Back matter · Evidence Register and Glossary
Evidence Register, Glossary, Sources, and Disclosures
I've asked you to take a lot on trust over eleven chapters. You shouldn't. Every claim I made traces to a row below, and every row traces to a source you can open yourself. This is the ai research evidence register for the whole report: the study, the narrowest claim it can carry, and the limit that claim runs into. If I stretched a finding somewhere, this is where you catch me.
When you build large power loads, you learn to trust the meter over the drawing. A summary is the drawing; the study is the meter. Here's the part most people skip: the limit column matters as much as the claim column. Check my work.
12.1 How to Use the AI Research Evidence Register#
The register indexes the major empirical and documentary evidence used in this report, organized by study program or record rather than by sentence. Exact estimates, denominators, methods, and citations remain in the chapter where each study is interpreted.
It does not make heterogeneous outcomes directly comparable, and one study cited in several chapters is still one study.
Here's how I'd read each row:
- Start with the ID. The letter gives the theme: E education, P pathways into work, H household, C capability, W work, V civic life, F finance, L later life, I infrastructure, G governance, T trackers.
- Open the source. Every row carries at least one link.
- Read the permitted claim as a ceiling, not a floor.
- Read the limit before you repeat the claim.
- Find the evidence class. The eight classes are defined in Chapter 11. A working paper is not a published trial, and a publisher statement is not verification.
Many rows are nulls, mixed results, or adverse findings, kept on purpose.
What the register contains, by domain
12.2 The Consolidated Evidence and Claim Register#
12.2.1 Education and Entry Into Work#
| ID | Evidence program | Principal permitted claim | Primary limit |
|---|---|---|---|
| E01 | Individualized kindergarten reading cluster trial (ERIC record) | A teacher-supported, assessment-informed package improved a defined reading measure in the studied setting. | Not unrestricted child-model interaction or a model-only effect. |
| E02 | Early mathematics application trial (full text) | Bounded applications improved selected early mathematics outcomes. | Specific applications and short study conditions. |
| E03 | Literacy-platform engagement trial (working paper) | Human engagement support can affect platform use and selected outcomes. | Working paper and program-specific bundle. |
| E04 | ASSISTments trial and independent review (ERIC, WWC review) | A small positive estimate exists, with the independent review retaining reservations and uncertain domain effects. | Do not omit the later review or present the estimate as universally established. |
| E05 | Tutor CoPilot study (preprint) | Adult-facing support improved selected tutor behavior and student mastery. | Preprint, specific tutoring setting, and no general teacher-replacement conclusion. |
| E06 | Khanmigo study (working paper) | A school implementation produced modest package-level results. | Attribution to the model alone is not established. |
| E07 | Teacher lesson-preparation trial (evaluation) | Preparation time fell for specified lessons under the tested support package. | Not total workload or student attainment. |
| E08 | High-school mathematics experiment (PNAS) | Unrestricted assistance improved assisted practice but reduced later unaided performance in the studied condition. | One subject, product configuration, and experiment. |
| E09 | Nigerian secondary English program (World Bank) | Supervised curriculum-focused use produced a positive English estimate. | Working paper, setting, program bundle, and transfer limits. |
| E10 | Cognitive Tutor Algebra I (RAND) | A historical adaptive system produced positive second-year results. | Not evidence that current conversational systems reproduce the effect. |
| P01 | College physics tutor (Scientific Reports) | A designed tutor improved immediate posttest performance in a bounded comparison. | Two lessons, one institution, and no long-term degree outcome. |
| P02 | College enrollment assistant (AERA Open) | Enrollment increased in the already-committed subgroup. | Not a significant full-sample general college-attendance effect. |
| P03 | Course outreach experiment (working paper) | Results are mixed and include nonsignificant adjusted pooled effects. | Do not elevate selected subgroup estimates. |
| P04 | Welding training trial (article) | The short comparison did not establish significant overall score differences. | A null difference is not proof of equivalence. |
| P05 | WeldAR crossover study (manuscript) | Immediate unassisted motion performance improved relative to video. | Not occupational qualification, long-term retention, or weld certification. |
| P06 | Apprenticeship evaluation (DOL report) | Registered participation was associated with positive estimated employment and earnings effects. | Quasi-experimental design and possible unobserved selection. |
| P07 | Year Up randomized evaluation (ACF report) | A substantial program package produced sustained earnings gains. | Components were not isolated and it was not a model intervention. |
| P08 | Writing-assistance hiring experiment (NBER) | Assistance increased a defined platform hiring outcome. | No economy-wide net-employment conclusion. |
| P09 | Customer-support workplace study (QJE) | Resolutions per hour increased on average with heterogeneous worker effects. | One occupation and quasi-experimental design; not an automatic wage effect. |
| P10 | Technical-learning experiment (preprint) | Assistance reduced immediate unassisted technical comprehension in a small randomized experiment. | Short, small, vendor-affiliated preprint. |
| P11 | Developer productivity update (METR) | Changing tools, task selection, and time measurement complicate a current uplift estimate. | The update explicitly does not provide one definitive productivity number. |
How to read this table. Packages beat products. The positive results (E01, E02, E06, E09, P07) mostly come from structured programs with people around the technology, and attribution to the model alone is not established. E08 and P10 cut the other way with equal weight: unassisted performance or comprehension fell. P04 is a null, not proof of equivalence. Ask what surrounds the tool, not just which tool it is.
