This source backbone collects the primary reading paths behind VS008 — The Confidence Gap. It is a public source map: a way to see the research, reporting, institutional documentation, and DFEI interpretation layers that informed the issue.
The backbone is not a claim-by-claim verification record and not a complete bibliography. It is a curated route through the sources most relevant to the issue's central questions:
Is AI deployment producing the labor and productivity outcomes the efficiency narrative claims? What is the structural gap between AI output confidence and the verification infrastructure available to evaluate it?
01 — AI Productivity Research
METR — Randomized Controlled Trial of AI Coding Assistants
Link: https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ Source type: Research organization / randomized controlled trial.
Why it matters: METR conducted a controlled trial of experienced open-source developers using state-of-the-art AI coding assistants on real-world tasks. Developers using AI tools took 19 percent longer to complete tasks than a control group working without them. When surveyed, AI-assisted developers believed they had been 20 percent faster — a 39-point perception gap. This is the sharpest documented instance of the gap between perceived and actual AI productivity in a controlled setting.
UC Berkeley / Yale — Embedded Workplace Study (2025–2026)
Link: https://newsroom.haas.berkeley.edu/ai-promised-to-free-up-workers-time-uc-berkeley-haas-researchers-found-the-opposite/ Source type: Academic embedded study.
Why it matters: Researchers tracked a 200-person technology company over eight months of voluntary AI adoption. 67 percent of workers who adopted AI tools reported working more hours by the end of the study period. The study also documented a 23 percent increase in sustained cortisol levels and a 40 percent rise in decision fatigue among AI-adopting workers. The study is the primary empirical grounding for the claim that AI deployment is increasing labor burden rather than reducing it.
ActivTrak — 2026 State of the Workplace Report
Link: https://www.activtrak.com/resources/state-of-the-workplace/ Source type: Vendor-produced workplace analytics report.
Why it matters: Drawing on over 443 million hours of digital workplace activity, ActivTrak's 2026 report documents AI's effect on work density, focus time, collaboration overhead, and weekend work patterns. Key findings cited in this issue: collaboration overhead increased 34 percent; weekend work became a baseline condition rather than a crunch artifact for AI-integrated workers; average daily focused time declined. The report provides the broadest dataset in the AI productivity research cluster.
BCG — AI Time Reinvestment Study
Link: https://www.bcg.com/publications/2026/ai-at-work-why-strategy-matters-more-than-tools Source type: Professional services research.
Why it matters: BCG research found that only 21 to 27 percent of employees reallocate AI-freed time to personal lives. The majority reinvest freed time into higher volumes of professional output. This finding is cited alongside the Adecco study as convergent evidence for the Jevons rebound in knowledge work.
Adecco — AI Time Reinvestment Study
Link: https://www.adeccogroup.com/our-thinking/flagship-research/workforce-trends-2026 Source type: Staffing and workforce research.
Why it matters: Adecco independently found the same 21 to 27 percent range as BCG for the proportion of employees who reallocate AI-freed time to personal lives. The convergence of two independent studies on this range strengthens the Jevons rebound claim.
Jevons, W.S. — The Coal Question (1865)
Reference: Jevons, W.S. (1865). The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines. Macmillan.
Source type: Historical economic text.
Why it matters: The original statement of what became known as the Jevons Paradox: improvements in steam engine efficiency increased coal consumption rather than reducing it, by making steam power viable for previously uneconomical applications. The paradox is the economic frame for the AI productivity rebound: lowering the marginal cost of cognitive output does not reduce demand — it expands it.
02 — Labor Market and Economic Terrain
PwC — 2026 Global AI Jobs Barometer
Link: https://www.pwc.com/gx/en/issues/artificial-intelligence/job-barometer/2026/2026-global-ai-jobs-barometer-full-report.pdf Source type: Professional services research report.
Why it matters: PwC's 2026 Jobs Barometer documents a bifurcating labor market: roles requiring demonstrated AI integration skills are growing eight times faster than the broader labor market and command a 62 percent wage premium. This is cited as the market's response to the confidence gap — the premium reflects demand for judgment to evaluate AI output, not enthusiasm for AI adoption.
