From the free book Delta: Closing the Specification Gap (Dhuri, 2026). The PDF is the canonical edition — figures, tables, and code formatting are simplified online.
J Research & Statistics Notes This appendix provides full citations, verification URLs, and context for every statistic cited in the book. Statistics are listed by chapter of first appearance. Where I quote a figure that does not have a peer-reviewed citation attached, it is a directional pattern based on my own reading of published research, post-mortem reports, and industry surveys on AI-assisted software development. These figures are indicative — useful for setting expectations and shaping decisions, but not the output of any independently audited survey or confidential dataset. The directional claims are aligned with the published research cited in this appendix, and any reader who wants to verify a specific number can follow the citation to the original work.
189
Meaning
Verified
Primary source publicly available online. URL provided. Directly verifiable.
Reported
Secondary citation. The statistic appears in the named report; the report is publicly available.
Illustrative figure. Directional pattern well established in the industry post-mortems on AI-assisted development. Not independently audited; not based on a private survey or confidential dataset.
Widely cited figure from industry without a single primary source. Noted with caution.
Chapter 1 — Why Prompting Is an Engineering Skill 55% faster task completion with AI coding tools Verified — GitHub Research Kalliamvakou, E. (2022). ”Research: Quantifying GitHub Copilot’s impact on developer productivity and happiness.” GitHub Blog.
URL: https://github.blog/2022-09-07-researchquantifying-github-copilots-impact-on-developerproductivity/ Methodology: Controlled study with 95 developers divided into Copilot and non-Copilot groups. Task: write an HTTP server in JavaScript. Copilot group completed 55% faster on average (1 h 11 min vs 2 h 41 min). 92% of US developers using AI coding tools J Research & Statistics Notes
190
Verified — GitHub Octoverse 2023 GitHub. (2023). ”Octoverse: The state of open source and rise of AI in 2023.” github.com/features/copilot. Applies to US-based developers specifically. Global figure from same survey: significantly lower. Context: adoption does not equal structured usage. The AI productivity paradox (Chapter 1.5) Directional / independent research. The claim that developers can feel faster while measured task completion is slower, and that AI-assisted code can carry higher defect and duplication rates, is drawn from independent (non-vendor) studies published 2025-2026. The specific percentages differ by study and methodology and should be read as directional, not audited constants. METR (2025-2026). ”Measuring the Impact of AI on Experienced OpenSource Developer Productivity.” Randomized controlled trial; experienced developers were measured taking longer on real tasks with frontier AI tools than without, despite perceiving a speed-up. URL: https:// metr.org/blog (see the developer-productivity study). Note: early result; the authors caution it reflects a specific population and period. GitClear (2024-2025). Longitudinal analysis of a large code corpus reporting increased code duplication and reduced refactoring over the AIadoption period. Independent analysis; figures are the authors’ and are cited here as directional. Independent PR/quality analyses (2025-2026), e.g. studies of AI-coauthored vs human-only pull requests reporting higher issue density in AI-assisted changes, and DORA/operational analyses reporting larger, slower-to-review PRs. Cited as directional, non-vendor evidence; vendorsponsored figures tend to be more favorable and are weighed accordingly. Interpretation note: the framing that ”the productivity paradox is a specification paradox” is the author’s own argument, not a claim made by the cited studies. The studies report the effect; the explanation — that the failures are upstream specification gaps — is this book’s thesis.
Chapter 2 — How LLMs Work ~90% cost reduction on prompt cache read tokens J Research & Statistics Notes
191
Verified — Anthropic API Documentation Anthropic. (2024–2025). ”Prompt caching.” Anthropic Documentation.
URL: https://docs.anthropic.com/en/docs/build-withclaude/prompt-caching Cache read pricing is a fixed percentage of the base input token price. Exact percentage varies by model tier. The ~90% figure represents cache read vs. uncached input token cost at standard tier pricing as of publication. Verify current pricing at anthropic.com/pricing. CO-STAR prompt-engineering framework (prior art for the Specification Frame) Verified — Towards Data Science (December 2023) Teo, S. (2023). “How I Won Singapore’s GPT-4 Prompt Engineering Competition.” Towards Data Science, 29 December 2023.