12.2.2 Household and Family Life#
| ID | Evidence program | Principal permitted claim | Primary limit |
|---|---|---|---|
| H01 | American Time Use Survey (BLS) | National time-use estimates provide a household-work baseline. | Observational time diaries, not an assistance effect. |
| H02 | Household-organization survey and five-person pilot (study) | Exploratory evidence identifies perceived uses and implementation questions. | No controlled objective time-saving result. |
| H03 | Blended parenting cluster trial (PLOS Digital Health) | Immediate caregiver-reported learning and violence outcomes improved under a structured digital-human program. | One preschool, self-report, and no long-term untreated comparison. |
| H04 | READY4K trial (CEPA study) | Small, actionable messages improved a defined literacy outcome. | Non-generative program and implementation limits. |
| H05 | SNAP outreach experiment (NBER Working Paper 24652) | Information and human application assistance affected program take-up. | Older eligible population and no general model intervention. |
| H06 | Home Energy Reports experiments (Allcott article) | Feedback caused a small average reduction in household electricity use. | Not a data-center impact or general-purpose model study. |
| H07 | Dementia-caregiver eHealth meta-analysis (JMIR) | A small pooled burden effect coexists with high heterogeneity and a wide prediction interval. | Varied programs and no general-purpose model conclusion. |
| H08 | Dementia-care recommender trial (Age and Ageing) | Expert-reviewed personalization can be evaluated for safety and burden. | Short selected sample and null between-group burden estimates. |
| H09 | Medical self-assessment trial (Nature Medicine) | General users did not reliably convert model access into better condition identification. | Simulated cases and tested 2024-era systems. |
| H10 | Personal-conversation experiment (CESifo) | Unstructured personal conversation produced adverse loneliness evidence in the tested encouragement design. | Working paper, one month, and not specialized therapy. |
| H11 | Parent survey (Lurie Children's) | Parents reported particular uses and perceived time effects. | Self-report adoption survey without causal measurement. |
How to read this table. READY4K (H04), SNAP outreach (H05), and Home Energy Reports (H06) are not model studies; they show what structured information and human help can do, a benchmark newer tools must meet. The rows that test model access are sobering: H09 found no reliable gain in condition identification, and H10 found adverse loneliness evidence. H01, H02, and H11 describe behavior and belief, not causation.
