Changelog

All notable changes to this research package are documented here. Versioning follows Semantic Versioning in spirit — a major version bump indicates a change to canonical figures or conclusions, not just prose.

[Unreleased] — targeted as v1.0.2

Fixed

• Fixed an arithmetic error in the delegated-agent band’s Hyperscale $/AI-working-hour figure: whitepaper §25 and 13-slide-deck-outline.md (and the deck built from it, before this fix) cited a stale $0.0252/hr figure and a “~150x” spread, which contradicted Section 17’s own matrix, the canonical CSV, and the correct arithmetic (400,000 tokens/hr × $0.133/M = $0.0532/hr). Corrected to the right $0.0532/hr and “~68x” figure in both files; 19-slide-deck.pptx already had the correct figure on slide 13 and did not need changing. Caught during the Finnish translation’s cross-check against the canonical model — fixed in both languages simultaneously.

Added — content

• Split the human-work comparison (whitepaper §19.2; 04-workbook-ai-working-capacity-conversion.md Part E.2) into two explicit columns: Human employer cost, fully loaded (internal hire — wage + benefits + overhead) and Human billable / externally purchased rate (agency/contractor/consultancy, which embeds its own margin and overhead on top). States the pairing explicitly: owned-AI production cost compares against human employer cost; retail API price compares against human billable rate — mixing the pairs mixes the same internal/external cost layers the paper already warns against mixing on the AI side alone (owned-production cost vs. retail API price). Added illustrative billable-rate bands (~2-3x the employer-cost bands) as a clearly-flagged scenario assumption.
• Added a compact “1 working hour / 10-team hour / 1 working day / 10-team day” table (whitepaper §16, after the usage-intensity band table; 04-workbook-ai-working-capacity-conversion.md new Part B.2) stating explicitly that orchestration does not create a different kind of hour — it creates multiple working hours in parallel, and that team working capacity is workers × hours on both the AI and human side, with no exception. Added the same table and framing to slide 12 of the deck (13-slide-deck-outline.md/19-slide-deck.pptx), and the employer/billable split to slide 14.
• Changed whitepaper §24’s SME/owner-manager lens description from “must govern an AI workforce it did not previously need to manage” to “must lead and govern an AI workforce it did not previously have” — makes “lead” the default term for the human-AI-team relationship in the maturity narrative, reserving “supervise” for the specific review/risk-cost component already used correctly in the usage-band descriptions (§16-17, Part E.3’s supervision-burden caveat) and “govern” for boundaries/accountability.
• Added whitepaper §3.6, “A living reference model — the framework is durable, the numbers are not”: states explicitly that the core chain, the AI-maturity axes (Advise→Cowork→Delegate→Lead; Individual→Team→AI workforce), and the evidence-class discipline are intended to stay stable, while every numerical input (hardware price, electricity price, financing cost, utilization, human employer/billable cost, retail token price) is expected to keep changing — and that any cited figure should identify its scenario, assumptions, and model version rather than being read as a fixed fact.

Added — GitHub Pages site architecture

• Native HTML pages for every markdown document. Built tools/build-pages.py (pandoc + a shared page template, tools/page-template.html) to generate a matching *.html page for all 21 tracked markdown documents (whitepaper, all workbooks, all appendices, README, CHANGELOG, etc.), with cross-references between documents rewritten from .md to .html so reading the package end-to-end no longer requires jumping to GitHub’s markdown viewer. index.html’s reading-path and package-contents links now point at these local pages instead of github.com/blob/main/*.md URLs. Extracted the landing page’s inline CSS into a shared assets/site.css, extended with article/table/TOC/breadcrumb styles for content pages.
• Diagrams gallery (diagrams.html) — all 11 diagrams inline with captions and descriptive alt text, replacing a link to GitHub’s directory listing (which 404s on Pages, since Pages does not browse directories).
• Perspectives & Writeups section (perspectives.html, perspectives/*.yaml) — a filterable index of original, independent, and critical publications derived from or challenging this research. Each entry is a YAML file (schema in perspectives/README.md); tools/build-perspectives.py renders them into a client-side-filterable page (by type and topic). Seeded with three entries for the author’s own prior published pieces already verified in the source register (“Who Owns Your AI,” the digital-sovereignty cooperative-cost article, and “The Overlooked AI Scale-Up Layer”). Added a submission call-to-action for third-party writeups.
• Backlink-discovery pipeline design (docs/backlink-discovery.md) — documents a discovery → classify → Perspectives-index architecture (Ahrefs/Semrush backlink APIs feeding a status: pending-review candidate queue, never auto-publishing) for later implementation; not yet wired up, since it requires a paid API subscription and credentials this repository does not hold. The perspectives/_pending/ directory and the classification step’s data flow are in place so this can be added without a redesign.
• Model-version banner. The landing page now states the last tagged model version (v1.0.1) explicitly and restates the durable-framework/changing-numbers principle in the page itself, not just in the whitepaper.

