Independent research · Open evidence

Research you can inspect

Independent audits, empirical registers and live tools. Each entry leads to the method, evidence or versioned record behind it.

TOOL 01 LIVE BOARD

The AI News Board and Sovereign Watch

A live, source-tiered front end for AI industry coverage. It separates reports by source class, shows claim-relative evidence and denominator checks, and links relevant portfolio research without treating the link as verification. Daily machine labels remain visibly unreviewed until checked by a person. Failed or deferred assessments remain visible, and incomplete coverage is reported. The publication path stops on invalid data or a failed verification step; this does not establish comprehensive coverage or classification accuracy.

Read the operations case study: incomplete coverage, recovery and explicit review boundaries.

TOOL 02 · DOI 10.5281/zenodo.21819102 PUBLIC-EVIDENCE EXPLORER

Article 50 Transparency Explorer

The corrected v1.1 paper preserves a historical public-evidence snapshot of twelve consumer AI products. It identifies unresolved cells, capture-state limits and the difference between a bounded negative search and an established absence. The linked explorer retains the historical scores. This is not a current compliance assessment or a completed repeat survey.

Vanilla JS · historical scored set · frozen CSV with declared exceptions · build_tool.py
TOOL 03 · DOI 10.5281/zenodo.20559430 DEFLATION CALCULATOR

Announced vs. Deliverable AI Power Demand

How much announced generation and data-centre load actually get delivered. A deflation calculator runs announced generation through measured PJM-queue attrition stages, and the load side displays operator-vetted PJM and ERCOT large-load screens with a two-rail envelope (≈20% floor to ≈48–63% ceiling).

TOOL 04 · DOI 10.5281/zenodo.21318960 WATER EXPLORER

AI Data-Centre Water Evidence Explorer

Which water figures are measured, which are planning assumptions, and what kind of water is involved. Reporting years and metrics are selectable, every site carries its evidence status, and unknown water sources stay visible rather than inferred.

TOOL 05 · ISSUE 6 · DOI 10.5281/zenodo.20541643 UNIT ECONOMICS

AI Cost Watch

Whether the AI buildout's unit economics are holding. Per-edition signal and developments, or one indicator tracked across issues. The signal it watches, a hyperscaler capex-guidance down-revision, has not fired in six issues.

TOOL 06 · DOI 10.5281/zenodo.20572928 FORECAST MATRIX

AI Energy-Demand Forecast Scorecard

How the field forecasts data-centre electricity demand. Dispersion is measured only within comparable slices, since units and scopes are not interchangeable, and each forecast is traced through a transparency funnel from verified to confirmable to reproducible.

TOOL 07 · DOI 10.5281/zenodo.20680108 GOVERNANCE AUDIT

CEO Pay-vs-Delivery Scorecard

A dated comparison of SEC Summary Compensation Table totals and Compensation Actually Paid. CAP is a fair-value remeasurement, not cash received. Board-target cases have separate coverage and are shown alongside peer-relative returns; the comparison does not establish that a CEO caused a return.

TOOL 08 · DOI 10.5281/zenodo.21543579 MODEL DIAGNOSTIC

Open-Weight Error-Detection Audit

The corrected v1.1 release distinguishes model-run provenance from answer checking in financial data extraction. Selected calculations and tables are reconstructed from stored outputs, with a corrected source label and explicit historical input limits. It reports open-weight experiments only; no new model runs or deployment validation are claimed.

TOOL 09 · DOI 10.5281/zenodo.21740081 DATASET DIAGNOSTIC

FineWeb-Edu Threshold Explorer

The filter behind a widely used training corpus keeps documents scoring 3 or more, and that 3 is a rounded integer. The raw cut is adjustable across the 233 score values observed in this sample, the rounding rule can be switched to move 458,461 documents across the boundary, and the rounding identity is shown in the page. The explorer uses corrected content-only token counts. Its historical fixed-score panel remains unavailable pending recalculation; the historical PDF and dataset are unchanged.

OPEN METHOD · v0.2-rc2 RESEARCH PROTOCOL

Divergence Registry

A six-stage research protocol and deterministic gate for recording scope choices, source custody, claim types and departures across independent runs. It makes process divergence inspectable; it does not establish that a finding is correct.

OPEN SOURCE · SOURCE ALPHA LOCAL INFRASTRUCTURE

PLAG IN, Private Local AI Gateway

A Linux source alpha for small organisations that need evidence of locality and reproducible model operation. It provides exact local model identity, bounded embedded libllama serving, authenticated loopback transport and metadata-only inference receipts. The accepted result covers one synthetic L3 trial, not production readiness, repeated-run reliability, model quality or comparative performance.

DOI 10.5281/zenodo.21325372 PUBLISHED 18 AUG 2026 EVIDENTIARY AUDIT

The AI Automation Claims Scorecard

A motive-neutral referee for the "AI writes most of our code" claims. It grades twelve vendor and adopter statements on four evidentiary tests: whether a denominator is stated, whether the figure is externally audited, whether an outsider could reproduce it, and whether a named party with no stake has checked that specific figure. One of twelve defines a rigorous denominator, none is audited, none is reproducible, and two carry an independent check.

