The block universe and the problem of change in cybersecurity

A paper from 1908 and a lecture in the same year

J. M. E. McTaggart published "The Unreality of Time" in 1908, arguing that the ordinary description of events as past, present or future was incoherent — every event is all three, depending on when you ask, which he took as contradiction. In the same year, Hermann Minkowski stood up in Cologne and told his audience that space and time, taken separately, were about to fade into mere shadows, with only a union of the two surviving. He was recasting Einstein's 1905 kinematics as a single four-dimensional geometry. Between them, McTaggart's B-series — events ordered merely as earlier and later, with no privileged now — found a physical home. C. D. Broad's growing block followed in 1923, adding events to the geometry as they occurred. Donald Williams, in 1951, called the sensation of time's passage a myth built on grammar.

None of this was about security. It was about how to describe a universe in which simultaneity is not absolute, so that two observers in different frames can disagree about which of two events came first without either being wrong. The block universe is the picture that survives that disagreement: past, present and future are regions of one four-dimensional whole, not stages a world passes through in turn. Change stops being a thing that happens and becomes a thing you read off by comparing slices, the way a road differs from county to county.

The problem the block left behind is the one this page is about. A slice — a single instant of the four-dimensional whole — contains no motion. It is a photograph, not a film. Yet whoever has to act, has to act now, in something that feels like duration, not in an instant with no width. Whether an instantaneous cut of the block can do anything at all is the standing difficulty, and it turns out to be an engineering difficulty as much as a metaphysical one.

The detection engineer's ninety days

A detection engineer inherits four kinds of stream, none of which stop. A disclosure feed — CVE entries, vendor advisories, exploit-in-the-wild flags — announces that something believed safe yesterday is not safe today. Telemetry — endpoint logs, network flow, authentication events — reports what is actually happening on the estate, continuously, at a rate no human reads directly. Malware corpora update as samples are triaged, reclassified, sometimes withdrawn when a signature turns out to fire on legitimate software. Configuration drift records the estate quietly diverging from whatever baseline was last agreed, as patches land unevenly and shadow changes accumulate.

The characteristic failure sits in the gap between two of these. A vulnerability is disclosed on day one. The audit that would catch an unpatched instance runs on a ninety-day cycle. For up to ninety days the exposure is real, documented, publicly known, and unaddressed on some fraction of the estate, because the system that would have caught it takes its slice too rarely. The disclosure feed knew on day one. The telemetry could, in principle, have shown the vulnerable configuration on day one. Nobody's model of the estate updated on day one. It updated on day ninety, if the audit happened to look in the right place, and even then the update was a fresh photograph, not a correction to the belief that had been standing wrong for three months.

This is the block universe's problem restated in a SIEM console. A slice of the estate, taken at audit time, tells you the state at that instant. It cannot tell you that the state was different eighty-nine days earlier, or when it changed, or which of the four streams should have told you sooner. An instant has no motion in it and, on this evidence, no memory either.

Three admissions of the block

Grade a system by how much of the block it is allowed to see, and a lineage appears.

A Large Language Model is a slab cut from the four-dimensional whole and then sealed. Its corpus stops at a training cutoff; everything after does not exist for it, and — this is the sharper point — nothing inside it is dated relative to anything else. Ask it about a vulnerability class and it answers from whatever mixture of true-then, true-now and no-longer-true happened to be in the text it was trained on, with no internal index distinguishing which. It is not old so much as atemporal: it holds what, not when.

A Large World Model recovers motion by shrinking the slice to a moving present: a scene, sensed live, closed the instant the sensor stops looking. Point one at a network for the duration of a capture window and it can describe what is happening, richly, while the window is open. Close the window and the model has nothing to say about what came before or what follows. It has bought duration by giving up persistence — exactly the trade a detection engineer cannot afford, because the ninety-day gap is not a moment to be sensed once; it is an interval to be tracked.

A Large Universe Model, on this reading, is not a slice at all but an extended segment of a worldline. The disclosure feed, the telemetry, the malware corpus and the drift log run continuously, and beliefs formed from them carry a timestamp and a source: this host was flagged vulnerable, from this feed, at this hour; this flag was revised, from this later scan, at that hour. Nothing closes. A belief that the estate is safe is always open to retraction by the next thing the stream says, and the retraction itself is recorded, not overwritten silently. That is the only architecture in which "day one to day ninety" is a tracked interval rather than a blind spot bounded by two unrelated snapshots.

The objection from discreteness

A continuous system is a fiction. Logs are append-only rows. Scanners poll on a schedule. Packet captures sample at some finite rate. Call it a Large Universe Model if you like; it is a fast Large World Model with a bigger buffer, not a different kind of thing.

The objection is right about the mechanics and wrong about what matters. Every stream feeding a detection stack is discrete: EDR beacons every few seconds, vulnerability scans run weekly or quarterly, threat-intel feeds push updates in batches. The distinguishing property was never sample rate. It is whether the system keeps a closing bracket. A scene-bound tool discards its state when the capture ends — the flow record ages out, the session view resets. A system built to intake continuously does the opposite: it keeps yesterday's belief about a host next to today's contradicting scan result, tags each with its source and moment, and treats the disagreement as something to reconcile rather than overwrite. That obligation — reconcile, don't overwrite — does not arrive at any particular polling frequency. It arrives when the design commits to never declaring the observation finished. A ninety-day exposure is trackable precisely because day one's disclosure and day ninety's scan can be held as two dated beliefs about the same host rather than two unrelated photographs.

The objection from sufficient statistics

Correlation engines already compress history into state. A risk score, a rolling baseline, an anomaly threshold — these are sufficient statistics. If the state captured today already reflects everything the past streams taught, then keeping the streams themselves is redundant. The Kalman filter never needs to remember the sensor readings once it has updated its mean and covariance.

This is the sharper objection, and it concedes the point it is trying to deny. A risk score is sufficient relative to a model someone fitted, and fitting that model — deciding that a login at 3 a.m. from a new geography matters more than one at 3 p.m. from a known one — took an interval of observed behaviour to establish. The compression is only as good as the assumptions baked into it, and those assumptions go stale exactly as the estate drifts. Worse, a rolling baseline typically cannot answer which stream produced the anomaly, when the signal that raised it was itself later retracted, or which earlier alert now needs downgrading because the sensor behind it was found miscalibrated last week. That is provenance, and provenance is precisely the dimension a sufficient statistic throws away in the act of being sufficient. Detection work is not just scoring the present state; it is revising yesterday's score in light of what today's feed says about yesterday, and a compressed statistic has already deleted the yesterday it would need to revise.

What the block universe actually licenses

None of this requires eternalism to be true. A presentist who thinks only the present instant exists can still grant that facts about change — a host going from patched to exposed — are facts about an interval, not a point, because no observer, on any theory of time, has access to a durationless instant. What the metaphysics licenses is narrower and more useful: a way of grading intake by how much of the causally available past and present a system is permitted to keep, rather than merely sense once and discard. Corpora keep none of it, indexed by when. Scenes keep it only while watched. A stream architecture that never closes the bracket, tags every belief with source and moment, and treats retraction as normal rather than exceptional is the only one built to hold a ninety-day gap as a single trackable interval instead of two disconnected photographs, three months apart, that happen to describe the same host.