The lecture opens with a brief, deliberate disclosure: its subject connects closely to the Evolutionary Intelligence Environment project, which works on infrastructure for machine auditing of complex systems. The disclosure is made openly at the very start, not tucked into a footnote — the lecture is explicit that it is examining general principles, not promoting any specific project, and that the same critical framework applied to every other example will be applied to this one too, in a dedicated block near the end.

The opening proper poses a single image: an inspector tasked with manually verifying a system processing millions of transactions a second — then asks whether the same limit applies once "transactions" becomes "decisions made by a complex public institution." The first substantive block answers with Charles Perrow's theory of normal accidents: systems that are simultaneously complex, tightly coupled, and carry catastrophic potential generate accidents that no amount of added caution can engineer away. Digital-era public procedures increasingly carry all three properties at once, which means classical periodic human oversight runs into a structural limit, not a fixable shortage of funding or training. A second, compounding limit follows: automation bias, the tendency of a human overseer to defer to an automated system's output even against contrary evidence — a bias that "explainability" alone does not resolve, since human cognitive limits on tracking complex causal chains are separate from mere access to information.

The lecture then turns to machine audit itself — already an existing, if still immature, practice, driven by regulation such as the EU's Digital Services Act. Its advantage follows directly from the previous block: no fatigue, no attention degradation, consistent criteria applied at a scale no human team could match. But a sobering empirical counterweight follows immediately: a large field study of practitioners across nearly two hundred organizations found the single greatest obstacle to good algorithmic audit was not technical — it was the audited party's simple refusal to consent to genuine review, connecting directly back to the previous lecture's account of institutional resistance. Machine audit expands technical capacity; it does not, by itself, create the institutional will to be checked.

The core block states the lecture's central question with full precision: if the auditor is itself a complex system, what guarantees it isn't itself biased or manipulated? Adding a second auditor to check the first only reproduces the same question at one remove — an infinite regress that more layers of the same kind of review cannot halt. A real, recent field study titled "Who Audits the Auditors?" documents this as an observed feature of current practice, not a hypothetical risk. The lecture's answer comes from a more mature, structurally parallel field — financial audit, and the model of the PCAOB: the regress is halted not by adding overseers but by structural independence — an auditor selected and paid through channels independent of the party being audited, subject to mandatory rotation and real disciplinary accountability. The lecture draws an explicit parallel to Kerckhoffs's principle from Lecture 4: reliability comes from the architecture of incentives, not from stacking additional secrecy or additional review.

The lecture then returns, as promised, to Evolutionary Intelligence Environment — describing its "regulatory loop" for continuous audit of autonomous AI agents, and applying to it, without exception, exactly the same questions posed of the PCAOB model: who funds the loop, does that funding create dependence on the companies it certifies, and is the loop itself subject to external rotation and disciplinary accountability. No project, the lecture insists, is exempted from these questions by good technical execution or good intentions alone — including the one named at the outset.

The lecture closes by tying its four theses together — the structural limits of human audit, the real but partial gains of machine audit, structural independence as the actual answer to "who audits the auditor," and that same standard applying to any specific audit infrastructure — and opens onto the question that carries into the next lecture: who decides which institutional model of audit applies where, and whether one single global model is even desirable.