12.2.3 Capability, Work, Civic Life, Finance, and Later Life#
| ID | Evidence program | Principal permitted claim | Primary limit |
|---|---|---|---|
| C01 | Professional writing experiment (Science) | Selected tasks were completed faster with higher assessed quality. | Not annual productivity, earnings, or job effects. |
| C02 | Consulting experiment (working paper) | Benefits occurred on inside-frontier tasks while accuracy fell on an outside-frontier task. | Bounded tasks, model, and working-paper estimates. |
| C03 | Experienced-developer randomized study (METR) | Assistance slowed completion in that early-2025 setting. | Small historical sample and changing technology. |
| W01 | ILO occupational exposure index (ILO) | Occupational task exposure is widespread and varies by category. | Exposure is not predicted displacement. |
| W02 | Danish labor-market study (NBER) | Early earnings and hours effects were near zero within the studied period while work changed. | Working paper, country and period limits, and no claim about future effects. |
| W03 | WorkAdvance ten-year trial (MDRC) | One provider produced a significant year-ten earnings gain while others did not. | Sectoral training, provider variation, and no model intervention. |
| V01 | Habermas Machine experiments (Science) | Generated group statements were preferred in the studied deliberation comparisons. | Not factual correctness, democratic legitimacy, or binding field decisions. |
| V02 | Consultation-analysis evaluation (DfT) | Assisted analysis showed measurable theme and mapping performance with material omissions and false themes. | Task-specific metrics and modeled rather than randomized savings. |
| V03 | MyCity audit and response (NYC Comptroller) | Public-system accuracy and governance require inspectable denominators and evaluation. | Audit dispute remains visible; program cost is not chatbot-only cost. |
| V04 | Participatory-budgeting review (PMC review) | Some settings show suggestive benefits from participatory institutions. | Predominantly nonrandomized, geographically concentrated, and not a model study. |
| F01 | Standardized financial-advice comparison (Journal of Financial Planning) | Recommendations differed materially across tools under common scenarios. | Few scenarios and no realized investor outcomes. |
| F02 | Life-cycle advice simulation (working paper) | Modeled advice can reflect broad principles while failing important contingencies. | Simulated lives and assumptions, not observed retirement wealth. |
| F03 | Pension-allocation experiment (preprint) | Recommendations influenced hypothetical incentivized choices. | No actual portfolio changes or long-term return benefit. |
| L01 | MASAI mammography trial (Lancet record) | A defined radiologist-supported workflow met noninferiority for interval-cancer outcome and had higher sensitivity. | Not reduced mortality or consumer self-diagnosis. |
| L02 | Physician diagnostic-reasoning trial (JAMA Network Open) | Model access did not significantly improve physician scores under the tested design. | Small vignette trial, not patient outcomes. |
| L03 | Older-adult randomized-trial review (BMC Geriatrics) | A small depression effect did not extend to a reliable pooled loneliness effect. | Heterogeneous, short interventions and few loneliness trials. |
| L04 | Older-adult pre-post review (Psychological Medicine) | Favorable within-group changes warrant further study. | Within-group comparisons do not isolate causal effects. |
How to read this table. The capability rows (C01 through C03) show assistance speeding one task, hurting accuracy on another, and slowing experienced developers in a third. W01 measures exposure, not displacement. V02 shows measurable performance alongside omissions and false themes. The finance rows show no realized investor outcome. L01 is a strong workflow trial that still does not show reduced mortality, and the nulls in L02 and L03 sit at full size.
12.2.4 Infrastructure, Governance, and Public Record#
| ID | Evidence program | Principal permitted claim | Primary limit |
|---|---|---|---|
| I01 | LBNL data-center energy report (LBNL report) | National electricity and water estimates establish material planning scale and scenario uncertainty. | Model-based national estimates, not site meter data or actual 2028 use. |
| I02 | Congressional Research Service water review (CRS) | National data sources do not provide a complete isolated inventory of data-center water use. | Policy synthesis, not new facility-level measurement. |
| I03 | Virginia JLARC evaluation (JLARC report) | Virginia faces material potential grid, fiscal, land-use, noise, and planning effects with explicit scenarios. | State-specific findings and forecasts, not universal project effects. |
| I04 | AEP Ohio data-center tariff (AEP Ohio, PUCO bulletin) | A documented tariff uses minimum commitments and risk-allocation provisions for large loads. | Regulatory and contractual record, not proof that every risk is eliminated. |
| G01 | NIST risk-management framework and profile (NIST framework, NIST profile) | Governance should connect context, measurement, management, and identified risk categories. | Voluntary guidance, not certification or deployment-effect evidence. |
| G02 | International AI Safety Report 2026 (report) | The report synthesizes pre-December-2025 scientific evidence on capabilities and risks. | Its cutoff cannot validate September 2026 releases. |
| G03 | Anthropic pacing statement (essay) | Leadership called for pacing while explicitly distinguishing it from halting all progress. | Policy argument and publisher claims, not an independently tested commitment. |
| G04 | OpenAI scaling statement (essay) | Leadership called for safety-conditioned scaling, shared bars, and possible withholding of further scaling. | No formal study, independent audit, or blanket no-release commitment. |
| G05 | September 22 product announcements (Anthropic, OpenAI) | Product updates occurred after public calls for safety-conditioned pacing. | Sequence alone does not establish contradiction or sufficient safety. |
How to read this table. This is my industry's table. I01 is model-based national scale, not site meter data. I02 says no complete national water inventory exists. I03 is Virginia only, and I04 is an allocation mechanism, not proof every risk is gone. G03 through G05 are publisher statements: what an organization says, not whether it did it. G05's sequence of events is neither proven contradiction nor proven safety.
