says Now: Energy, Carbon, Water and Materials

  • File: 02_what_the_evidence_says.md
  • Status: draft
  • Purpose: Report what the evidence says now about AI's energy, carbon, water and materials, with sizes, shares and uncertainty. Separate the footprint of data centres in general from AI in particular, vendor-reported from independent figures, measurement from projection, and operational from lifecycle. Give the reader a table of boundaries to hold while reading any figure, and show with one worked contrast how two honest figures can differ by an order of magnitude.
  • Builds on: AI Platform and Security, Chapter 22; AI Operations, Chapter 11; AI Governance, Chapter 22
  • Environmental functions: E1 Measure
  • Sections: Why boundaries come first; Energy: sizes and shares; Carbon; Water; Materials and hardware; Measured, modelled, projected, asserted; Two figures that disagree; What this means for the buyer
  • Practical guides: Reading a figure: the six-question boundary check, applied to three published figures
  • Worked examples: The same prompt counted at three boundaries, illustrative arithmetic with every assumption stated
  • Nordvik Region and Hansa Mutual: Nordvik Region: uses providers' published figures for its citizen-service assistant without boundaries; the chapter's check shows what they cannot answer. Hansa Mutual: its internal GPU estate gives a facility-level number it has never compared with a provider's per-prompt figure. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Usage modes differ by an order of magnitude in volume and in controllability: Mode 0 Ambient AI is largely invisible to the buyer; Modes 4 and 5 carry the volume; a table of what a buyer can see in each.
  • Observable, expected, speculation: Observable: measured per-query figures from some providers; independent benchmarks of open models. Expected: more standardised disclosure and further efficiency gains per task. Speculation: aggregate demand in 2030 and beyond.
  • Figure: Why two figures differ: the boundaries of a footprint figure, from active accelerators to the whole facility
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Google, 'Measuring the environmental impact of delivering AI at Google Scale' (Elsworth et al., arXiv 2508.15734, Aug 2025): median text prompt 0.24 Wh comprehensive and 0.10 Wh active-accelerator boundary, vendor-reported, May 2025 data [verify; dossier 02 read the abstract only].
  • The 2026 Joule paper on inference energy (median about 0.31 Wh, IQR 0.16-0.60; boundary and authorship unconfirmed) [verify; page not read].
  • IEA, Energy and AI (2025): data centres just over 1% of world electricity in 2024 and about 945 TWh by 2030 in the central case [verify; dossier 05 saw a secondary reproduction by Carbon Brief].
  • Mistral/Carbone 4 lifecycle assessment (2025) as the example of a lifecycle boundary [verify; dossier 02 read only the method page, numbers are from snippets].
  • Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 3: The Benefit Side: What AI Can Do for the Environment and How Much Is Shown

  • File: 03_the_benefit_side.md
  • Status: draft
  • Purpose: Describe the benefit side with the same discipline as the footprint side: what AI is used for to the benefit of the environment (energy systems and grids, buildings, industry, transport, agriculture, monitoring, forecasting, the organisation's own operations), how much of the benefit is shown by measurement in deployment versus modelled, projected or asserted, and what a critical reading finds in the corporate claims. The chapter neither inflates nor dismisses the benefit; it grades it.
  • Builds on: AI Governance, Chapter 22; AI Platform and Security, Chapter 22; AI Leader, Chapter 20
  • Environmental functions: E2 Justify
  • Sections: A claim in a sustainability report; Where AI helps; The evidence grades; What the IEA and others model; What critics found; Handprint, if a source uses the word; Benefit within reach; A fair summary
  • Practical guides: Grading a benefit claim: five questions and the grade they give
  • Worked examples: Three benefit claims graded, one per grade, with the evidence each would need to move up
  • Nordvik Region and Hansa Mutual: Nordvik Region: an AI-assisted route optimisation for regional public transport and an energy-use forecast for its buildings are candidate benefits; neither has a baseline yet. Hansa Mutual: claims-photo assessment that reduces site visits, and a catastrophe-risk model used in climate adaptation pricing, are candidates; the benefit claims are in the sustainability report without a boundary. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Benefit also differs by mode: ambient features rarely produce a measurable environmental benefit; built systems (Mode 4) and controlled agents (Mode 5) in operations can, and carry a footprint of their own.
  • Observable, expected, speculation: Observable: deployed AI for forecasting and control in energy and industry with reported savings. Expected: more applications with measurement as reporting rules mature. Speculation: the size of AI's net contribution to climate goals by 2035.
  • Figure: The evidence grades for a benefit claim, from asserted to measured in deployment
  • Planned length: 4,500-6,000 words of prose
  • Evidence to verify:
  • IEA, Energy and AI (2025), chapter on AI and climate: about 5% of energy-related emissions in 2035 under widespread adoption of existing applications, as quoted in search summaries; baseline not seen [verify against the primary text; the page returned 403 to the dossier agent].
  • Beyond Fossil Fuels / Friends of the Earth / Stand.earth, 'AI Climate Hoax' (Joshi, Feb 2026): 154 company claims tested [verify; PDF unreadable to the dossier agent].
  • Rolnick et al., 'Tackling Climate Change with Machine Learning' (ACM Computing Surveys, 2022) as the standard catalogue of possible uses [verify].
  • Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 4: The Method: Five Functions, the Optimisation Hierarchy and the Net-Benefit Principle

  • File: 04_the_method.md
  • Status: draft
  • Purpose: Define the method of the book: the five environmental functions (E1 Measure, E2 Justify, E3 Design, E4 Operate, E5 Account), the optimisation hierarchy (necessity, sufficiency of model, efficient serving, place and time, hardware and lifecycle), the net-benefit principle with its three tests (Count, Compare, Cap), the environmental budget with a stop path outside the system it stops, the rule that neither side of the ledger pays for the other, and the book's artefacts (footprint, benefit, budget, placement, supplier disclosure, claim and hardware lifecycle records, and the environmental register). State the delegation ceiling for agents that can spend resources.
  • Builds on: AI Operations, Chapter 4; AI Operations, Chapter 17; AI Purple, Chapter 15; AI Leader, Chapter 4; AI Developer, Chapter 4; AI Governance, Chapter 2; AI Governance, Chapter 22
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: From principle to practice; The five functions; The net-benefit principle; The optimisation hierarchy; The use is the unit; The environmental budget and the stop path; The delegation ceiling; The artefacts; Claims need evidence, and evidence expires
  • Practical guides: The method on one page: the three tests as a checklist for any proposed, scaled or renewed use
  • Worked examples: A minimal environmental register with two rows, one per organisation, illustrative
  • Nordvik Region and Hansa Mutual: Nordvik Region and Hansa Mutual are each given their starting register row and the question each function will answer in Parts II to VI. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Defines the table used throughout: degrees of delegation per class of environmental action (measure, recommend, schedule, provision, scale, stop, publish) with the usage modes concerned.
  • Observable, expected, speculation: The principle and hierarchy are the book's own conclusions; statements about the future are labelled as in AI Governance, Chapter 22; the author's projections are kept to Chapters 25 and 29.
  • Figure: The optimisation hierarchy as five ordered levers and the three tests of the net-benefit principle
  • Planned length: 5,000-6,500 words of prose
  • Evidence to verify:
  • Software Carbon Intensity, ISO/IEC 21031:2024, as the form of a rate per functional unit [verify; dossier 02 cites the Green Software Foundation page]. Life cycle assessment ISO 14040/44 for system boundary and functional unit [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part II: Measuring

