# Mind the Gap — Factcheck & Proof Register

**Scope:** `MTG_PDF_Reader_Rel2_V0_0.pdf` (15 chapters + front matter) and `1780159905368_MTG_BackMatter_v1.docx` (bibliography, about, colophon).
**Date of check:** 30 May 2026. **Method:** every stat / percentage / date / quote-attribution / named external claim extracted, then the highest-consequence and most-checkable verified against primary or strong secondary sources. Each entry carries a status and, where relevant, proposed softening wording.

**Status key**
- **GREEN — verified.** Source found, claim stands as written.
- **AMBER — Paul's call.** Defensible but contestable, slightly off, stale, or unsourced. Soften or confirm.
- **RED — fix before publication.** Wrong as written.
- **NOT VERIFIED THIS PASS.** Couldn't reach a source in this session; flagged for Paul / ChatGPT cross-check. Not asserted either way.

---

## 1. RED — fix before publication

### R1. "Google's Gemini 4 has a two million token context window." (Ch 10)
Wrong on both counts as of May 2026.
- There is **no Gemini 4**. Google's current line is **Gemini 3** (Nov 2025) and **Gemini 3.1 Pro** (19 Feb 2026), built on Gemini 3 Pro.
- Current Gemini context window is **1 million tokens**, not two. The 2M window belonged to the earlier Gemini 1.5 / 2.5 Pro generation.
- Also internally inconsistent: Ch 6 correctly states frontier windows run "200,000 to 1,000,000 tokens."

**Proof:**
- Google DeepMind model card, Gemini 3.1 Pro — "token context window of up to 1M," "based on Gemini 3 Pro." https://deepmind.google/models/model-cards/gemini-3-1-pro/
- Google blog, "Introducing Gemini 3" — "1 million-token context window." https://blog.google/products/gemini/gemini-3/
- Google Cloud docs, Gemini 3 Pro — "1M token context window." https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/3-pro

**Proposed fix (pick one):**
- Version-pinned: *"Google's Gemini 3 holds a one-million-token context window."*
- Future-proof (recommended — avoids re-dating at every model bump): *"The leading models now hold context windows of a million tokens or more."*

---

## 2. AMBER — Paul's call (soften, confirm, or reframe)

### A1. Frankl — "between stimulus and response there is a space" (Ch 3; Bibliography annotation)
Presented as Frankl's. The exact line does **not appear anywhere in Frankl's works**. It was popularised by **Stephen Covey**, who said he found it unattributed in a library book and never recorded the source. This is a famous misattribution — and it is famous *inside Paul's own field* (coaching/therapy), which is exactly where it costs credibility. *(Consequence is high enough that Paul may choose to treat this as Red.)*

**Proof:**
- Quote Investigator — "Researchers have been unable to find this passage in the works of Viktor E. Frankl… popularized by… Stephen R. Covey; however, he disclaimed authorship." https://quoteinvestigator.com/2018/02/18/response/
- Viktor Frankl Institute (via Check Your Fact) — "The statement appears nowhere in Frankl's written works… Covey… did not note down the book's author and title." https://checkyourfact.com/2019/09/27/fact-check-viktor-frankl-stimulus-response-space-growth-freedom/

**Proposed fix:** keep the idea, drop the quotation framing. e.g. *"Frankl's insight — that between stimulus and response lies a space, and in that space our freedom — "* → reframe as *"the principle most associated with Frankl: that between stimulus and response there is a space…"* and, if you want to be unimpeachable, add a half-line acknowledging Covey popularised the phrasing. The concept is genuinely Franklian; the verbatim line is not his.

### A2. "Google's data centres consumed six and a half billion gallons in 2023" (Ch 15)
Figure is slightly high and slightly mis-scoped. Google's **2023 total operations** were **6.4 billion gallons**; of that, **~6.1 billion (95%) went to data centres.** "Six and a half billion … data centres" rounds the *total* up and attributes it to *data centres*.

**Proof:** "In 2023, Google operations worldwide consumed 6.4 billion gallons of water (24.2 billion liters), with 95%, 6.1 billion gallons, used by data centers." https://www.civilbeat.org/2025/08/data-centers-consume-massive-amounts-of-water-companies-rarely-say-exactly-how-much/ ; corroborated https://gijn.org/stories/researching-water-consumption-data-centers/

**Proposed fix:** *"Google's data centres consumed around six billion gallons of water in 2023, most of it potable."* ("most of it potable" is supported.)

