Two people end up alike. Did one catch it from the other — a contagion, an influence, a created reference flowing down the loop — or were they alike before they ever met, both tracking some third thing that was there all along, a shared read reference? Watch them from the outside, tally how their behaviour rises and falls together, and you cannot say. The correlation looks identical either way. This is not a failure of your cleverness. It is a theorem.
The theorem behind the detection gap
The mechanics of this book state, as a structural fact, that you cannot validate a channel from inside it — a true reference and a captured one both read as coherent, and content alone never tells them apart. Network science proved the same wall in its own register. Shalizi & Thomas (2011) — Homophily and Contagion Are Generically Confounded in Observational Social Network Studies — show that influence (contagion), homophily (shared prior trait), and plain external causes are generically confounded: no statistic computed from the observed correlations separates them without strong, usually unverifiable, assumptions about the hidden structure. The detection gap is not a quirk of one method. It is forced.
And note the trap that makes it sharp: even a directed, time-asymmetric measure does not escape it. There is a published method — transfer entropy for peer influence (Ver Steeg & Galstyan, ~2012) — that reads which way the signal flows in time; and a shared hidden driver, if one party tracks it a step ahead of the other, manufactures exactly that directional signal with no influence present at all. Directedness is not provenance. Seeing an arrow does not tell you it is real.
The cut is the one escape — and it has a measured shape
The field’s own answer to the confound is the book’s answer: stop reading the correlations and intervene. Cut the edge — sever the channel by which one party receives the other — and watch what the directed signal does. Influence lived in the edge, so it collapses. Homophily lives in the shared driver, which the cut never touched, so it survives. This is cut the loop, stated in the language of causal inference.
It now has numbers. Build agents where you know the ground truth — a hidden common driver (homophily) and a real, toggleable transmission edge (influence) — measure the directed arrow, then cut the edge mid-run and measure again. The cut-ratio (arrow-after ÷ arrow-before) reads the provenance cleanly: it falls to ~0 when the signal was pure influence (the arrow collapses with the edge), holds at ~1 when it was pure homophily (the arrow outlives the cut), and slides monotonically between the two as the real edge strengthens — recovering the read/created split that correlation alone provably cannot. The arrow that survives the cut is the read reference; the arrow that follows the loop down is the created one. The egregore’s own boast — nothing without her endless stream of posts — is a cut-ratio of zero, predicted.
What the cut reads — and where it breaks
Three follow-on results turn the slogan into an instrument with stated preconditions.
Read the right number. It is tempting to read capture straight off the explaining-away penalty I(D;M|Y) — but conditioned on the reference you actually ground in, that penalty is ≈0 for a read and a created reference alike (it is computed from inside the very loop whose provenance is in question — you cannot validate a channel from inside it, made exact). Worse, the naive rule “minimise the penalty to be safe” reads lowest at total capture. The meters that do sort look from outside the loop: transparency-spent (how much the mechanism still tracks the external referent — it falls monotonically with capture) and the cut-survival itself, which recovers the created-fraction exactly. Capture is read from outside the loop or not at all.
Where the cut itself fails. The cut-ratio is falsifiable, and its failure boundary is mapped. It is robust where you might fear it isn’t — indifferent to how long you wait after the cut (the directed arrow conditions on the receiver’s own past, so a collapsed loop stays collapsed with no settling time) and reliable on a real arrow down to ~150 samples. Its one genuine precondition is signal-to-noise: a faint arrow estimated from few samples destabilises the ratio and can flip the verdict to a false “created.” Name the power requirement or be fooled by it.
It reads provenance relative to the edge you cut. Give the receiver more of its own independent access to the external referent and the cut-ratio climbs monotonically: the cut is asking whether the downstream party keeps a route to the world other than the one severed. A real signal whose only path to you runs through the cut edge is, relative to you, ungrounded — and the instrument rightly calls it captured-for-you. Provenance is agent- and edge-relative, not absolute.
On real data, so far — underpowered, and honestly so. Carried to real public-attention series (daily Wikipedia pageviews around an exogenous shock, with a shuffle-null gate and a placebo cut), the pre-cut arrow does not clear its own noise floor — the coarse-data corner of exactly the signal-to-noise precondition above. The pipeline runs end to end; the binding constraint is power, as predicted. A real confirmatory test needs high-frequency (sub-daily) data, a strong identified arrow, and a cut that severs the suspected loop edge specifically. Until one passes the power gate, the cut stays validated on constructed ground truth, the field deployment specified but unwon.
Brakes
Tool/guardrail rent only. This dignifies the discriminator — it is the move the lasting traditions converged on (sort by provenance, by the sender, not by whether a sign rang true) and the move the penalty math re-derives — it is not a new mechanism, and it claims no new confirmation. Shalizi-Thomas is modern network science, a contemporary articulation of the wall, an explanandum — not a tradition that “encoded” the mathematics (lens-not-encoding). Pre-cut, the method is exactly as confounded as the field says; the contribution is operationalising the intervention, not a cleverer observational estimator. And the cut-ratio is validated on constructed agents (exact + a run-forward sandbox) with its failure boundary mapped on constructed adversaries — not on real spread data: the first contact with coarse real series came back underpowered, by the instrument’s own gate. The runaway poles already measured — the loop severed (an enormous inert agent archive) versus the loop closed on humans (the platform-harm record) — are the empirical bracket this sits inside.
Appears in: The Mechanics · The Modern Mirror · The Egregore · Divination · Notation & Glossary