Ummid News — AI is creating a 'two-track' labour market
Link: https://ummid.com/news/2026/6/19/ai-is-creating-a-two-track-labour-market-jobs-barometer.html Source type: News coverage (secondary reporting on PwC Barometer).
Why it matters: Secondary reporting on the PwC findings, providing a public-facing framing of the two-track labor market dynamic.
03 — AI Deployment Reliability
Anthropic — Claude Code Postmortem (April 23, 2026)
Link: https://www.anthropic.com/engineering/april-23-postmortem Source type: Primary institutional postmortem.
Why it matters: Anthropic's April 23, 2026 postmortem documented three compounding configuration changes to Claude Code introduced between March 4 and April 16, 2026: a reduction in reasoning depth (March 4), a bug causing mid-session reasoning history discard (March 26), and a 25-word response cap between tool calls (April 16, reverted April 20). The tool continued responding fluently throughout. The postmortem is the issue's SIGNAL: the most precisely documented public case of AI quality regression that projected confidence while degrading outputs, over a 50-day window, with no external detection mechanism.
Anthropic — Dario Amodei on production code authorship (May 2026)
Link: https://venturebeat.com/technology/anthropic-says-80-of-its-new-production-code-is-now-authored-by-claude-how-your-enterprise-can-keep-up Source type: Executive disclosure.
Why it matters: Amodei disclosed that more than 80 percent of the code merged into Anthropic's production systems in May 2026 was authored by Claude. This is cited not as a critique of Anthropic's deployment — which occurs under exceptional evaluation infrastructure — but as a reference point for the scale of AI-authored production code at the organization best positioned to evaluate it.
04 — Deployment Terrain: Platforms and Tools
Microsoft — Agentic Copilot in Office
Link: https://www.microsoft.com/en-us/microsoft-365/blog/2026/04/22/copilots-agentic-capabilities-in-word-excel-and-powerpoint-are-generally-available/ Source type: Platform announcement.
Why it matters: Microsoft Agentic Copilot's embedding in Office is cited as the clearest instance of the lay deployment gap: the most widely deployed enterprise software suite now includes autonomous, multi-step agentic capability available to users across every level of technical sophistication.
Google — Gemini 3.5 Flash
Link: https://stajic.de/blog/google-io-2026-gemini-omni-and-gemini-3-5 Source type: Technical analysis / Google I/O 2026 coverage.
Why it matters: Gemini 3.5 Flash's autonomous interface operation (Computer Use) is part of the terrain establishing that agentic capability has become a standard subscription feature, not a specialized deployment.
Google — I/O 2026 Official Blog
Link: https://blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-collection/ Source type: Primary institutional announcement.
Why it matters: The I/O 2026 collection documents Google's agentic and multimodal product releases, providing the primary sourcing for Google's Gemini and agentic product references in this issue.
OpenAI — GPT-5.5 and UI automation
Link: https://openai.com/index/introducing-gpt-5-5/ Source type: Platform announcement.
Why it matters: GPT-5.5's benchmark-leading UI automation performance is part of the terrain establishing the current accessibility floor for agentic deployment.
Lovable — Natural language to production code
Link: https://lovable.dev Source type: Product.
Why it matters: Lovable is cited as an example of a platform that lowers the barrier to production code deployment without requiring users to understand the underlying code, illustrating the non-technical accessibility of consequential AI capabilities.
Tech Jacks Solutions — Microsoft MAI-Thinking-1
Link: https://techjacksolutions.com/ai-brief/ai-models-news-mai-thinking-1-official-microsofts-first-in-h/ Source type: Technical coverage.
Why it matters: MAI-Thinking-1 and Microsoft's broader MAI stack document the move toward proprietary enterprise AI stacks, relevant to the institutional terrain layer.
05 — Regulatory and Governance
Morgan Lewis — June 2 Executive Order on AI Security
Link: https://www.morganlewis.com/pubs/2026/06/executive-order-promotes-public-private-cooperation-on-ai-innovation-and-security Source type: Legal analysis.
Why it matters: Morgan Lewis provides a detailed legal breakdown of the June 2 Executive Order: mandated AI integration into federal cyber defenses, 30-day pre-release vetting requirement for frontier models, and the mechanism by which the administration requested OpenAI limit GPT-5.6's rollout.