### A3. "A single ChatGPT query draws roughly ten times the electricity of a Google search" (Ch 15)
True to the *2024* IEA/Goldman framing (2.9 Wh vs 0.3 Wh), but **superseded**. Epoch AI (Feb 2025) put a typical GPT-4o query at ~0.3 Wh — *ten times less* than the old estimate, i.e. roughly the same as a search. Sam Altman (2025) cited ~0.34 Wh. An informed reader will know the 10× figure has been walked back.

**Proof:**
- Epoch AI — "We find that typical ChatGPT queries… likely consume roughly 0.3 watt-hours, which is ten times less than the older estimate." https://epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
- "The often-repeated claim that ChatGPT is ten times more energy-intensive than a Google search is no longer supported by current data." https://www.zmescience.com/science/news-science/how-much-energy-chatgpt-query-uses/

**Proposed fix:** attribute and date it — *"On the figures that circulated through 2024, a single ChatGPT query was estimated to draw roughly ten times the electricity of a Google search — an estimate later revised sharply downward as models and hardware became more efficient."* Or drop the multiplier and keep the directional point.

### A4. "thirty minutes of generative AI use consumes around a sixth of a gallon of water" (Ch 15)
In the ballpark of the widely-cited (and itself contested) water estimates (~500ml per session in the Li et al. line of work; Altman's ~0.000085 gal/query). The "sixth of a gallon / 30 minutes" basis isn't cleanly sourced and the field's figures vary widely.

**Proposed fix:** attribute/hedge — *"by some estimates, around a sixth of a gallon of water"* — or cite the specific source you drew it from.

### A5. [name held] et al. (2019) — "a 77-year-old brain" (Ch 15; Bibliography annotation)
Paper and citation are otherwise **correct** (see GREEN G6). But the study reports its oldest specimen as **78 years** ("20 gestational weeks to 78 years of age"), and frames persistence "into old age." The recurring "77-year-old" (used in Ch 15 and the bib) may come from a specific figure in the paper, but the headline age in the source is 78.

**Proof:** UCSF / Nature Communications coverage — "postmortem human amygdala tissue from 49 human brains — ranging in age from 20 gestational weeks to 78 years of age." https://www.ucsf.edu/news/2019/06/414756/mood-neurons-mature-during-adolescence

**Proposed fix:** confirm the exact specimen age against the paper's figures. If no 77-year-old specimen is specifically shown, change to **78**. (CL-31 is built on this line, so worth nailing.)

### A6. WaaS deployment figures — "Accuracy rates at 99.6%. Cycle times reduced by 68%. Cost to serve down 57%." (Ch 13)
Three precise percentages presented as generic ("early WaaS deployments") with no source named. Precise figures with no attribution are the single most exposed kind of claim in the book. **NOT VERIFIED** — they read like one specific vendor case study generalised.

**Proposed fix (in order of preference):**
1. Name the source/case study these came from, or
2. Soften to ranges and label as illustrative: *"Early deployments have reported accuracy in the high-90s, cycle times cut by more than half, and cost-to-serve down by roughly a half"* — and make clear these are vendor-reported.

### A7. Bibliography — Klein, *The Writings of Melanie Klein, Volume III* dated **(1946)**
The collected *Writings of Melanie Klein* (Hogarth) were **published 1975**; 1946 is the date of the key paper ("Notes on Some Schizoid Mechanisms") *within* the volume. As a book citation, the year is off.

**Proposed fix:** cite as *Klein, M. (1975). The Writings of Melanie Klein, Vol. III: Envy and Gratitude and Other Works 1946–1963. London: Hogarth.* (Keep 1946 only if citing the paper, not the volume.)

### A8. Bibliography — "Compass of Shame — Donald Nathanson (1992)"
Part One annotation titles the work *Compass of Shame*; Part Two correctly lists the actual book, *Shame and Pride: Affect, Sex, and the Birth of the Self* (1992). "Compass of Shame" is the **concept within** that book, not a separate title. Minor internal inconsistency.

**Proposed fix:** in the annotation, *"Shame and Pride (and the 'compass of shame' it introduces) — Donald Nathanson (1992)."*

### A9. Buddha — "we are what we think" (Ch 3, the five anchors)
Popular attribution, drawn loosely from a Dhammapada translation; the wording and direct attribution are contested among translators. Low consequence, but it sits in a list beside the Frankl line, so worth a consistent decision.