Chapter 5: The Footprint of a Model: Training, Inference and Embodied Impact

  • File: 05_the_footprint_of_a_model.md
  • Status: draft
  • Purpose: Show how to count the footprint of a model: training, fine-tuning, inference over its service life and embodied impact of the hardware, with the allocation choices (amortisation period, share of training attributed to a use) made explicit. Explain what providers disclose in model cards and environmental reports, what is missing, and how a buyer who does not train a model should use the figures.
  • Builds on: AI Platform and Security, Chapter 7; AI Platform and Security, Chapter 22; AI Operations, Chapter 11
  • Environmental functions: E1 Measure
  • Sections: A number in a model card; Training; Fine-tuning and adaptation; Inference over the life; Embodied impact; Allocation choices; What providers disclose; What a buyer can do
  • Practical guides: A model footprint sheet: the eight fields a buyer asks a provider to fill
  • Worked examples: Two model cards compared at their stated boundaries, with what cannot be compared said plainly
  • Nordvik Region and Hansa Mutual: Nordvik Region: buys a general-purpose model through a framework and has no training figure; the sheet shows which fields the supplier cannot give. Hansa Mutual: fine-tunes an open-weight classifier and can count its own training run in full. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Training footprint is the provider's and the builder's concern, inference the user's; the table shows who can count what in Modes 1 to 5.
  • Observable, expected, speculation: Observable: provider disclosures of training figures for some models. Expected: disclosure of energy for general-purpose models under the AI Act's documentation duties [verify the article and annex, Chapter 21]. Speculation: future training scale.
  • Figure: The life of a model: training, fine-tuning, inference and hardware, with the footprint counted at each stage
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • Meta Llama 3.1 405B model card: about 8,930 tCO2e location-based and 0 market-based for training, electricity only [verify; dossier 02 saw search summaries only].
  • NVIDIA product carbon footprint for H100-class cards (164 kgCO2e per card, as used in a modelling paper) [verify; low confidence].
  • Regulation (EU) 2024/1689 documentation duties for general-purpose AI models, including energy consumption, Annex XI [verify with fetch_eu_law; Article numbers as amended by Regulation (EU) 2026/1744]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 6: The Footprint of a Use: Per Query, Per Task and Per Agent

  • File: 06_the_footprint_of_a_use.md
  • Status: draft
  • Purpose: Show how to count the footprint of a use: per query, per task and per agent run. Explain the multipliers that make a per-query figure misleading (reasoning, context length, retrieval, output length, modality, retries, tool calls, background work), the difference between a median and a mean, and the boundaries of provider per-prompt disclosures. Give the reader the method to estimate the footprint of their own use at a stated boundary and to say how uncertain it is.
  • Builds on: AI Operations, Chapter 13; AI Operations, Chapter 15; AI Platform and Security, Chapter 11
  • Environmental functions: E1 Measure
  • Sections: The unit of decision; What provider figures say; The multipliers; The agent run; Estimating your own use; Distribution and averages; What not to do; The footprint record FPR
  • Practical guides: Estimating a use's footprint in five steps with a stated boundary and an uncertainty range
  • Worked examples: Illustrative: the citizen-service assistant's monthly footprint as range, with the three assumptions that dominate it
  • Nordvik Region and Hansa Mutual: Nordvik Region: the citizen-service assistant on reserved accelerators in a Nordic data centre is counted at facility boundary with its hosting provider's grid and PUE figures; FPR-NR-0001 illustrative. Hansa Mutual: the claims-triage agent run is counted per claim, with the retrieval and retry multipliers; FPR-HM-0001 illustrative. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Footprint per use differs by mode: Mode 1 chat versus Mode 3 agentic work versus Mode 5 agents that start their own work; the table gives the multiplier ranges the writer can verify.
  • Observable, expected, speculation: Observable: independent benchmarks showing reasoning multiplying energy by tens of times for open models. Expected: standard per-task disclosure. Speculation: per-task cost trends for agents.
  • Figure: From a single prompt to an agent run: where the multipliers enter a use's footprint
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Hugging Face AI Energy Score v2 (4 Dec 2025): reasoning on versus off, 154x, 514x, 697x for three open models, GPU energy only; about 30x on average [verify; dossier 06 read the blog text, dossier 02 saw summaries].
  • Chung et al., ML.ENERGY, 'Where Do the Joules Go?' arXiv 2601.22076 (Jan 2026) [verify; PDF not decodable to the dossier agent].
  • Agent-level preprints: arXiv 2605.22883 (goal-level energy accounting) and arXiv 2611.12128 (per-task carbon), both preprints [verify; unread].
  • Google (Elsworth et al., 2025) and the Joule paper (2026) as the two per-prompt boundaries to contrast [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 7: Water, Materials and Land: Beyond Carbon

  • File: 07_water_materials_and_land.md
  • Status: draft
  • Purpose: Cover what carbon leaves out: water (withdrawal versus consumption, on-site versus off-site, local stress), materials (embodied minerals, hardware manufacture, e-waste), land and local effects (noise, heat, grid connection, diesel backup, competition for power and water). Describe local conflicts over data centres by what was reported and by whom, and show how an enterprise or authority counts water and materials for an AI use without false precision.
  • Builds on: AI Platform and Security, Chapter 22; AI Operations, Chapter 11
  • Environmental functions: E1 Measure
  • Sections: A litre is not a litre; What providers report; What regulators and the Commission report; Materials and e-waste; Land and local effects; Counting water for a use; What to ask a supplier; Limits of the evidence
  • Practical guides: A water and materials disclosure checklist for a supplier, mapped to the SDR record
  • Worked examples: Illustrative: the same assistant counted for on-site and total water in a Nordic and a continental region
  • Nordvik Region and Hansa Mutual: Nordvik Region: a Nordic hosting region with cool climate and low water stress, and a resident debate about a regional data-centre plan, reported by the local press and the municipality. Hansa Mutual: a mix of regions including a continental one with stated water stress; its Scope 3 figure does not yet carry water. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Water and materials differ little by mode at the level of a buyer; they differ by where and how capacity is hosted; the table maps the mode to the party who can count.
  • Observable, expected, speculation: Observable: provider water reports, rising consumption at some providers, local conflicts in Europe. Expected: facility-level water reporting under the EU data-centre scheme. Speculation: water becoming binding in permitting at scale.
  • Figure: Withdrawal and consumption, on-site and off-site: the water boundaries of an AI use
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • Google 2026 Environmental Report: 10.9 billion gallons (41 million m3) of water consumption in 2025, up 34%, replenishment about 78%; median text prompt 0.26 mL (May 2025 method) [verify; dossier 03 read it].
  • COM(2026) 500 final, Commission report under Article 12(5) of the Energy Efficiency Directive, 21 Sept 2026: first analysis of the EU data-centre database (770 data centres, about 36% of the estimated population), a common rating scheme (Commission Delegated Regulation C(2026) 3472 final of 21 September 2026, adopted by the Commission, not yet in force; the number 2026/7000 in the report is a placeholder) and a move towards minimum performance standards on PUE, WUE and REF.
  • EU average WUE of 0.58 m3/MWh is a 2024 reporting-year value [verify; dossier 03].
  • Spain (Madrid draft rules, Oct 2026) and Germany (EnEfG heat-reuse duties from 1 July 2026) were snippet-only items [verify before any use]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 8: Measuring in Practice: Telemetry, Standards, Supplier Figures and Evidence Age

  • File: 08_measuring_in_practice.md
  • Status: draft
  • Purpose: Show how to measure in practice. Where the data comes from (cloud carbon tools, provider reports, telemetry, tagging, billing as a proxy, supplier disclosures), the standards and methods available (SCI and SCI for AI, ISO life cycle assessment, GHG Protocol, ISO/IEC 30134 data-centre KPIs), their status and limits, and how to attach evidence age to every figure. End with the footprint record and the supplier disclosure record in use, and with what to do when a supplier cannot give a figure.
  • Builds on: AI Operations, Chapter 11; AI Operations, Chapter 20; AI Platform and Security, Chapter 21; AI Purple, Chapter 15
  • Environmental functions: E1 Measure
  • Sections: Measurement is a system; Direct measurement; Provider tools and reports; Standards and methods; Estimating when you cannot measure; Evidence age; When the supplier cannot say; The records
  • Practical guides: A measurement plan for one AI use in one page: boundary, data source, owner, evidence age, check
  • Worked examples: The footprint record for one use, filled with the evidence-age field and the independent check
  • Nordvik Region and Hansa Mutual: Nordvik Region: relies on supplier figures through a framework; its SDR-NR-0001 lists what the supplier gave and what it did not. Hansa Mutual: instruments its own GPU estate and compares cloud carbon-tool output with its own meter readings; FPR-HM-0002 illustrative. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: What can be measured by whom across Modes 0 to 5, and which degree of delegation applies to an agent that collects or reports measurements (not Level 4 for publishing).
  • Observable, expected, speculation: Observable: SCI is an ISO/IEC standard; GHG Protocol guidance on AI called inconsistent in 2026 working-group material. Expected: standards under the AI Act and a Scope 3 revision. Speculation: a single accepted per-task metric.
  • Figure: A measurement system: from meter and tag to footprint record, with evidence age and an independent check
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Green Software Foundation, SCI and SCI for AI [verify; ratification date of SCI for AI inconsistent across sources; ISO submission listed Q2 2026].
  • CEN-CENELEC, 23 Oct 2025: prioritised AI Act deliverables by Q4 2026; request M/593 as amended by M/613, expiry 28 Feb 2027; whether an energy deliverable has been published is unconfirmed [verify].
  • GHG Protocol Scope 3 working-group presentation, 30 Apr 2026 [verify].
  • ISO/IEC TR 20226 (2025) status unchecked; AFNOR SPEC 2314 and ITU-T L.1801 not verified at all [verify or drop]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part III: Justifying

