/* ============================================================ The three deep pages behind the landing: /evidence — every case with its five-question verdict /verify — the method, in the two registers at once /limits — what we cannot do, with the numbers ============================================================ */ function PageHead({ eyebrow, title, lead }) { return (
{eyebrow}

{title}

{lead}

); } /* ---------------------------------------------------------------- */ /* --------------------------------------------------------------- THE CASE, ON ONE SCREEN (Mo, 16.09.). The film carries the story and the only headline on this page; this column carries the few numbers that matter, each with an icon instead of a sentence. It must fit without scrolling, so every line here earns its place. */ const CaseIcon = { Voice: ({ className = '' }) => ( ), Link: ({ className = '' }) => ( ), }; function RoleTag({ children }) { return ( {children} ); } function CaseFigure({ icon, v, k }) { const I = icon; return (
{v}
{k}
); } function CaseResult({ r }) { return (
{r.v} {r.b}
after {r.t} · {r.s}
); } /* The five checks, in place — a dialog, not a second page. */ function CheckDialog({ onClose }) { const { FIVE } = window.GD; React.useEffect(() => { const esc = e => { if (e.key === 'Escape') onClose(); }; window.addEventListener('keydown', esc); return () => window.removeEventListener('keydown', esc); }, [onClose]); return (
e.stopPropagation()} className="w-full max-w-[600px] max-h-[82vh] overflow-y-auto rounded-2xl bg-paper p-7" style={{ boxShadow: '0 24px 60px rgba(0,0,0,0.28)' }}>

Before a finding may be shown, it has to survive five questions.

    {FIVE.map(f => (
  • {f.q} {f.ask}
    • ))}

Ten dated events went through it. Four came through, three left an indication, two moved nothing at all, one was too small to judge — and all four kinds are published.

); } function CaseAutomotivePage() { const { CASE_AUTO, CONTACT } = window.GD; const [checks, setChecks] = React.useState(false); const figures = [ { icon: CaseIcon.Voice, v: CASE_AUTO.kpis[0].v, k: 'customer reviews' }, { icon: Icon.Car, v: CASE_AUTO.kpis[1].v, k: 'dealerships' }, { icon: Icon.Settings, v: CASE_AUTO.kpis[2].v, k: 'causes you can steer' }, { icon: CaseIcon.Link, v: CASE_AUTO.kpis[3].v, k: 'of relationships hold' }, ]; return (
CASE STUDY · AUTOMOTIVE AFTER-SALES

We read what customers write —{' '} and find the causes you can steer.

Public reviews from car dealerships, put through our causal AI pipeline — and one measure followed from the workshop floor all the way to the money.

{checks && setChecks(false)}/>} ); } function VerifyPage() { const { FIVE } = window.GD; return ( We try to disprove our own findings. Five times.} lead="A child should understand every visible sentence; a data scientist should find nothing to take apart. The two registers say the same thing here — the simple version may simplify, never overstate." />
THE FRAME, IN ONE PARAGRAPH

We measure the way a natural experiment works: sites where something happened, against comparable sites where nothing happened — same brand, same market, same quarter. Then we try to disprove the result five times over. Does the difference count? Could it have been luck? Are we measuring something else, or ourselves? Does it show up more than once? Do we understand why?{' '} What survives that, we call a demonstrated effect.

THE FIVE QUESTIONS

Each one can fail a finding on its own.

{FIVE.map((f, i) => (
{String(i + 1).padStart(2, '0')} · {f.q}
{f.ask}

{f.detail}

EXAMPLE{f.example}

))}
THE LADDER

Four rungs. Never one higher than earned.

Every result is placed on exactly one of these, in the demo, in a report and in a sales conversation. The wording is fixed, so nobody has to guess how strong a statement is meant to be.

THE OBVIOUS OBJECTION

“Is this proven the way a laboratory proves things?”

No — and we do not claim it is. In a laboratory you randomise the treatment; nobody can randomise which dealerships rebuild their premises. So we randomise the test instead: the same finding is recomputed in hundreds of random cohorts without the event, one year before the event, without the reviews that mention the event itself, and in other countries. A false finding fails one of those — and ours did: two measures came back with nothing, and one case failed to replicate abroad. Where the outside world keeps a register, we check ourselves against it: an official recall wave shows up in our data at the predicted dose. This is not a laboratory. It is what a court would call a chain of evidence, with the thresholds written down.

); } /* ---------------------------------------------------------------- */ function LimitsPage() { const { LIMITS, EXPLAIN } = window.GD; return ( State the edge, earn the middle.} lead="This page exists because almost nobody can name these numbers, and most people do not know they exist. Whoever states the limit of their model has also told you where to trust it." />
{LIMITS.map((l, i) => (
LIMIT {String(i + 1).padStart(2, '0')}

{l.t}

{l.b}

))}
THE NUMBER ALMOST NOBODY PUBLISHES

How much of what happens is explainable at all?

{EXPLAIN.intro}

OUTCOME CEILING WHY IT IS NOT 100%
{EXPLAIN.rows.map(([t, ceil, why]) => (
{t} {ceil} {why}
))}

Read every explainability figure against that ceiling, never against 100%. Half to three quarters of quarter-to-quarter movement is sampling noise — no model should explain it, and none can.

{EXPLAIN.claims.map(([n, b], i) => (
{n}

{b}

))}
WHY THIS IS AN ARGUMENT

A product that closes its own data gap.

Under 1% of the movement traces back to measures we can name today. That is not a verdict on the measures — the demonstrated ones work — it is a statement about visibility: we only see what leaves a trace in what customers write. Every measure a customer records in the loop moves from invisible medication into the tested catalogue. The number on this page is the one we expect to change, and we would rather you watch it move than never have seen it.

); } /* ---------------------------------------------------------------- */ window.CaseAutomotivePage = CaseAutomotivePage; window.VerifyPage = VerifyPage; window.LimitsPage = LimitsPage;