Vatican Press — Magnifica Humanitas (May 15, 2026)
Link: https://www.vatican.va/content/francesco/en/encyclicals/magnifica-humanitas.html Source type: Primary institutional document (papal encyclical).
Why it matters: Pope Leo XIV's encyclical establishes a moral framework for AI development from an institution with 1.4 billion adherents. The issue treats it as a distinct governance instrument — non-binding, diffuse in influence, but durable in ways that regulatory instruments often are not.
UN University — Environmental Cost of AI (2030 projections)
Link: https://unu.edu/inweh/news/environmental-cost-of-AIs-Enrgy-use-carbon-water-and-land-footprints Source type: UN research institution report.
Why it matters: The UNU report projecting 945 TWh electricity consumption and 2.5 million tonnes of e-waste from AI data centers by 2030 is part of the broader institutional terrain establishing regulatory pressure on compute scaling.
EU — Cloud and AI Development Act (CADA)
Link: https://digital-strategy.ec.europa.eu/en/library/proposal-cloud-and-ai-development-act-cada Source type: Regulatory proposal.
Why it matters: CADA advances EU infrastructure sovereignty and data center capacity as strategic assets, with long-term implications for how AI systems are built and deployed within European jurisdictions. The issue correctly frames it as an infrastructure and sovereignty instrument, not a liability or deployment-behavior instrument.
UN AI Governance Dialogue — Geneva, July 6, 2026
Link: https://www.un.org/global-dialogue-ai-governance/en Source type: Intergovernmental event.
Why it matters: The Geneva dialogue opens the issue's coverage window. Its significance is framed as process — shared definitional frameworks and political groundwork for eventual treaty-level instruments — rather than immediate binding authority.
06 — Infrastructure and Capital Terrain
Bloomberg — Nvidia / IREN $5.5B Infrastructure Deal
Link: https://www.bloomberg.com/news/articles/2026-06-29/nvidia-iren-infrastructure-deal Source type: Financial / business reporting.
Why it matters: The Nvidia/IREN deal contributes to the terrain establishing the scale of AI infrastructure capital expenditure in the coverage window.
Built In — SpaceX IPO
Link: https://builtin.com/articles/spacex-ipo Source type: Business coverage.
Why it matters: The SpaceX IPO ($75B) and its xAI/Starlink compute integration is part of the financial consolidation terrain in the coverage window.
SmartAsset — OpenAI IPO filing
Link: https://smartasset.com/investing/openai-stock-ipo Source type: Financial overview.
Why it matters: OpenAI's IPO filing ($1T valuation target) and Anthropic's filing ($965B target) document the capital consolidation layer.
07 — Historical Reference
Gartner — Supply Chain Macro Trends 2026
Link: https://www.gartner.com/en/supply-chain/trends-2026 Source type: Industry analyst report.
Why it matters: Gartner's identification of collaborative multiagent systems as the dominant force restructuring 2026 supply chains supports the deployment terrain establishing the scope of agentic expansion.
08 — DFEI Interpretation Notes
Several issue-level constructs combine external source material with DFEI interpretive framing. In these cases, external sources establish the terrain; DFEI supplies the operating language and diagnostic frame.
Examples include:
- "The confidence gap" — not a term from the external literature; DFEI names the structural gap between AI output confidence and organizational verification capacity.
- "Organizational hallucination" — a deliberate DFEI reframing of the technical AI term. The external sources (METR, UC Berkeley/Yale, Anthropic postmortem) document the phenomenon; DFEI names the organizational-level condition.
- "Verification tax" — the concept of review cost appears in the METR literature; the framing as a Jevons rebound mechanism is DFEI interpretation.
- "Lay deployment gap" — DFEI term for the structural gap between tool accessibility and accountability knowledge; not sourced to an external category.
- "Adaptation throughput" and "adaptation debt" — DFEI Field Artifact constructs derived from Table synthesis; not external research instruments.
- "The six conversion layers" — DFEI diagnostic framework. Individual layers reference established governance and accountability concepts; the framework itself is DFEI synthesis.
These constructs are not presented as direct quotations or established categories from the linked sources. They are DFEI operating frames built from the broader source terrain above.
DFEI.008 :: Source Backbone :: The Confidence Gap Dispatches From Emerging Intelligence :: Vector Intelligence Studio