Chapter 9: Is This Use Worth It? Benefit per Footprint, Baselines and the Non-AI Alternative

  • File: 09_is_this_use_worth_it.md
  • Status: draft
  • Purpose: Teach the second test, Compare, in full. Define a benefit per footprint, set a baseline, compare with the non-AI alternative (including doing nothing and a simpler method), and write the benefit record. Show how to decide whether a use is worth its footprint, when it is not worth measuring precisely, and how benefits other than environmental ones (service quality, cost, safety) enter without being used to hide a footprint.
  • Builds on: AI Governance, Chapter 8; AI Governance, Chapter 22; AI Leader, Chapter 12; AI Developer, Chapter 20
  • Environmental functions: E2 Justify
  • Sections: A pilot that worked; Compare means three comparisons; Benefit per footprint; Baselines; The non-AI alternative; A decision rule; Other benefits; The benefit record BEN
  • Practical guides: A benefit case in one page: baseline, alternative, measure, owner, review date, decision
  • Worked examples: Illustrative: two uses compared with their non-AI alternatives, one accepted and one declined
  • Nordvik Region and Hansa Mutual: Nordvik Region: a route-optimisation assistant against a spreadsheet method; BEN-NR-0001. Hansa Mutual: claims-photo assessment against site visits, with the visits avoided counted; BEN-HM-0001. Both illustrative. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: what an AI system may do in preparing a benefit case (draft, recommend), who decides (a named owner), and why a system never approves its own benefit case.
  • Observable, expected, speculation: Observable: few deployments publish benefit with baseline. Expected: benefit records becoming part of governance. Speculation: standard benefit metrics across domains.
  • Figure: A benefit case: baseline, non-AI alternative and rebound compared with the footprint
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • IEA and other scenario studies as examples of modelled benefit without deployment baselines [verify]; the 'AI Climate Hoax' report's grading of claim types [verify].
  • Evidence from deployed case studies of AI in energy and buildings with a stated baseline (to be found by the writer; none is named in the dossiers). Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 10: Rebound and Growth: Induced Demand and the Limits of Efficiency

  • File: 10_rebound_and_growth.md
  • Status: draft
  • Purpose: Treat rebound and growth: induced demand, the Jevons paradox in AI inference, efficiency versus volume, and what has been measured. State what is known from energy economics, what is measured for AI (little), the difference between per-unit and total, and the organisation's own rebound (more uses, longer outputs, more agents). Give a practical way to measure rebound at organisation level.
  • Builds on: AI Operations, Chapter 15; AI Platform and Security, Chapter 22; AI Governance, Chapter 22
  • Environmental functions: E2 Justify
  • Sections: A halved cost, a doubled volume; Kinds of rebound; What is known for AI; Inside the organisation; Measuring rebound; Responding; Limits of the claim; What follows for the benefit case
  • Practical guides: A rebound check: three series (units, volume, total) for each use, reviewed at renewal
  • Worked examples: Illustrative: a model swap that halves per-query footprint and doubles volume, with the net result
  • Nordvik Region and Hansa Mutual: Nordvik Region: its citizen-service assistant doubles in volume after a campaign; per-query efficiency gains do not offset it. Hansa Mutual: a cheaper claims model leads handlers to use it for new tasks; the rebound is measured by the register. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Rebound differs by mode: ambient features switch on for all users; agents multiply work without a person asking; the table says who can cap each.
  • Observable, expected, speculation: Observable: aggregate data-centre demand rising alongside falling per-task figures, as reported by the IEA [verify]. Expected: continued growth. Speculation: whether efficiency gains will keep pace. No measured total-inference rebound is claimed.
  • Figure: Per-unit footprint, volume and total footprint over time: how rebound shows in the three lines
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • IEA update (Apr 2026) reportedly showing per-task power falling while data-centre demand rises [verify; secondary report only in dossier 05].
  • Jevons-in-inference preprints (Research Square, SSRN, arXiv, 2026), none measuring total inference energy directly [verify].
  • A Nature-family paper (Sep 2026) modelling AI data centres at about 1% of global electricity demand by 2030 [verify venue and figures; dossier 05 could not confirm the venue]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 11: Crisis or Manageable Cost? A Debate

  • File: 11_crisis_or_manageable_cost.md
  • Status: draft
  • Purpose: Stage the debate: is AI's environmental cost a crisis or a manageable cost? State the case for concern and the case for calm each at its strongest, separate what a source reports from interpretation, show where the positions meet, and end with what an organisation can do now. The verdict, labelled observable, expected or speculation, concerns the organisation's decisions and not the future of the planet.
  • Builds on: AI Governance, Chapter 22; AI Leader, Chapter 25; AI Operations, Chapter 15; AI Platform and Security, Chapter 22
  • Environmental functions: E2 Justify
  • Sections: Two readings of the same numbers; The case for concern; The case for calm; Where the positions meet; What can be done; What would change the verdict; Verdict
  • Constructive debate: Concern: Aggregate demand is rising fast and is concentrated in specific grids and basins; disclosure is thin and mixed in boundary; rebound is unmeasured; hardware and water impacts are underreported; local conflicts are real; an unproven benefit does not offset any of it. Calm: AI is a small share of global electricity and emissions; per-task efficiency has improved by large factors in provider reports; most of the electricity growth in the IEA scenarios comes from other sources; real benefits exist in energy systems; disclosure is improving; the right response is measurement and budgets, not alarm. Where they meet: Both sides agree that disclosure is poor, that boundaries matter, that rebound decides the net result, that local effects are real and that benefit must be shown, not assumed. What can be done: Count, Compare and Cap per use; put disclosure in contracts; place and time workloads; decline low-value uses; keep a register; revisit the verdict when the markers move.
  • Practical guides: A watch list of five markers that would move the organisation from calm to concern, or back
  • Worked examples: Illustrative: the board of each organisation reads the same paper and reaches different first questions
  • Nordvik Region and Hansa Mutual: Nordvik Region: the council debate follows a regional data-centre plan; the authority's answer is a procurement requirement. Hansa Mutual: the board reads the group's reporting under the CSRD regime and asks what AI adds to Scope 2 and 3. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Final table: for each position, which delegation levels it would accept per class of environmental action, and what each position says about Mode 5.
  • Observable, expected, speculation: The verdict separates what is observable, what is expected and what remains speculation; the author's own projections are reserved for Chapters 25 and 29.
  • Figure: The debate in four steps: concern, calm, where the positions meet and what can be done
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Carbon Brief 'AI: Five charts that put data-centre energy use and emissions into context' (15 Sep 2025, modified 15 Jul 2026) reproducing IEA figures [verify against the IEA text].
  • Beyond Fossil Fuels 'Data centres and AI at COP31' and Food & Water Europe 'The Urgent Case Against Data Centres in Europe' (Oct 2026) as advocacy positions, labelled as such [verify; unread by the dossier agent].
  • A RETN survey (Nov 2026, vendor-commissioned) on public attitudes [verify; label as vendor-commissioned or drop]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part IV: Designing for Less

Chapter 12: Choosing the Right Model: Right-Sizing, Small Models, Routing and Cascades