POLICY ARTICLE AUGUST 2026 GRID REFORM

Announced against financed: Britain's data centre grid queue

The UK counterpart to the infrastructure analysis. Britain's electricity demand connection queue grew from 41 GW to 125 GW in seven months, with roughly 73 GW in data centres. Evaluates the gap between speculative announcements and binding final investment decisions.

Ofgem demand connections · draft London Plan 2026 · Planning Inspectorate casework
DOI 10.5281/zenodo.20512703 PUBLISHED 08 JUL 2026 ENERGY BOUND

How Much Electricity Does Bot Web Traffic Actually Use?

A transparent order-of-magnitude bound on the electricity used by automated web traffic. A naive multiply yields 257 TWh a year; two distinct methods bracket the figure at order 10 TWh, exposing the parametric sensitivity of traffic-to-energy models.

DOI 10.5281/zenodo.20763480 PUBLISHED 18 JUL 2026 RELIANCE TRILOGY

AI Reliance: how deep, and how reversible

A reproducible measure of systematic AI dependency across four layers (reliability, downstream attribution, penetration, reversibility), with a dated off-switch register and demand-side cohort analysis from provider telemetry and government data.

Python stdlib · Provider status JSON · OMB / GAO data
DOI 10.5281/zenodo.20807278 PUBLISHED 19 JUL 2026 CONSUMER RELIANCE

Consumer Reliance Architecture

How individual AI dependency is built, priced, and made hard to leave: a reproducible profile of the commercial machinery (pricing tiers, caps, metering, portability) across US, EU, Chinese and companion providers, with a revealed-elasticity test of user retention.

Python stdlib · Public pricing pages · Google Trends
DOI 10.5281/zenodo.20941003 PUBLISHED 26 JUN 2026 HOUSEHOLD INCIDENCE

The Consumer Incidence of the AI Buildout

A reproducible ledger of how the AI data-centre build-out reaches a single household (device prices, AI-laden subscriptions, electricity, public subsidy and retirement exposure), with each channel's AI-attributable share bounded across two archetype households.

Python stdlib · EIA · PJM IMM · BLS · Good Jobs First
DOI 10.5281/zenodo.20790896 PUBLISHED 20 JUL 2026 AGENTIC COMMERCE

Who Does Your Shopping Agent Work For?

An evidence review of agent-mediated retail: why the neutral agent is a structural fallacy, who controls downstream purchase funnels, and which consumer protection boundaries remain legally enforceable.

Evidence review · Attack-surface matrix · Argumentative paper
DOI 10.5281/zenodo.20788844 PUBLISHED 19 JUL 2026 CONSUMER PROTECTION

Designed, Not Accidental

An empirical survey of commercial choice architecture and deceptive design patterns in digital retail environments, mapping why individual consumer vigilance fails against systematic behavioural steering.

Evidence review · Enforcement casework · Argumentative paper
DOI 10.5281/zenodo.20778968 PUBLISHED 04 JUL 2026 LABOUR ECONOMICS

AI-Literacy Load by Occupation

A reproducible map of where AI-fluency becomes part of the job: occupation-level AI-usage penetration crossed with the augment/automate mix, derived from the open Anthropic Economic Index.

Python stdlib · Anthropic Economic Index · O*NET data
DOI 10.5281/zenodo.20586863 UPDATED 23 SEP 2026 · v7.1 CAPEX INFRASTRUCTURE

AI Infrastructure: Commitments, Delivery and Access

A dated assessment of selected commitments, delivery stages and access obligations. The accompanying bundle recalculates named balances, rebuilds selected numerical content and tables, and lets readers inspect a quantity with its source and qualifications. These checks do not establish source truth or a causal account of the infrastructure build-out.

DOI 10.5281/zenodo.20672581 PUBLISHED 14 JUL 2026 VALUATION BENCHMARK

The AI Private-to-Public Ledger

A pre-registered benchmark scoring private AI venture marks against primary SEC filings, IPO pricing mechanisms, and post-listing lock-up expirations.

DOI 10.5281/zenodo.22015796 PUBLISHED 19 AUG 2026 INDUSTRY SCAN

The AI Industry: Architecture, Chokepoints, and Deflation (v2.1)

A dated assessment of infrastructure constraints and investment claims. Version 2.1 adds an offline reproduction bundle to the v2 analysis. The deposit identifies unresolved sources and calculations included only in the bundle; the reproduction result is not verification of every source claim.

Python stdlib · Primary regulatory and trade filings

About the evidence

Research Records
Versioned
Versioned records; reproduction coverage and unresolved checks are stated per project
Jurisdiction Evidence
Source-linked
Versioned local packs; custody checks do not establish legal correctness or complete coverage
Compensation Evidence
SEC filings
Dated issuer universe; CAP, cash received and board targets are distinct measures
Documents Audited In One Filter
84M
FineWeb-Edu, across 94 Common Crawl snapshots