URL: https://towardsdatascience.com/how-i-wonsingapores-gpt-4-prompt-engineering-competition34c195a93d41 Six-element prompt structure: Context, Objective, Style, Tone, Audience, Response. Originally developed by data scientist Sheila Teo on the GovTech Singapore Data Science & AI team. The Specification Frame in this book builds on the same general principle — structured separation of the elements a prompt must convey — but uses a different four-element decomposition (role, context, task, constraints) chosen for enterprise codegeneration specifically. RISEN prompt-engineering framework (prior art for the Specification Frame)
192
Verified — promptentrepreneur (Kyle Balmer, 2023–2024) Balmer, K. (2023–2024). The RISEN framework. promptentrepreneur (TikTok and LinkedIn).
URL: https://www.linkedin.com/in/kylebalmer (creator's public profile) Five-element prompt structure: Role, Instructions, Steps, End Goal, Narrowing. The framework was popularized through Balmer’s promptentrepreneur content on social platforms (TikTok, LinkedIn) rather than a single citable publication. The Specification Frame in this book draws on the same intuition — that prompts perform better when each component is named and separated — while using a four-element decomposition tailored to enterprise code generation rather than general-purpose tasks.
Chapter 4 — Code Generation Domain-specific corrections required by AI-generated code
Author note About these numbers: the percentages above are indicative — what I have seen across enterprise .NET and Java codebases, not measurements from a controlled study. By “domain-specific correction” I mean any PR review comment that requires a change to a type, pattern, or constraint that is specific to the organization’s standards and not present in the generation prompt. The directional claim — that AI-generated code in regulated enterprise contexts usually needs domain corrections, and that structured prompts cut that revision rate substantially — has published support; see Sajid et al. (2023, arXiv:2311.07599) for the closest peerreviewed evidence. The 68% / 19% figures are illustrative magnitudes, not data points from a single dataset. Use them as a starting hypothesis. Measure your own team’s revision rate before and after adopting the Frame; that is the figure that should drive your decisions.
193
76% of developers use or plan to use AI coding tools Verified — Stack Overflow Annual Developer Survey 2024 Stack Overflow. (2024). ”Stack Overflow Developer Survey 2024.” survey.stackoverflow.co/2024.
URL: https://survey.stackoverflow.co/2024/ai Survey of 65,000+ developers worldwide. 76% reported using or planning to use AI tools in their development workflow in 2024. This is the adopted statistic replacing a prior unverified figure in the development manuscript.
Chapter 10 — Security Injection attacks: OWASP Top 10 for Web Applications Verified — OWASP OWASP Foundation. (2021). ”OWASP Top 10: 2021.” owasp.org/Top10.
URL: https://owasp.org/Top10/ A03:2021 Injection. Prompt injection is a new sub-category not explicitly listed in OWASP Top 10 2021 but addressed in the OWASP LLM Top 10 2025 (LLM01: Prompt Injection). See owasp.org/www-project-top-10for-large-language-model-applications. OWASP LLM Top 10: LLM01 Prompt Injection Verified — OWASP LLM Top 10 OWASP Foundation. (2025). ”OWASP Top 10 for LLM Applications 2025.” genai.owasp.org.
Chapter 14 — The Toolchain J Research & Statistics Notes
194
Reported — IDC IDC. (2024). ”Worldwide AI and Generative AI Spending Guide.” International Data Corporation. IDC research products are subscription-based. This figure appears in multiple IDC press releases (publicly available). The projection is cited as an indicative order-of-magnitude figure. Verify current IDC projections at idc.com.
Observation-Based Statistics: Methodology Some figures in this book do not come from a single citable study. Where I quote a number that I describe as illustrative or observational, it is a directional pattern from three sources: my own independent experimentation with AI coding tools; my reading of published research, post-mortem reports, and industry surveys on AI-assisted development; and iterative testing of every prompt template in this book against code written for this book in the four covered languages. These figures are useful for setting expectations and shaping engineering decisions. They are not the output of a peer-reviewed study, a private survey, a consulting engagement, or any confidential dataset. Where a directional claim has a specific published source, that source is cited with the relevant entry below. Marker
Verified
Author has accessed and confirmed the public source. URL provided.
Cite directly with confidence.
Reported
Industry research from a named source. Cited but not independently audited by the author.
Useful as directional. Verify against current data if precision matters.
Author’s professional pattern recognition. Presented as illustrative magnitudes, not measurements.
Treat as engineering rule-of-thumb. Validate against your team’s data before quoting.
Author-constructed teaching example (the “Pattern” boxes). Demonstrates failure mechanisms documented in published literature, not actual incidents.
Use as a teaching aid. Do not cite as a case study.