  • File: 12_choosing_the_right_model.md
  • Status: draft
  • Purpose: Teach model choice as the second lever of the hierarchy: sufficiency of model. Right-sizing, small and task-specific models, routing and cascades, open-weight versus proprietary, and the evidence on how much each saves at what quality cost. Show that parameter count is not a reliable proxy for energy, how to choose by task benchmark and measured energy, and how to record a model choice with its requirement.
  • Builds on: AI Platform and Security, Chapter 10; AI Platform and Security, Chapter 7; AI Operations, Chapter 13; AI Developer, Chapter 19
  • Environmental functions: E3 Design
  • Sections: The biggest model by default; Writing the requirement; Small and task-specific models; Routing and cascades; Open-weight and proprietary; Size is not a proxy; Reasoning effort as a model choice; The model-choice record
  • Practical guides: A model-choice protocol: requirement, three candidates, quality test, energy test, decision
  • Worked examples: Illustrative: correspondence routing at Nordvik Region with a fine-tuned small classifier versus a general model
  • Nordvik Region and Hansa Mutual: Nordvik Region: the correspondence-routing classifier (a fine-tuned open-weight classifier) is the small-model case. Hansa Mutual: routing claims enquiries between a small model and a larger one, with an escalation rate recorded. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Model choice by mode: Mode 1 users choose by default; Mode 4 builders choose by requirement; Mode 5 agents may route among models only within a mandate that fixes the allowed set.
  • Observable, expected, speculation: Observable: small models meet many task requirements in benchmarks. Expected: routers become standard in provider products. Speculation: whether small models close the quality gap on complex tasks.
  • Figure: A cascade: small model first, escalation on a stated rule, and the escalation rate as the footprint driver
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • UNESCO/UCL 'Smarter, Smaller, Stronger' (8 Jul 2025, page updated 10 Apr 2026): up to 90% less energy with small task-specific models, shorter prompts over 50%, compression up to 44%; baselines undefined [verify the report, not the press page; dossier 06].
  • Hugging Face AI Energy Score v2 on size not being a reliable proxy [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 13: Efficient Inference: Quantisation, Caching, Batching and Context Discipline

  • File: 13_efficient_inference.md
  • Status: draft
  • Purpose: Cover efficient inference as the third lever: quantisation, caching, batching, speculative and other serving techniques, context discipline (shorter prompts and outputs, retrieval of less) and reasoning effort. For each technique give the measured saving, its baseline, its quality trade-off and whether a buyer or a supplier controls it. Show how to turn a lever into a requirement in a contract or a configuration.
  • Builds on: AI Platform and Security, Chapter 10; AI Platform and Security, Chapter 21; AI Operations, Chapter 13
  • Environmental functions: E3 Design
  • Sections: Where the energy goes in a request; Reasoning effort and thinking budget; Context discipline; Quantisation and compression; Caching and batching; Speculative decoding and other techniques; From lever to requirement; Testing quality after the lever
  • Practical guides: An inference efficiency checklist, ordered by saving and by who controls the lever
  • Worked examples: Illustrative: effort, context and caching changes on one assistant, with measured effect and quality check
  • Nordvik Region and Hansa Mutual: Nordvik Region: the citizen-service assistant gets an output-length limit and a cache for frequently asked questions; its supplier controls quantisation. Hansa Mutual: claims handlers' assistant defaults reasoning effort to low and escalates on request. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Levers by mode and delegation: whether an agent may change its own effort, context or model, which it may not do beyond its mandate.
  • Observable, expected, speculation: Observable: measured effects of reasoning on and off in benchmarks. Expected: effort controls exposed by more providers. Speculation: the size of future serving gains.
  • Figure: Levers on the path of a request: effort, context, cache, batch and model, with the party that controls each
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • Hugging Face AI Energy Score v2 (Dec 2025): reasoning on versus off and gpt-oss low versus high effort [verify; GPU energy only, quality not measured in the post].
  • Microsoft Cloud blog 15 Jun 2026 and the Joule paper: projected 8x to 20x per-query reduction, labelled a modelling estimate, with its own worked example near 2.3x [verify; dossier 06].
  • ML.ENERGY leaderboard and the benchmark paper (Chung et al., arXiv 2505.06371) [verify; unread]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 14: Training, Fine-Tuning and Retrieval: When Not to Train

  • File: 14_training_fine_tuning_and_retrieval.md
  • Status: draft
  • Purpose: Answer the question of when not to train. Compare pre-training, fine-tuning, parameter-efficient adaptation, retrieval-augmented generation, prompting and no AI, by footprint and by fit. Show the footprint of retrieval infrastructure (indexing, embeddings, storage, queries), when fine-tuning pays back against inference volume, and how to record the decision.
  • Builds on: AI Platform and Security, Chapter 7; AI Platform and Security, Chapter 10; AI Governance, Chapter 5
  • Environmental functions: E3 Design
  • Sections: A training run nobody needed; The ladder of approaches; Retrieval's footprint; When fine-tuning pays; Governing experiments; Reuse and sharing; Data and its footprint; Decision record
  • Practical guides: A when-not-to-train checklist: five questions before any training or fine-tuning is approved
  • Worked examples: Illustrative: the correspondence classifier trained once and the retrieval index refreshed monthly, with footprints compared
  • Nordvik Region and Hansa Mutual: Nordvik Region: the casework retrieval assistant and its index refresh cadence; the fine-tuned correspondence classifier (AIDR-0042 in the profile) as the training case. Hansa Mutual: its fine-tuned classifier and the retrieval over policy wordings; AIDR-1350 to AIDR-1399 reserved for its environmental decisions. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: an agent that launches training or fine-tuning jobs is capped by a budget set by someone else; Level 4 never for provisioning compute beyond it.
  • Observable, expected, speculation: Observable: retrieval and adapters widely used instead of training. Expected: more reuse of shared adapters. Speculation: the future share of training in total AI energy.
  • Figure: A ladder of approaches from no AI to training, with the footprint rising at each step
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • Elsworth et al. 2025 on serving versus training boundaries, and provider model cards for training figures [verify].
  • Evidence on the footprint of retrieval and embedding infrastructure: none in the dossiers; the writer finds and verifies a primary source or states that none was found. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 15: Agents and Resource Budgets: Loops, Tool Calls and Background Work

  • File: 15_agents_and_resource_budgets.md
  • Status: draft
  • Purpose: Treat agents: loops, tool calls and background work. An agent that can start work can spend resources, so it carries a resource budget set by someone else, measured outside the agent, with a stop path outside the system it stops. Cover how agents multiply footprint (loops, retries, parallel calls, scheduled and event-driven work), the budget dimensions (tokens, tool calls, wall-clock, cost, energy where measurable), the mandate manifest with a resource section, and the delegation ceiling.
  • Builds on: AI Platform and Security, Chapter 11; AI Platform and Security, Chapter 19; AI Operations, Chapter 17; AI Purple, Chapter 15; AI Developer, Chapter 9; AI Governance, Chapter 13
  • Environmental functions: E3 Design; E4 Operate
  • Sections: The agent that kept going; Where agents spend; What has been measured; The budget; The stop path; The mandate manifest; The delegation ceiling; Coding agents
  • Practical guides: A resource budget for an agent in one page: dimensions, limits, measurement point, stop path, owner, test
  • Worked examples: Illustrative: the claims-triage agent's mandate with a resource section, and one stop path test
  • Nordvik Region and Hansa Mutual: Nordvik Region: no Mode 5 agents; agentic work tools (Mode 3) limited to a small platform team, whose compute use is budgeted; MND-NR-ENV-0001. Hansa Mutual: the claims-triage agent (Mode 5) with a per-claim budget and a stop path in the platform, outside the agent; MND-HM-ENV-0001. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Table of the degrees of delegation (Level 0 Advisory to Level 4 Autonomous) for what an AI system may do in this chapter's subject, with the usage modes concerned (Mode 0 Ambient AI to Mode 5 Autonomous production agents) and the Level 4 prohibitions that apply: no Level 4 for an agent that provisions or scales its own compute, or starts work, beyond a budget set by someone else; for an action that disables the measurement, the budget or the stop path; or for publishing an environmental claim. The table is the chapter's centrepiece.
  • Observable, expected, speculation: Observable: first measurements of agent task energy in preprints; vendor tools for budgets. Expected: budgets in agent platforms by default. Speculation: agent-to-agent economies and their footprint.
  • Figure: An agent with a budget: spend measured outside the agent and a stop path outside the system it stops
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • arXiv 2605.22883 (v1 20 May 2026, v2 12 Oct 2026) and arXiv 2611.12128 (10 Nov 2026) [verify; preprints, unread by the dossier agent].
  • Series foundations: observer principle (AI Operations, Chapter 17), independence principle (AI Purple, Chapter 15), agents that exceed their mandate (AI Platform and Security, Chapter 19). Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 16: Hardware and Lifecycle: Utilisation, Embodied Impact, Reuse and Retirement