**Proposed fix:** *"the idea, often rendered from the Dhammapada, that we are what we think."* Or leave as a deliberately informal anchor and accept the looseness.

### A10. Ferrucci, *What We May Be* — dated **(1984)**
Original US publication was **1982** (Tarcher). "1984" is plausibly a UK / 3rd-edition printing (the bib says "3rd ed., Turnstone Press"), so likely fine — but worth confirming the printing you're citing.

---

## 3. NOT VERIFIED THIS PASS — recommend Paul / ChatGPT cross-check before print

These are checkable but I didn't reach a source this session. Listed so nothing is silently assumed true. (Several are in global memory as "verified May 2026" — re-confirm against a live source before KDP, as figures move.)

- **Ch 4** — Uber CTO disclosed full-year 2026 AI budget exhausted in first four months (book says *April 2026*; memory note says source is *The Information*). Confirm date + figure.
- **Ch 4** — Microsoft to cancel internal AI coding-tool licences for ~100,000 engineers, effective end June (book: "mid-May 2026").
- **Ch 4** — Anthropic decoupling agent usage from chat usage with its own metered credit pool from **15 June 2026** (memory: billing announced 13 May, live 15 June).
- **Ch 4** — FinOps teams "doubled in twelve months from thirty-one per cent to sixty-three per cent of large organisations." Specific percentages — confirm against the FinOps Foundation / source survey.
- **Ch 4** — Cybersecurity launch partner surfaced "approximately ten thousand vulnerabilities… first month saw ninety-seven patched." Confirm against the published source.
- **Ch 4** — SWE benchmark projections "non-specialised agents at fifty-four per cent and state-of-the-art at eighty-seven per cent" at start of 2026. Confirm against the benchmark cited.
- **Ch 4** — Pricing: "$15 per million output tokens at the top end… mid-tier… three dollars." Roughly consistent with current market (e.g. Gemini 3.1 Pro $12/M out; reasoning tiers ~$15/M) but confirm the figures you want to stand behind.
- **Ch 4 / Ch 6** — Anthropic **"Dreams"** feature and the quoted description ("let Claude reflect on past sessions to curate an agent's memory and surface new insights"); and the **"Code with Claude"** conference quote ("getting much closer to the feeling of an infinite context window"). Confirm both quotes verbatim against Anthropic's own docs/transcript — they're direct quotations, so wording matters.
- **Ch 10** — OpenAI **/Goal** feature ("announced this week"). Confirm it exists under that name.
- **Ch 4** — Training-token scale: early ChatGPT "~300 billion tokens / ~225 billion words"; GPT-4 "~13 trillion." These match credible leaked estimates but are not official; the book already hedges ("reportedly," "credible estimates"), which is the right posture.
- **Ch 4** — "Library of Congress… ~20 billion words across its entire collection." Contested figure; estimates vary enormously. The book uses it as an order-of-magnitude illustration, which is defensible, but the "20 billion words" specifically is soft.
- **Bibliography** — Bond, *Behind the Mask* (1998, Atlow Mill); Firman described as "a student of Ferrucci" (lineage claim, hard to source). Low consequence.

---

## 4. GREEN — verified, stands as written

### G1. Kohls v. Ellison (Ch 8; Bibliography)
Stanford professor **Jeff Hancock**, founding director of the **Stanford Social Media Lab**, used GPT-4o (i.e. ChatGPT); declaration contained **two non-existent citations** (plus one mis-cited); court **excluded the testimony**; order filed **10 Jan 2025**; Judge Provinzino named the irony explicitly. All accurate. (2025 WL 66514, D. Minn.)
**Proof:** https://reason.com/volokh/2025/01/10/misinformation-experts-citation-to-fake-ai-generated-sources-in-his-declaration-shatters-his-credibility-with-this-court/ ; https://www.law.berkeley.edu/wp-content/uploads/2025/12/Kohls-v-Ellison.pdf

### G2. Mata v. Avianca (Ch 8; Bibliography)
Citation **678 F. Supp. 3d 443 (S.D.N.Y. 2023)** correct. **Six** fabricated cases generated by ChatGPT; the three named (Varghese v. China Southern Airlines, Shaboon v. Egyptair, Petersen v. Iran Air) are all genuine fabrications from the case; lawyers sanctioned ($5,000). Accurate.
**Proof:** https://en.wikipedia.org/wiki/Mata_v._Avianca,_Inc. ; https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/