12.2.5 The Seven Savrn Tracker Records#
| ID | Tracker | What the publisher-described record can support | What it cannot support alone |
|---|---|---|---|
| T01 | Capital Atlas | Identification and classification of financing and capital disclosures | Local spending, jobs, tax benefit, or return |
| T02 | Delay Watchlist | Dated evidence of identified delay and related statuses | Cause, permanent failure, or a lost household benefit |
| T03 | Moratorium Tracker | Status and scope of indexed state and local actions | Complete census or cross-jurisdiction applicability |
| T04 | Water Tracker | Categorized water claims, records, and measurement boundaries | Local impact without utility and facility evidence |
| T05 | Grid Operator Watchlist | Authorities, tariffs, dockets, queue, and curtailment records | Energization, household rate effect, or site reliability |
| T06 | Permits and Power Development Tracker | Selected permits and development signals by stage | Construction, operation, revenue, or net benefit |
| T07 | Scarcity Tracker | A modeled signal under published assumptions | Spot price, liquid settlement, or guaranteed revenue |
How to read this table. The trackers are not studies, so the columns change. Savrn publishes them and has a commercial interest in what they cover. A tracker entry starts a question and never ends one: an announced dollar is not a local job, and a queue record is not energization. The four-layer chain in Chapter 11 runs from tracker record to local record to causal or allocation analysis to decision; skipping steps produces confident, unsupported conclusions.
12.2.6 What the Register Supports#
The evidence supports conditional, bounded conclusions and keeps favorable, null, mixed, and adverse findings side by side. It does not establish that every proposed workflow works, or that a useful application automatically justifies a facility. Capability is not benefit. The next step is implementation testing under the shared protocol of Chapter 11, not stronger adjectives.
12.3 Glossary of AI Research Terms#
Assistance, delegation, substitution. Defined in the Terminology and Conventions section of The Research Thesis. Assistance: a system prepares, drafts, or explains under human review. Delegation: a human authorizes a specific action. Substitution: the system performs a function previously done by a person without per-case authorization.
Behind-the-meter. Power supplied on the customer's side of the utility meter rather than drawn entirely through the public grid.
Confidence interval. The range of values consistent with the data at a stated confidence level. A 95 percent interval spanning benefit and harm, such as the loneliness estimate's -2.57 to 1.23, means the study cannot tell them apart.
Denominator discipline. Reporting every figure with its population, comparator, and limit in the same breath.
Difference-in-differences. A quasi-experimental method comparing how an outcome changed for a group that got something against a group that did not, over the same period. P09 used it. It works only if both groups would otherwise have moved in parallel.
Effect size (Hedges' g, Cohen's d). Differences between groups in standard-deviation units, so studies on different scales can be compared. Values near 0.2 are typically called small.
Evidence class. One of eight editorial labels, from randomized comparison through proposed workflow, defined in Chapter 11, stating what evidence can and cannot establish.
F1 score. The harmonic mean of precision and recall. The 0.59 F1 in the DfT theme test (V02) reflects weakness in both directions at once.
Forbidden inference. A cross-layer conclusion the evidence does not support, such as announced capital equaling local jobs.
Heterogeneity. Variation in results across people, settings, or studies. High heterogeneity (as in H07) means one pooled average hides very different effects.
Intention-to-treat. Comparing groups as randomized, regardless of what participants actually used. Analyzing only engaged users can manufacture effects.
Interconnection. The process and agreement for connecting a large generator or load to the grid, usually through a queue and studies. A queue position is not energization.
Interval cancer. A cancer diagnosed between scheduled screening rounds; the MASAI trial's primary outcome (L01).
Noninferiority. A design testing whether a new approach is no worse than a standard by more than a preset margin. Meeting it is not demonstrating improvement.