Regenerates 01-whitepaper.pdf (89 pages, up from 88), 19-slide-deck.pptx, all markdown-sourced *.html pages, and perspectives.html; syncs the Desktop PDF copy; regenerates CHECKSUMS.sha256.

[1.0.1] — 2026-08-14

Added — content

• Replaced the front-matter “About the author and method” section with an expanded “About the author, perspective and method” — covers the author’s broader entrepreneurship/innovation-ecosystem and advisory background (30+ ecosystems, government/EU/university advisory work) alongside the AI-infrastructure-specific disclosure, and names ValtoAI and the EIOS reference framework (organizations/governed information/agent-ready context) alongside the previously-named Prifina, Digiole, and PIOS. Propagated the same venture list to Method §3.4’s disclosure, 12-executive-brief.md, 15-shortform-ownership.md, 13-slide-deck-outline.md/19-slide-deck.pptx (slide 15), and README.md, replacing all remaining “peecos/PIOS” references.
• Added a new conceptual-bridge section, “From AI working capacity to new value,” between Part IV (§25) and Part V of the whitepaper — explicitly tagged INTERPRETATION throughout, introduces no new cost figures. Distinguishes “can the capacity be created” (this paper’s cost/capacity model, Parts I–IV) from “can that capacity create and capture new value” (invention, innovation, adoption, and revenue-model questions this paper’s cost tables cannot answer), and notes the same boundary applies to any future Scenario Explorer built on this paper’s workbooks. Added a corresponding paragraph to 12-executive-brief.md and a new slide 18 to 13-slide-deck-outline.md/19-slide-deck.pptx (deck is now 26 slides, up from 25; slides 18–25 renumbered to 19–26).
• Added Diagram 11 (“Creating Growth Based on New Value”) — the author’s own pre-existing framework diagram, supplied directly and used as-is at assets/diagrams/diagram-11-creating-growth-from-new-value.jpg. Unlike Diagrams 1–10, this was not built from a from-scratch designer brief; 17-visual-asset-briefs.md’s Diagram 11 entry instead documents the actual image’s structure (two columns converging on a shared “New Value” band) for editors and consistency checking. A provenance-framed PNG derivative (diagram-11-creating-growth-from-new-value.png, evidence-class tag, attribution, source note, standard footer) was added for use in the PDF, deck, and Pages site; the original JPG is kept as the raw source.
• Fixed a stale cross-reference in README.md’s disclosure line (pointed to whitepaper “§4,” which no longer exists as a numbered section; corrected to “Method §3.4”).

Added — release mechanics (in response to external audit)

• tools/requirements.txt — pinned dependency versions (openpyxl, python-pptx, pypdf) for a reproducible consistency-check environment.
• Rewrote tools/check-canonical-consistency.py: required figures are now derived at runtime from data/canonical-cost-model.csv rather than hard-coded; added a PDF-text check (previously only markdown/xlsx/pptx were checked, so a stale PDF that wasn’t regenerated after a markdown fix could pass silently); every check now fails closed — a missing dependency or a broken Python environment (e.g. a pyexpat symbol mismatch) is reported as a failure, never silently skipped, and an unexpected exception inside any check is caught and reported rather than crashing the script.
• index.html + .nojekyll — a GitHub Pages landing page: reading paths, full package-contents tables, version/citation/licence links, and setup instructions for the consistency checker. Markdown documents link to their GitHub-rendered view; binaries (PDF/xlsx/pptx/checksums/diagrams) are served directly from the Pages site.
• 01-whitepaper.pdf now embeds its four most load-bearing diagrams (end-to-end chain, ownership stack, scale spectrum, new-value bridge) directly at their point of use in the text, so a reader holding only the PDF sees them without the repository; fixed a pdf-style.css gap (no img sizing rule existed) that was causing embedded images to overflow the page width before this was caught and corrected.
• CHECKSUMS.sha256 no longer contains comment/header lines, so it validates without warnings under both macOS shasum -a 256 -c and GNU sha256sum -c; the explanatory header moved to README.md.

Fixed

• main had drifted from CITATION.cff/CHECKSUMS.sha256/README.md’s v1.0.0 version claims after the content additions above — resolved by this v1.0.1 release.
• index.html: the “Diagrams (11)” link pointed to assets/diagrams/, which 404s on GitHub Pages (it does not browse directories) — changed to the GitHub tree view.
• README.md: fixed a stale “82-page document” reference (the PDF is now 88 pages) and a stale “not deployed to a public site (no GitHub Pages)” line — the Pages site has been live since this release’s preceding commit.
• 17-visual-asset-briefs.md: fixed “All 10 diagrams” to “All 11 diagrams” in the cross-diagram consistency notes.