  • File: 16_hardware_and_lifecycle.md
  • Status: draft
  • Purpose: Cover hardware and lifecycle as the fifth lever: utilisation, embodied impact, procurement of accelerators, extending life, reuse, refurbishment, take-back and retirement, and the evidence on embodied carbon and minerals. Show how an enterprise or authority, which mostly buys capacity, uses hardware facts in contracts, and how an organisation with its own GPUs manages the lifecycle. Introduce the hardware lifecycle record HWL.
  • Builds on: AI Platform and Security, Chapter 13; AI Platform and Security, Chapter 14; AI Operations, Chapter 11
  • Environmental functions: E3 Design
  • Sections: A rack at the end of its contract; Embodied impact; Utilisation; Extending life; Retirement and take-back; Buying capacity; Efficiency of new generations; The hardware lifecycle record HWL
  • Practical guides: A hardware lifecycle checklist for buyers of capacity and for owners of GPUs
  • Worked examples: Illustrative: Hansa Mutual's refresh decision: replace or re-tier a GPU cluster, with embodied and operational effects
  • Nordvik Region and Hansa Mutual: Nordvik Region: buys reserved accelerators from a hosting provider and has no hardware of its own; HWL-NR-0001 is a clause set, not an inventory. Hansa Mutual: owns a GPU cluster and faces a refresh decision; HWL-HM-0001. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Hardware decisions by delegation: an agent may forecast capacity and recommend; provisioning decisions beyond budget stay with named owners.
  • Observable, expected, speculation: Observable: vendor product carbon footprints and rising e-waste concerns. Expected: ecodesign and take-back rules for servers and data-centre equipment [verify in Chapter 21]. Speculation: lifetimes of accelerators.
  • Figure: Hardware over its life: manufacture, utilisation, second life and retirement, with decision points
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • NVIDIA H100-class product carbon footprint 164 kgCO2e per card as used in a modelling paper [verify; low confidence].
  • Mistral/Carbone 4 LCA: hardware about 11% of GHG, 5% of water and 61% of mineral depletion, as reported by Projet Celsius [verify in the Mistral report; secondary in the dossier].
  • Gaps flagged in dossier 03: EU Critical Raw Materials Act, chip-fab energy, WEEE statistics, liquid cooling [find primary sources or state the gap]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part V: Operating

Chapter 17: Where and When: Grid Carbon Intensity, Water Stress and Carbon-Aware Scheduling

  • File: 17_where_and_when.md
  • Status: draft
  • Purpose: Cover place and time: grid carbon intensity, water stress, carbon-aware scheduling and the Nordic case. Explain location- and market-based accounting as it affects the choice of region, hourly and marginal intensity, which workloads can be shifted (training, batch, indexing) and which cannot (interactive), latency and data-residency constraints, and the evidence on how much shifting saves. Write the placement policy PLC.
  • Builds on: AI Platform and Security, Chapter 13; AI Platform and Security, Chapter 14; AI Operations, Chapter 11; AI Governance, Chapter 6
  • Environmental functions: E4 Operate
  • Sections: Same workload, two grids; Grid carbon intensity; Water stress and place; Shifting in time; Shifting in space; What the savings are; Market-based claims and the physical grid; The placement policy PLC
  • Practical guides: A placement policy in one page: which workloads, which regions, which hours, what exceptions
  • Worked examples: Illustrative: shifting a nightly indexing job and a weekly evaluation run, with the estimated saving and its range
  • Nordvik Region and Hansa Mutual: Nordvik Region: hosted in a Nordic data centre with low grid intensity, with sovereignty and procurement requirements binding the region; PLC-NR-0001. Hansa Mutual: a mix of German, Dutch and Polish capacity with very different grids; PLC-HM-0001 routes batch work by hour and region. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: an agent may schedule within a policy, never change the policy or move data across residency limits; the stop path for a scheduler is outside it.
  • Observable, expected, speculation: Observable: large differences in grid intensity between and within countries. Expected: more granular, hourly data. Speculation: how far shifting scales before grids change.
  • Figure: Placement by place and time: region, hour, carbon intensity and water stress as inputs to a placement rule
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • Grid data sources and their status (European and national system operators, ENTSO-E, electricity maps) [find and verify; no source in the dossiers; dossier 06 flags Nordic grids as thinly read].
  • GHG Protocol Scope 2 guidance on location- and market-based methods [verify].
  • Sweden as the worked national case: national grid intensity and data-centre policy [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 18: Data Centres in Europe: Power, Cooling, Heat Reuse and Grid Connection

  • File: 18_data_centres_in_europe.md
  • Status: draft
  • Purpose: Describe data centres in Europe as a buyer and operator's concern: power, cooling, heat reuse, grid connection, efficiency metrics (PUE, WUE, REF, ISO/IEC 30134), the EU reporting and rating scheme, national rules (Germany, Ireland, the Netherlands, France, Sweden), and local conflicts as reported. Not a guide to building data centres, but to what a buyer or operator should know and ask.
  • Builds on: AI Platform and Security, Chapter 13; AI Platform and Security, Chapter 22; AI Operations, Chapter 11
  • Environmental functions: E4 Operate
  • Sections: A data centre in the news; Power and grid connection; Cooling and water; Efficiency metrics; Heat reuse; The EU scheme; Local effects and conflicts; What a buyer or operator asks
  • Practical guides: Ten questions to a data-centre provider and what a good answer contains
  • Worked examples: Illustrative: the same assistant hosted in a facility with and without heat reuse, with what is counted and what is not
  • Nordvik Region and Hansa Mutual: Nordvik Region: the regional context includes municipal district heating and a local debate; its hosting provider's heat-reuse answer is evaluated; no facility of its own. Hansa Mutual: a colocation facility in Germany subject to the heat-reuse duty from July 2026 and its own server room being decommissioned. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: building-management agents and cooling controllers are Mode 5 systems in the facility; Level 4 limits and independent stop paths apply.
  • Observable, expected, speculation: Observable: EU reporting database and rating scheme adopted. Expected: minimum performance standards and a Cloud and AI Development Act proposal [verify COM(2026) 502]. Speculation: whether permitting becomes the binding constraint.
  • Figure: A data centre's resource flows: power in, heat out, water and grid, with the points a buyer can ask about
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • COM(2026) 500, 501 and 502 final (21 Sept 2026): EED report, roadmap for digitalisation and AI in energy, and proposal for a Cloud and AI Development Act [verify identifiers and status].
  • Commission Delegated Regulation C(2026) 3472 final of 21 September 2026 on the rating scheme (not yet published in the Official Journal; the number 2026/7000 is a placeholder); the earlier dossiers conflicted (one source calls it a proposal; one cites C(2026) 3472) [verify in the Official Journal before any statement].
  • Germany EnEfG: heat-reuse duties for new data centres from 1 July 2026 and a cabinet amendment draft [verify the law's current text].
  • EU average WUE 0.58 m3/MWh (2024 reporting year) [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 19: Cloud, Sovereign or On-Premises: Choosing on Environmental Evidence. A Debate