Pass-through estimate. In the pension experiment (F03), how much a recommended allocation change showed up in participants' choices: 0.368, 95% CI 0.303 to 0.433, in hypothetical incentivized choices.
Percentage point. The plain difference between two percentages. Moving from 10 to 15 percent is 5 percentage points, or a 50 percent relative increase.
Precision and recall. Precision is the share of flagged items that are correct; recall is the share of correct items that were flagged. Recall of 0.75 means a quarter of validated themes were missed.
Prediction interval. In a meta-analysis, the range where a new, similar program's effect would be expected to fall. A wide one (as in H07) means a small average benefit promises nothing in your setting.
Preprint. A paper posted publicly before, or without, journal peer review. E05, P10, and F03 are preprints.
Scenario, forecast, simulation. A conditional result under stated assumptions; a planning input, not an observation. The LBNL 2028 electricity range and the JLARC 2040 bill range are scenarios.
Sensitivity and specificity. Sensitivity is the share of true cases detected (MASAI: 80.5 percent supported versus 73.8 percent standard); specificity is the share of non-cases correctly cleared (about 98.5 percent in both arms).
Standard deviation. How spread out values are around their average; the unit effect sizes are measured in.
Tariff. A utility's regulator-approved schedule of rates and terms for a customer class, such as the AEP Ohio data center tariff (I04). Tariffs can change by commission order.
The seven Savrn trackers. The Capital Atlas, Delay Watchlist, Moratorium Tracker, Water Tracker, Grid Operator Watchlist, Permits and Power Development Tracker, and Scarcity Tracker: an evidence-discovery instrument.
TWh (terawatt-hour). One billion kilowatt-hours. LBNL estimated about 176 TWh of data center use in 2023, with 2028 scenarios of roughly 325 to 580 TWh.
Withdrawal and consumption (water). Withdrawal is water taken from a source; consumption is water not returned. Both must be distinguished from discharge, permits, and modeled estimates.
Within-group comparison. Change in one group over time with no control group. It cannot separate the intervention from time, attention, or regression to the mean (see L04).
Working paper. A paper circulated by an author or institution such as NBER or CESifo before or instead of journal publication. Estimates can change.
12.4 Source Verification: How I Checked Every AI Research Source#
12.4.1 The Verification Method#
On September 23, 2026, every distinct URL in the report's citation ledger received an automated browser-like request, with redirects followed and a 25-second timeout, plus manual spot-checks of representative blocked sources and one replacement search. Three rules governed drafting:
- Blocked sources are cited with an access note, and no claim may depend on content that could not be inspected directly or through an independent fetch path.
- Pages that change after the cutoff are quoted as of the cutoff.
- A source that fails verification is logged and downgraded to "as cited in the September 23, 2026 edition," never silently dropped or upgraded.
12.4.2 Verification Results#
| Result | Count | What it means |
|---|---|---|
| HTTP 200, accessible | 64 | Page retrieved; content available for inspection. |
| HTTP 403, blocked or paywalled | 17 | Live pages that block automated access (major publishers, government sites, OpenAI). Spot-checks confirmed content matches; OpenAI "Introducing Superalignment" was verified verbatim. Cited with an access note. |
| HTTP 203, nonstandard success | 2 | PubMed record for MASAI (PMID 41620232) and JMIR e78568; both live. |
| Recovered on retry | 1 | ERIC EJ963694 (kindergarten reading trial) returned HTTP 200 on retry. |
| HTTP 404, dead link | 1 | SNAP outreach manuscript on economics.mit.edu (truncated filename). Replaced. |
12.4.3 The One Correction: SNAP Outreach Experiment#
The SNAP outreach experiment (H05) originally linked to a manuscript on the MIT economics site that returned a 404. I replaced it with the canonical location of the same study: Finkelstein and Notowidigdo, "Take-up and Targeting: Experimental Evidence from SNAP," NBER Working Paper 24652, later published in the QJE.
The full text was retrieved and figures confirmed. The study covered about 30,000 elderly individuals likely eligible for SNAP. Over nine months, enrollment was 5.8 percent in the control group, 10.5 percent with information only, and 17.6 percent with information plus application assistance. The verification log rounded those to 6, 11, and 18 percent; this report uses the figures as published. Chapter 6 cites the NBER record as a working paper with later QJE publication.