### G3. Moffatt v. Air Canada (Ch 8; Bibliography)
Citation **2024 BCCRT 149** correct. Chatbot advised a refund could be claimed **within 90 days**; actual policy required claiming before travel; tribunal rejected Air Canada's "separate legal entity" argument; airline held liable. Accurate.
**Proof:** https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html ; https://www.americanbar.org/groups/business_law/resources/business-law-today/2024-february/bc-tribunal-confirms-companies-remain-liable-information-provided-ai-chatbot/

### G4. IEA data-centre projection (Ch 15)
"Around 945 terawatt-hours by 2030 — equivalent to the entire current electricity demand of Japan." Verbatim match to the IEA *Energy and AI* report (Apr 2025).
**Proof:** https://www.iea.org/reports/energy-and-ai/executive-summary ; https://www.scientificamerican.com/article/ai-will-drive-doubling-of-data-center-energy-demand-by-2030/

### G5. OpenAI / Jony Ive device (Ch 4)
"$6.5 billion acquisition," "screenless, voice-first, pocket-sized," "calm computing," "late 2026 release" — all match reporting. io Products acquisition value $6.5bn (sometimes reported $6.4bn all-stock); H2 2026 debut confirmed at Davos.
**Proof:** https://www.benzinga.com/markets/tech/25/07/46335969/openai-finalizes-6-5-billion-deal-to-acquire-jony-ives-ai-hardware-startup ; https://introl.com/blog/openai-consumer-device-jony-ive-hardware-2026

### G6. Meta Ray-Ban smart glasses "seventy-five to eighty per cent of their category" (Ch 4)
Verified (75–80% market share).
**Proof:** https://introl.com/blog/openai-consumer-device-jony-ive-hardware-2026 ("Meta's Ray-Ban smart glasses captured 75-80% market share").

### G7. [name held] et al. (2019) citation (Bibliography)
*Nature Communications*, **10(1), 2748**, DOI 10.1038/s41467-019-10765-1, full author list, title "Immature excitatory neurons develop during adolescence in the human amygdala" — all correct. (Only the "77-year-old" detail needs the A5 check.)
**Proof:** https://www.nature.com/articles/s41467-019-10765-1 ; https://pubmed.ncbi.nlm.nih.gov/31227709/

### G8. GPT-5.5 attribution (Front matter / Colophon)
"Images by ChatGPT (OpenAI, GPT-5.5)" — GPT-5.5 is real, **released 23 April 2026**. Plausible for images generated late in the project.
**Proof:** https://openai.com/index/introducing-gpt-5-5/ ; https://www.cnbc.com/2026/04/23/openai-announces-latest-artificial-intelligence-model.html

### G9. Claude Opus 4.7 (throughout) — consistent with the current model landscape (Opus 4.x line; Mythos Preview referenced in coverage). Stands.

### G10. ISBN 978-1-0369-9170-8 (Colophon)
Check digit is mathematically **valid** (computed check digit = 8, matches). Structurally sound.

### G11. ChatGPT join date "Sunday 11 December 2022" (Ch 9 / Ch 10)
ChatGPT opened 30 Nov 2022; +11 days = 11 Dec 2022, which **was a Sunday**. Internally and externally consistent.

### G12. Internal-consistency checks that pass
- Train fare £18.80 vs breakfast £16.00; "£2.80 on a train fare" (Ch 13) = the £2.80 *shortfall* (18.80 − 16.00). Consistent.
- 76,000,000 words ÷ 783,000 (KJV) = 97.06 → "ninety-seven Bibles… call it a hundred." Consistent.
- 900,000 words ÷ 45,000 = 20 → "roughly twenty to one." Consistent.
- Token/word ratios (0.75): 200k tok ≈ 150k words; 200k–1M tok ≈ 150k–750k words; 300bn tok ≈ 225bn words. All consistent.
- Biography: apprentice 1978 at 16 → born ~1962 → age 63 in 2026. Consistent across chapters.

---

## 5. One cross-document note (not a fact, but worth a glance)
The back-matter Colophon states "**culminating in V69**"; the live manuscript baseline is **V76**. The back-matter file is `v1` and predates the current manuscript, so this is expected — but the version numbers (and the "forty-five-thousand-word" / "twenty-seven versions / six weeks" figures) will need a final sync before print. **Amber.**

---

*Prepared as a verification reference for Mind the Gap. Drafted with Claude (Anthropic, Opus 4.8). Sources are live URLs as of 30 May 2026; figures in fast-moving areas (model versions, token pricing, deployment stats) should be re-confirmed immediately before KDP upload.*