  • File: 19_cloud_sovereign_or_on_premises.md
  • Status: draft
  • Purpose: Stage the debate: cloud, sovereign or on-premises, chosen on environmental evidence. Hyperscale cloud (efficiency, utilisation, renewable procurement, opacity), sovereign and regional cloud (residency, smaller scale, variable disclosure), and on-premises (control, measurable, low utilisation risk). Separate what sources report from interpretation, find where the positions meet, and give a decision method that does not decide the question by label.
  • Builds on: AI Platform and Security, Chapter 14; AI Platform and Security, Chapter 13; AI Governance, Chapter 6; AI Governance, Chapter 7; AI Operations, Chapter 11
  • Environmental functions: E4 Operate
  • Sections: Three options, three claims; The case for hyperscale cloud; The case for sovereign and regional cloud; The case for on-premises; Where the positions meet; Choosing on evidence; What can be done; Verdict
  • Constructive debate: Concern: Hyperscale cloud discloses little at the level a buyer needs, concentrates demand, and its renewable claims are often market-based; sovereign choices may be smaller, less efficient and less transparent; on-premises GPUs idle. Calm: Hyperscale cloud achieves high utilisation and efficiency and publishes per-prompt figures; sovereign and regional providers may sit on clean grids and reuse heat; on-premises gives full measurement when the load is steady. Where they meet: Label does not decide: utilisation, grid, cooling and disclosure do; comparability needs the same boundary; sovereignty and environment can both be written into the contract. What can be done: Ask the same ten questions of each option; require facility-level and market- and location-based figures; pilot measurement; record the decision with its evidence age; revisit at renewal.
  • Practical guides: A comparison sheet: the same ten questions put to three options, with the boundary stated for each answer
  • Worked examples: Illustrative: a three-option comparison for Hansa Mutual's next GPU capacity
  • Nordvik Region and Hansa Mutual: Nordvik Region: the choice is among a regional cloud framework, a hyperscale offering and a sovereign provider, with public procurement and sovereignty constraints; its decision is recorded as AIDR-1300 illustrative. Hansa Mutual: the choice is between its own cluster, hyperscale burst capacity and a regional provider; AIDR-1350 illustrative. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Final table: for each option, who can measure, who can budget, who can stop, by mode.
  • Observable, expected, speculation: Observable: provider disclosures and the EU data-centre scheme. Expected: comparable facility-level data from 2027. Speculation: relative efficiency of the three options in 2030.
  • Figure: Three hosting options compared on the same evidence: utilisation, grid region, disclosure and control
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Provider environmental reports (Google 2026 Environmental Report read in dossier 03) and the EU rating scheme as comparable data [verify].
  • No independent comparison of hyperscale, sovereign and on-premises AI hosting was found in the dossiers; the writer finds one or states that none exists. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 20: Operating Sustainably: GreenOps, Observability and Stop Paths

  • File: 20_operating_sustainably.md
  • Status: draft
  • Purpose: Cover operating sustainably: GreenOps in practice. Observability of footprint and budgets, showback and chargeback, dashboards, alerts, budget enforcement and quotas, stop paths, reviews at renewal, incident handling for a budget breach, and the role of an operations team. Refer back to FinOps and capacity practice in earlier volumes without repeating it, and show how the environmental register becomes an operating instrument.
  • Builds on: AI Operations, Chapter 11; AI Operations, Chapter 15; AI Operations, Chapter 17; AI Operations, Chapter 20; AI Platform and Security, Chapter 21; AI Platform and Security, Chapter 22
  • Environmental functions: E4 Operate
  • Sections: A dashboard nobody opened; The loop; Observability; Budgets and quotas; Stop paths; Breach and incident; Reviews at renewal; GreenOps for small teams
  • Practical guides: A budget breach runbook in one page: detect, stop, assess, restart, review
  • Worked examples: Illustrative: a quota breach on the citizen-service assistant and the stop path exercised
  • Nordvik Region and Hansa Mutual: Nordvik Region: a small platform team and a supplier-managed service; it enforces quotas at the supplier gateway and reviews the register quarterly; EBG-NR-0001. Hansa Mutual: a platform team with showback, budgets per agent and a tested stop path in the orchestration layer; EBG-HM-0001. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Table of the degrees of delegation (Level 0 Advisory to Level 4 Autonomous) for what an AI system may do in this chapter's subject, with the usage modes concerned (Mode 0 Ambient AI to Mode 5 Autonomous production agents) and the Level 4 prohibitions that apply: no Level 4 for an agent that provisions or scales its own compute, or starts work, beyond a budget set by someone else; for an action that disables the measurement, the budget or the stop path; or for publishing an environmental claim. The table covers enforcement, stop and restart per mode.
  • Observable, expected, speculation: Observable: FinOps practice extended to carbon and to AI cost. Expected: budgets and stop paths in agent platforms. Speculation: autonomous optimisers operating under budgets.
  • Figure: The environmental operating loop: measure, budget, observe, stop and review, with the stop path outside the system it stops
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • FinOps Foundation FinOps for AI and Sustainability work [verify; not in dossiers].
  • Series practice: AI Operations, Chapters 11, 17 and 20; AI Platform and Security, Chapter 21. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part VI: Accounting and Governing

Chapter 21: Regulation: What Applies to AI and the Environment in 2026

  • File: 21_regulation.md
  • Status: draft
  • Purpose: Map the regulation that bears on AI and the environment in 2026, with each statement attributed to instrument and article and framed as an architectural and organisational implication, not legal advice. Cover the AI Act (documentation of energy consumption for general-purpose AI models; codes of practice; standardisation requests for energy and resource efficiency; high-risk systems), the Energy Efficiency Directive and its data-centre reporting, the EU rating scheme, CSRD and ESRS and their 2026 state, the Taxonomy, ecodesign, consumer-claims rules (the Green Claims proposal and the Empowering Consumers Directive), and national rules in Sweden and in Germany, Ireland, the Netherlands and France.
  • Builds on: AI Governance, Chapter 3; AI Operations, Chapter 23; AI Leader, Chapter 24; AI Developer, Chapter 24; AI Governance, Chapter 17
  • Environmental functions: E5 Account
  • Sections: A date in the calendar; The AI Act and the environment; Energy Efficiency Directive; CSRD, ESRS and the omnibus changes; Taxonomy and ecodesign; Claims law; National rules; Implications for the architecture
  • Practical guides: A regulatory register for AI and the environment: instrument, article, applicability, owner, evidence, review date
  • Worked examples: Illustrative: the same assistant mapped to the obligations that reach each organisation
  • Nordvik Region and Hansa Mutual: Nordvik Region: a Swedish regional authority; public procurement law, public-access rules and the national transposition of EU rules are the setting; whether CSRD reaches it depends on its form [verify]. Hansa Mutual: EU entities subject to the reporting regime in its 2026 form, plus UK and Swiss entities outside EU law [verify]. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: no AI system decides on compliance or publishes a regulatory report; drafting and monitoring may be delegated to Level 2 or 3 with a named human answerable.
  • Observable, expected, speculation: Observable: the AI Act's text as amended, the EED reporting database. Expected: standards under Article 40 and minimum performance standards. Speculation: the content of future Green Claims and Cloud and AI Development Act rules.
  • Figure: The regulatory map: AI Act, Energy Efficiency Directive, CSRD, ecodesign and claims rules, by who is bound
  • Planned length: 5,000-6,500 words of prose
  • Evidence to verify:
  • Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744: verify every cited article with bookengine run fetch_eu_law get ai-act --article N and fetch_eu_law.py check [verify; the CEN-CENELEC request M/593 as amended by M/613 and Article 40(2) per dossier 02].
  • Directive (EU) 2023/1791 (EED), Article 12 and Delegated Regulation (EU) 2024/1364 [verify]; COM(2026) 500 and 502 [verify].
  • Directive (EU) 2022/2464 (CSRD) and the omnibus amendments; Directive (EU) 2024/825 (Empowering Consumers); Green Claims proposal status [verify each; none read by the dossier agents].
  • Swedish statutes via fetch_eu_law.py sfs get where relevant [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 22: Reporting and Claims: Accounting Methods and How to Avoid Greenwashing