12.5 Sponsorship and Independence Disclosure#
In my own words: this research was prepared for Savrn, the company I run. Savrn operates in an industry this report examines and publishes the seven trackers used throughout. Read every page with that in mind. What I did about it was keep the unfavorable findings at full size and label Savrn's own design goals as publisher statements. What I can't do is certify my own independence. Nobody can.
Formal statement. This report was prepared for Savrn and concerns an industry in which Savrn has a commercial interest, including through the seven Savrn trackers described throughout. Savrn's publisher role and commercial interest are stated wherever tracker findings support an argument, consistent with the public-claim controls set out in Chapter 11. This report does not describe itself as an independent institutional review. No outside peer-review panel is represented as having approved it. The proposed publication policy, preservation of material unfavorable findings even when they weaken a commercial narrative, has been applied throughout the drafting. The sponsor-conflict review listed among the publication gates in Chapter 11 is an external requirement for any institutional use of this report, not a step this report can perform on itself.
12.6 Evidence-Cutoff Statement#
The evidence cutoff for this edition is September 23, 2026. All findings, product references, regulatory records, tracker statuses, and URLs reflect that date. The cutoff is a boundary, not a warranty: linked pages may change, records may be corrected, and later evidence is not incorporated. Product configurations, tariffs, permits, and live project status should be rechecked against original records before consequential use.
12.7 Corrections Policy#
The refresh policy uses triggers rather than pretending every source goes stale on the same schedule. Product configurations, tariffs, permits, and live project status should be rechecked before consequential use; the AEP Ohio tariff in Chapter 10 is exactly the kind of record that can change by commission order. Historical trials remain historical records, but they should be checked for corrections, retractions, and material follow-up before their estimates are quoted as current.
Each correction should identify the affected claim, original wording, corrected wording, reason, and version. A material correction should propagate to summaries and outreach content, not stay hidden in the longest document.
12.8 About the Author#
Chad Everett Harris is the Founder and CEO of Savrn. He is a serial infrastructure entrepreneur based in Dallas. He has spent his career building large power infrastructure. He now builds Savrn, and he works from an "operator first" philosophy: no infrastructure without customers, and vertical integration as a defense.
12.9 About Savrn#
Savrn designs, builds, and delivers AI factories, purpose-built data centers. Its design is built around behind-the-meter power and closed-loop cooling with a zero-makeup-water design goal. Those are publisher statements, held to the same evidence test as every other organization's claims. Savrn also publishes the seven Savrn trackers listed in the register.
12.10 Frequently Asked Questions#
What is an AI research evidence register?
It is an index of every major study and record behind this report. Each row gives the source link, the narrowest claim the evidence permits, and its primary limit. Exact estimates and denominators stay in the chapter that interprets each study, and one study cited in several chapters is still one study.
How were the sources in this AI report verified?
Every distinct URL in the citation ledger got an automated browser-like check as of September 23, 2026, plus manual spot-checks of blocked sources. Sixty-four loaded directly, 17 were blocked but confirmed another way, 2 returned a nonstandard success code, 1 recovered on retry, and 1 was dead and replaced.
What was the SNAP study correction?
The dead MIT economics link now cites NBER Working Paper 24652 by Finkelstein and Notowidigdo. Among about 30,000 elderly people likely eligible, nine-month enrollment was 5.8 percent for controls, 10.5 percent with information, and 17.6 percent with information plus application help. It tested human assistance, not an AI model.
Who sponsored this AI research report, and why does it matter?
No. It was prepared for Savrn, which has a commercial interest in the data center industry the report examines and publishes the seven Savrn trackers. It is not an independent institutional review, no outside peer-review panel approved it, and an external sponsor-conflict review is required before institutional use.
Why are the Savrn trackers in the evidence register?
Because the report uses them, and anything the report uses belongs where you can check it. The tracker table lists what each record can support and what it cannot support alone. An announced dollar is not a local job. A tracker entry starts a question and never ends one.
What does the evidence cutoff date mean?
The cutoff is September 23, 2026, and findings, records, tracker statuses, and URLs reflect that date. It is a boundary, not a warranty: pages can change and later evidence is not included. Recheck tariffs, permits, product configurations, and live project status before any consequential decision.
