The numbers, and where they come from.
Every figure on this site carries a footnote. This page is the footnotes.
F1Review reality
Global industry survey (McKinsey, State of AI): 27% of organizations using generative AI review all AI-created content before use; roughly as many review at most a fifth. Full review is documented standard in healthcare (the clinician reviews and signs every AI-drafted note before it enters the record) and in US securities (a registered principal pre-approves retail communications; the approver’s name and date are archived; firms are fully responsible for AI-generated content). In July 2026 the regulator, FINRA, proposed risk-based selection instead of pre-review of everything, explicitly because AI volumes have made full review unmanageable.
F2The middle lane and the error-cap span
5–10x (the middle lane’s escalation saving) and the twenty-fold error-cap span: measured in replayed tests on real model data (July 2026), at the same measured error cap.
F3Model-swap detection
Model swaps were detected in testing within a median of 36–59 responses; the step-down covers detected swaps and gross shifts. Silent drift without ground truth is the meter’s job (see For insurers & brokers).
F4The guard caught our own error
One of our calibrated limits let through 1.6x more alarms than promised; the guard’s alarm was correct, and the correction is computed and confirmed in a follow-up control study.
F585–94% automation
85–94%: measured automation share at error caps of 1% and 5% respectively, in replayed tests (July 2026) on two open language models (70–72 billion parameters) and public test data. Your number depends on your case mix and is measured in a pilot.
F64.0–4.4% against a promised 5.0%
4.0–4.4% against a promised maximum of 5.0%: measured share of false alarms in four separate test flows on real models, with statistical margins reported.
F7The missing record
Industry analyses and the reading of the EU’s automatic event-logging requirements: operator-mutable application logs do not hold up as strong evidence; NIST’s framework for AI agents defines the record required instead: who approved an action, under which policy, when, with what outcome.
F8EU AI Act timeline
Adopted and in force: Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026, in force since 27 July 2026. Transparency obligations apply from 2 August 2026; the Article 50(2) machine-readable marking duty is deferred to 2 December 2026, with grace for systems already on the market. High-risk obligations carry fixed dates: 2 December 2027 for Annex III stand-alone systems, 2 August 2028 for Annex I embedded systems; the earlier conditional trigger mechanism was dropped.
F9The insurance market
Specialist programs at Lloyd’s write up to $25M per organization (early 2026); insurers already back AI performance warranties for vendors; the largest European insurers do not yet have their own AI liability lines.
F10The meter
Two studies in simulation on frozen real model data: 60,000 runs across 72 conditions, then confirmed in an independent repetition with corrected reference flows (24 control conditions). Medians 141–157 against a computed information-theoretic floor of 145; false alarms 1.7–4.0% against a 5% budget, versus 57% for the check-every-time method.
Questions about a number?
We're happy to walk through any figure, method or margin in detail.