  • File: 22_reporting_and_claims.md
  • Status: draft
  • Purpose: Show how to report and claim without greenwashing. Cover accounting methods (location-based and market-based, hourly matching, power purchase agreements and certificates, allocation of AI in Scope 2 and Scope 3 Category 1), how to disclose AI in a sustainability report, what a defensible environmental claim looks like, and the claim record CLM. State what a claim must contain (boundary, year, baseline, method, independent check) and what no claim may do (offset an unproven benefit against a measured footprint).
  • Builds on: AI Governance, Chapter 11; AI Leader, Chapter 12; AI Leader, Chapter 13; AI Operations, Chapter 19
  • Environmental functions: E5 Account
  • Sections: A claim that was true and misleading; Location-based and market-based; Hourly and regional matching; AI in Scope 2 and 3; Claims; Greenwashing patterns; Who may approve a claim; The claim record CLM
  • Practical guides: A claim review checklist: seven questions every environmental claim must answer before publication
  • Worked examples: Illustrative: a claim in Hansa Mutual's annual report rewritten from asserted to defensible; the claim Nordvik Region declines to make
  • Nordvik Region and Hansa Mutual: Nordvik Region: the authority's annual report carries a statement about its assistant; CLM-NR-0001 records the boundary it can defend and the claim it withdraws. Hansa Mutual: its sustainability report counts AI in Scope 2 and 3 and CLM-HM-0001 records the method; the board approves. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: drafting and checking of claims may use AI at Level 1 to 3; publishing never at Level 4; a named officer answers (answerability principle).
  • Observable, expected, speculation: Observable: market-based zero training emissions alongside large location-based figures; criticism of AI-climate claims. Expected: stricter claims law. Speculation: hourly matching becoming mandatory.
  • Figure: A claim from figure to publication: boundary, baseline, independent check and approval by a named person
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Meta Llama 3.1 405B model card: 8,930 tCO2e location-based and 0 market-based [verify; snippet-only in dossier 02].
  • GHG Protocol Scope 3 revision decks (Apr 2026) and the corporate accounting preprints arXiv 2606.10660 and arXiv 2608.06733 [verify; unread].
  • Beyond Fossil Fuels 'AI Climate Hoax' (Feb 2026) on claim patterns [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 23: Procurement and Public Authorities: Clauses, Disclosures and Evidence from Suppliers

  • File: 23_procurement_and_public_authorities.md
  • Status: draft
  • Purpose: Cover procurement and public authorities: how to buy AI with environmental evidence. Contract clauses (disclosure of boundary-stated footprint, hosting region and grid, water, hardware lifecycle, reporting cadence, audit, change of model), award criteria and technical specifications under public procurement law, supplier disclosure records SDR, how to handle suppliers that cannot or will not disclose, and the particular duties and constraints of a public authority (transparency, equal treatment, proportionality).
  • Builds on: AI Governance, Chapter 7; AI Governance, Chapter 17; AI Platform and Security, Chapter 14; AI Leader, Chapter 18
  • Environmental functions: E5 Account
  • Sections: A framework with no figures; What to ask for; The legal frame; Award and specification; Supplier disclosure records SDR; When suppliers cannot say; Contract performance; A model clause set
  • Practical guides: A tender annex for environmental disclosure with a scoring guide
  • Worked examples: Illustrative: an SDR for the citizen-service assistant's hosting provider, with what it gave and what it did not
  • Nordvik Region and Hansa Mutual: Nordvik Region: the central case: public procurement through frameworks and a regional cloud; SDR-NR-0001 and a clause set for the next call-off. Hansa Mutual: a private buyer with more bargaining position on its own GPU capacity and cloud commitments; SDR-HM-0001 for its main provider. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation: AI may draft a tender annex or score disclosures at Level 1 to 2; award decisions stay with named officials; no Level 4.
  • Observable, expected, speculation: Observable: public procurement criteria for data centres and green public procurement guidance. Expected: disclosure clauses in AI frameworks. Speculation: standard model clauses adopted across the sector.
  • Figure: From tender to contract: disclosure items, award criteria and the supplier disclosure record in the lifecycle of a purchase
  • Planned length: 4,500-5,800 words of prose
  • Evidence to verify:
  • Directive 2014/24/EU and EU green public procurement criteria for data centres and cloud [verify the current criteria and dates].
  • Swedish public procurement act and Swedish public-sector guidance on environmental criteria [verify via the Riksdag SFS database and the national procurement authority].
  • The state of supplier disclosure: provider environmental reports and per-prompt figures as the baseline of what is available [verify]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 24: Governance: Roles, Budgets, Decision Records and the Environmental Register

  • File: 24_governance.md
  • Status: draft
  • Purpose: Cover governance: roles, budgets, decision records and the environmental register. Define who owns the footprint, the benefit, the budget and the claim; the decision forums and their tiers; AI Decision Records for environmental decisions (AIDR-1300 to AIDR-1399); the environmental register as the single place where every AI use shows owner, footprint record, benefit record, budget and evidence age; the role of the board under the answerability principle; and how competence is maintained. Apply the net-benefit principle to a new, scaled or renewed use.
  • Builds on: AI Governance, Chapter 13; AI Governance, Chapter 9; AI Leader, Chapter 12; AI Leader, Chapter 13; AI Operations, Chapter 22; AI Developer, Chapter 4
  • Environmental functions: E2 Justify; E5 Account
  • Sections: A register with no owner; Roles; Forums and tiers; The environmental register; The gate; Decision records; Answerability; Competence and assurance
  • Practical guides: The gate in one page: three tests, evidence required, decision, record
  • Worked examples: Illustrative: a register extract for each organisation with one row failing the Compare test
  • Nordvik Region and Hansa Mutual: Nordvik Region: the register is owned by the sustainability lead with the platform team; decisions routed through the regional AI board; AIDR-1301 to AIDR-1349 illustrative. Hansa Mutual: the register is owned by the group sustainability office and the platform owner; the board approves claims; AIDR-1351 to AIDR-1399 illustrative. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Delegation per decision class (measure, recommend, schedule, provision, scale, stop, publish) with owners and the Level 4 prohibitions.
  • Observable, expected, speculation: The register and the gate are the book's own conclusions; statements about the future of governance are labelled as in AI Governance, Chapter 22.
  • Figure: The environmental register and the gate: a row per use, three tests, owners and evidence age
  • Planned length: 4,800-6,200 words of prose
  • Evidence to verify:
  • Series practice: decision forums and tiers (AI Governance, Chapter 13), governance as code (Chapter 9), the answerability principle (AI Leader, Chapter 4) [verify the chapter references against SERIES_FOUNDATIONS before drafting].
  • ISO 14001:2026 (April 2026) for environmental management systems, if the chapter relates the register to it [verify; search summary only]. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 25: Efficiency or Sufficiency? A Debate on Limits

  • File: 25_efficiency_or_sufficiency.md
  • Status: draft
  • Purpose: Stage the debate on limits: efficiency or sufficiency. Can efficiency gains alone bring AI within environmental limits, or must organisations also choose to do less (sufficiency: decline uses, cap volume, stop)? State both positions at their strongest, separate what is reported from interpretation, find where they meet, and end with what an organisation can do. This is one of the two chapters (with Chapter 29) where the author's own projections appear, labelled as such.
  • Builds on: AI Governance, Chapter 22; AI Leader, Chapter 25; AI Operations, Chapter 15
  • Environmental functions: E2 Justify; E5 Account
  • Sections: A choice between two kinds of progress; The case for efficiency; The case for sufficiency; Where the positions meet; What can be done; The author's projections; Verdict
  • Constructive debate: Concern: Efficiency gains have been outrun by demand in other sectors; per-unit improvements without a cap do not reduce totals; local limits (grid, water, land) bind regardless of per-unit progress; unproven benefits are not a reason to scale. Calm: AI is young and its efficiency is improving by large factors; hardware and software gains plus cleaner grids can outpace demand; benefits may reduce emissions elsewhere; sufficiency without evidence can forgo real benefit. Where they meet: A cap is a design choice, not an ideology; necessity and sufficiency of model are the first and second levers of the hierarchy; both camps accept measurement and review at renewal. What can be done: Set caps per use and in total; decline uses that fail Compare; review at renewal; publish the criterion; revisit when the markers move.
  • Practical guides: A sufficiency review: for each use, what happens if it stops, shrinks or waits
  • Worked examples: Illustrative: a use that each organisation chose to decline or cap, with the reasons recorded
  • Nordvik Region and Hansa Mutual: Nordvik Region: a public authority answering to residents and politicians chooses, openly, which uses it will not scale; it publishes the criterion. Hansa Mutual: the board decides what growth in AI use it will not fund on environmental grounds. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: Final table: efficiency measures and sufficiency measures against the delegation levels of the agents concerned.
  • Observable, expected, speculation: Author's projections are labelled and confined to this chapter and Chapter 29. Everything else is labelled observable, expected or speculation.
  • Figure: Efficiency and sufficiency: two ways of keeping total footprint within a limit, and the Cap that joins them
  • Planned length: 5,000-6,500 words of prose
  • Evidence to verify:
  • The IEA's scenarios (Base, Lift-Off, High Efficiency 2035) as the quantified frame: High Efficiency about 20% below Base [verify; dossier 05 saw a secondary reproduction].
  • Energy economics literature on rebound (to be located by the writer; none is cited in the dossiers) [verify].
  • Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Part VII: Experience and Conclusions