[1.0.0] — 2026-08-13

Initial public release.

Package

• Full whitepaper (01-whitepaper.md / .pdf, 45 sections, 82 pages) tracing energy → hardware → compute → tokens → AI working capacity → work → outcomes → value → agency, from public statements by Jensen Huang, Sam Altman, and Mark Zuckerberg plus independently verified hardware specs and pricing.
• Six companion workbooks with fully shown formulas (Global Baseline, AI Working-Capacity Conversion, Token-Factory Scenarios, Investment-Thesis Notes, Humanoid Working-Capacity, Localized EUR-Finland template).
• A live-formula Excel workbook (18-companion-data-model.xlsx, 10 sheets), a 25-slide deck (19-slide-deck.pptx), and 10 diagrams (assets/diagrams/).
• A full source register (67 checks across 10 research clusters) and a consolidated assumption register.
• Five evidence classes (Observed Fact, Attributed Statement, Derived Calculation, Scenario Assumption, Interpretation) applied consistently across every asset.

Fixed during pre-release verification (three external review rounds)

• Corrected a Hyperscale-tier cost/scope error: the paper’s canonical full-layer figure ($0.091–$0.312/M tokens, mid $0.133/M) was, in earlier drafts, inconsistently compared against a partial-layer figure that excludes opex — now clearly separated everywhere with explicit scope labels.
• Corrected an invalid extrapolation from OpenAI’s own self-reported “60+ agent-hours/day” telemetry that had been used to imply a general-population usage statistic.
• Resolved an apparent ~2x discrepancy between two previously published cooperative-cost figures (€42 vs. €19.50–23/member/month) — they describe two different hardware tiers, not competing estimates of one cooperative.
• Confirmed the previously UNVERIFIABLE “$0.123/M tokens” GB300 inference figure directly against NVIDIA’s own site, with an explicit scope note restricting it to the specific rack/interactivity configuration it describes.
• Unified a cluster of stale Home-tier ($0.6–$2/M, superseded) and Cooperative-tier ($0.77–$1.20/M, superseded) figures across the whitepaper, workbook, slide deck, and diagrams to the single canonical Home ($1.37–$11.89/M) and Cooperative ($1.99–$7.62/M) ranges, and recomputed every dependent $/AI-working-hour figure from the canonical ranges.
• Fixed a 1,000x unit error in the Bitcoin-mining energy-monetization comparison; the corrected arithmetic inverts that passage’s original illustrative conclusion (production-cost-valued AI tokens exceed, rather than fall below, efficient Bitcoin-mining revenue per MWh).
• Fixed two prose-label cells in the companion Excel workbook that were stored as broken formulas (would render #NAME? in Excel).
• Fixed one Hyperscale-tier utilization assumption in the Excel workbook that was set to the high-case (90%) instead of the mid-case (60%) the canonical figure is built on.
• Fixed five diagram layout defects found during visual verification (label/caption collisions, an empty grid missing its cell labels, subplot overlap, off-canvas clipped content) across diagrams 1, 2, 3, 5, 9, and 10.

Added for this release

• 00-how-to-use-this-research.md — one-page orientation note.
• 20-appendix-known-limitations.md — consolidated list of live uncertainties (model performance, utilization, financing, power pricing, adoption timing, Professional-tier evidence gap).
• data/canonical-cost-model.csv — the single canonical source for every tier’s cost figures.
• tools/check-canonical-consistency.py — automated drift check across the whitepaper, workbooks, deck, and xlsx.
• CITATION.cff — machine-readable citation metadata.
• LICENSE — full CC BY 4.0 legal text plus a scope note distinguishing this author’s original content from quoted third-party statements, trademarks, and external data.
• CHECKSUMS.sha256 — SHA-256 manifest for every release asset.
• Reframed the paper’s opening question from an implicit bubble/transition binary to “what would need to be true for this spending to make economic sense” (whitepaper §6/§7, executive brief, slide deck).
• Added reading-path guidance to the README distinguishing the whitepaper (authoritative map), executive brief (entry point), workbooks (audit/exploration layer), and slide deck (conversation starter).

Known limitations

See 20-appendix-known-limitations.md for the full list. Highlights: DGX Spark throughput figures rest on a single unaudited community benchmark; all utilization, financing, and electricity-price inputs are Scenario Assumptions, not measured or quoted real-world figures; the Professional tier (HGX B300 node) has no public price or power spec and is flagged throughout as the least-evidenced tier; adoption-timing claims from named executives are Attributed Statements, not independently verified forecasts.