Chapter 26: Failures and Anti-Patterns

  • File: 26_failures_and_anti_patterns.md
  • Status: draft
  • Purpose: Collect failures and anti-patterns in AI environmental practice, by mechanism and with attribution to what was reported: figures without boundaries, a per-query figure used as a total, market-based claims presented as physical, benefits offset against footprints, dashboards with no owner, efficiency work that was outrun by volume, an agent without a budget, a supplier disclosure that was never read, a claim approved by nobody, hardware refreshed for an efficiency gain smaller than its embodied cost. Each pattern is linked to the function and the test it violates and to the record that prevents it.
  • Builds on: AI Governance, Chapter 20; AI Leader, Chapter 30; AI Developer, Chapter 30; AI Operations, Chapter 28
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: Ten patterns; Figures without boundaries; Offsetting and greenwashing; Efficiency outrun by volume; Agents without budgets; Disclosures nobody read; Hardware churn; Public cases as reported; A pattern check
  • Practical guides: A pattern check of ten questions and the register field that answers each
  • Worked examples: Illustrative composite cases, labelled illustrative, one per pattern group
  • Nordvik Region and Hansa Mutual: Nordvik Region and Hansa Mutual each carry two of the patterns in their early registers, told without blame. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: A table of patterns by usage mode and delegation level at which each occurs.
  • Observable, expected, speculation: Observable: the reported cases; expected and speculation labelled where the chapter looks ahead.
  • Figure: Ten anti-patterns mapped to the function they break and the record that prevents them
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • Public cases must be found and verified by the writer in primary sources (company reports, regulator decisions, court or advertising-standards rulings on environmental claims); none is named in the dossiers. Dossier items marked as search snippets are leads, not sources: the writer verifies each against the primary document, states boundary, year and retrieval month, and drops what cannot be verified.

Chapter 27: A Year of Measuring and Optimising: Nordvik Region and Hansa Mutual

  • File: 27_a_year_at_nordvik_and_hansa.md
  • Status: draft
  • Purpose: Tell a year of measuring and optimising at Nordvik Region and Hansa Mutual, quarter by quarter: what each counted first, which levers they pulled in which order, what the records showed, where rebound appeared, which claim was withdrawn, and what changed in the contracts. All figures of the two organisations are illustrative and labelled so; they follow the profiles and the artefact identifiers of the book.
  • Builds on: AI Operations, Chapter 29; AI Developer, Chapter 31; AI Leader, Chapter 31
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: Quarter one: counting; Quarter two: justifying; Quarter three: designing and operating; Quarter four: accounting; What the register shows; What went wrong; What they would do differently
  • Practical guides: A year plan template: four quarters, one function per quarter, with gates
  • Worked examples: Two illustrative registers at year end; one agent budget breach and its review
  • Nordvik Region and Hansa Mutual: This is the book's full narrative of the two organisations. Nordvik Region: a small team, supplier-reliant, public duties. Hansa Mutual: own and cloud capacity, board reporting. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: A table of delegation changes during the year per class of action, and the reasons.
  • Observable, expected, speculation: The account is illustrative; statements about what the organisations expect are labelled expected or speculation.
  • Figure: A year in four quarters: the functions applied in order at each organisation
  • Planned length: 5,000-6,500 words of prose
  • Evidence to verify:
  • No external evidence is needed beyond the earlier chapters; every figure is illustrative and flagged. Facts must agree with docs/EXAMPLE_ORGANISATIONS.md and the AIDR allocation (1300 to 1349 Nordvik Region, 1350 to 1399 Hansa Mutual).

Chapter 28: Nordvik Region and Hansa Mutual: Conclusions for Two Organisations

  • File: 28_conclusions_for_nordvik_and_hansa.md
  • Status: draft
  • Purpose: Draw conclusions for the two organisations: what each should do next, in order, with the evidence it needs, the decision records it should hold and the budget and stop path it should have. Mark what is the same for both and what differs because of the public or private setting, the legal regime and the way each buys or runs capacity.
  • Builds on: AI Platform and Security, Chapter 24; AI Operations, Chapter 30; AI Developer, Chapter 32; AI Leader, Chapter 32
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: Nordvik Region; Hansa Mutual; What is common; What differs; A first ninety days; What would change the plan
  • Practical guides: A ninety-day plan template
  • Worked examples: The two plans, illustrative, side by side
  • Nordvik Region and Hansa Mutual: The chapter is about the two organisations. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: A table of the target degrees of delegation per class of action for each organisation.
  • Observable, expected, speculation: Conclusions are the book's own, drawn from the evidence of Parts II to VI; any statement about the future is labelled.
  • Figure: Two plans in parallel: functions, owners, records and first ninety days at each organisation
  • Planned length: 4,200-5,600 words of prose
  • Evidence to verify:
  • None external; the chapter draws on Chapters 1 to 27 and the organisation profiles.

Chapter 29: Conclusion

  • File: 29_conclusion.md
  • Status: draft
  • Purpose: Conclude the book. Judge the five-part thesis against the evidence: the net-benefit principle, the use as the unit, the optimisation hierarchy, agents with budgets, and claims with evidence. Say what the evidence supports, what it does not, and what remains open. This is the second chapter (with Chapter 25) where the author's own projections appear, labelled observable, expected or speculation, and the last section gives the reader something to do on Monday.
  • Builds on: AI Governance, Chapter 22; AI Governance, Chapter 25; AI Leader, Chapter 33; AI Developer, Chapter 33
  • Environmental functions: E1 Measure; E2 Justify; E3 Design; E4 Operate; E5 Account
  • Sections: The thesis judged; What remains open; The author's projections; What to do on Monday; What would change my mind; A last word
  • Practical guides: A one-page Monday plan per reader type
  • Worked examples: The two organisations' registers at the book's end, illustrative
  • Nordvik Region and Hansa Mutual: The two organisations close the book with the one decision each would take first. The details follow the environmental profiles in docs/EXAMPLE_ORGANISATIONS.md; the figures of the two organisations are illustrative and labelled so.
  • Usage modes and degrees of delegation: A final table across usage modes and degrees of delegation.
  • Observable, expected, speculation: Author's projections are labelled and confined to this chapter and Chapter 25.
  • Figure: The thesis judged against the evidence: five claims and the strength of support for each
  • Planned length: 4,000-5,400 words of prose
  • Evidence to verify:
  • None external; the chapter draws on the whole book. Refresh the dated snapshots listed in HANDOVER.md before release.

Appendices

  • 30_glossary.md: Defines the vocabulary of the book: environmental footprint, environmental benefit, the five environmental functions, the optimisation hierarchy, the net-benefit principle (Count, Compare, Cap), environmental budget, stop path, boundary, functional unit, operational and lifecycle, location-based and market-based, withdrawal and consumption, PUE, WUE, REF, SCI, embodied impact, rebound, GreenOps, the book's artefacts and the series' usage modes and degrees of delegation.
  • 31_appendix_a_code_examples.md: Appendix A: complete versions of the book's record formats: footprint record (FPR), benefit record (BEN), environmental budget (EBG), placement policy (PLC), supplier disclosure record (SDR), environmental claim record (CLM), hardware lifecycle record (HWL), the environmental register, the mandate manifest with a resource section (MND-NR-ENV, MND-HM-ENV) and a budget-breach runbook (RBK).
  • appendix_templates_and_tools.md: Appendix B: the checklists and templates of the book (method on one page, boundary check, claim review, tender annex, ninety-day plan) and the categories of tools with what each must provide, without recommending a supplier.
  • 32_references.md: The only bibliography of the book.
  • 33_about_the_author.md: About the author of The Transformed Architecture, placed at the very end of the book, and a note on how the book was made.