The most violent collapse of Earth's magnetic field in 40,000 years was, technically, a blip. From inside the event, nothing distinguishes an excursion from a reversal. Your dashboard has the same problem.
Forty-one thousand years ago, Earth's magnetic field began to fail. Over roughly three centuries, a geological eyeblink, the dipole that steers every compass collapsed to somewhere between five and ten percent of its modern strength. The familiar two-pole field came apart into a tangle of weaker poles, so that a compass needle, had anyone been holding one, would have pointed to different norths in different places. A 2025 reconstruction in Science Advances traced the aurora as it broke from its polar mooring and wandered over latitudes that never see it; ice cores record the atmosphere absorbing more cosmic radiation as the planet's shield sagged.
Here is the question: was that the beginning of a polarity reversal, the field flipping, north becoming south for the next few hundred thousand years, or a swerve that would come home?
You cannot answer from inside the event; nothing in the measurements distinguishes the two. We only know because we have the ending: after about two thousand years, the field climbed back and settled into the same polarity it had before. North stayed north. The episode is called the Laschamps excursion, after lava flows in the French Massif Central where its trace was first read, and "excursion" is the geophysicist's word for exactly this, a departure, even a near-total collapse, that returns.
Sit with the uncomfortable part. The most dramatic geomagnetic deviation in the recent geological record, a ninety-percent collapse of a planetary field, was, in the technical sense, a blip.
Geomagnetism sorts large polarity events into two bins. A reversal flips and stays: the last full one, the Brunhes–Matuyama reversal about 780,000 years ago, swapped magnetic north and south and held. An excursion departs and returns: the pole wanders far from the rotation axis, sometimes more than 45 degrees, the dipole weakens drastically, the field may splinter into multiple poles, and then the whole thing comes home to the polarity it left. The literature's own names for excursions give the game away: "failed reversals," "an aborted polarity state." They run short, under five thousand years, and they happen more often than true reversals do.
So the taxonomy is clean. The diagnosis is not. The field's own literature concedes that "it is not always easy to distinguish between an excursion and a short polarity event." Those tidy labels are assigned at the finish line, by people who already know how the story ended. While the dipole is collapsing, an excursion and an early-stage reversal don't merely look the same, for all anyone can measure, they may be the same, a system departing from baseline with its ending not yet written.
That is not a gap in the data that better instruments will close. It is a property of the early signal.
Why do excursions return at all? The explanation favored in geodynamo modeling is one of the loveliest details in geophysics. Earth has two cores: a liquid outer core, where churning iron generates the field and where magnetic changes work themselves through in something like 500 years, and a solid inner core, where changes diffuse far more slowly, on the order of 3,000 years. During an excursion, the field reverses in the fast, volatile outer core, but the slow inner core doesn't follow. It still holds the old field, the way a thick wall holds the day's heat into the night, and it pulls the system back. The inner core remembers.
A deviation only becomes a regime change when it persists long enough to convert the slow layer too. Until that happens, the slow layer is actively dragging the system home. Keep that mechanism in your pocket; we'll need it again.
This is not just deep-past lore; we are living inside the ambiguity right now. Over the roughly 175 years of systematic measurement, the dipole has lost about nine percent of its strength, a rate near five percent per century since 1840. The weakening is not uniform: it concentrates in the South Atlantic Anomaly, a growing weak patch stretching from South America toward southern Africa, the place where satellites shut off their sensitive instruments to ride out the elevated radiation. I've written before about what the anomaly's geography teaches, that systems fail locally before they fail globally. This essay is about the other lesson, the temporal one: what are we allowed to conclude from the trend?
The popular-science answer arrives every few months with the same headline: the field is weakening, the patch is growing, is the flip beginning? The professional answer is a model of discipline. Researchers at the GFZ, the German Research Centre for Geosciences, concluded the anomaly is "probably no evidence" of an impending reversal; a 2016 review in Frontiers in Earth Science put it that the anomaly "does not necessarily indicate an impending pole reversal," because the current decay sits within the range of ordinary variation during stable polarity epochs. Weak patches like this one come and go without anything flipping.
And there is a precedent, the kind of control case the universe rarely hands you. A study in PNAS of recurrent ancient field anomalies found that a near-identical weak patch sat over the South Atlantic region around 48,000 years ago. No reversal followed. The patch faded; polarity held. Today's scariest-looking deviation has an older twin, and the twin went home.
The geophysicists are resisting two temptations, the two most common errors anyone makes in front of a deviating metric.
The first is linear extrapolation. Nine percent down in 175 years invites a tidy arithmetic: at this rate, the field is gone in roughly two thousand years. The inference is unsound, and geomagnetists say so flatly. The geodynamo is a chaotic, nonlinear system; the dipole has sagged and recovered repeatedly within the current stable epoch without ever approaching zero. A trend is not a trajectory. The software twin needs no introduction: "error rates climbed nine percent this week, at this rate we're down by the fourteenth." Most systems are nonlinear in exactly the way that makes this projection worthless; they mean-revert far more often than they run away. Straight-line forecasts are only as good as the mechanism's linearity, and almost no interesting mechanism is linear.
The second is the gambler's fallacy dressed up as prudence. The last reversal was about 780,000 years ago; the average gap between reversals over the last few million years runs roughly 200,000 to 300,000 years. So we're overdue, right? But reversals are not periodic. They have no schedule to be behind. The Cretaceous Normal Superchron ran nearly 37 million years, from about 120 to 83 million years ago, without a single flip. A process with no clock cannot owe you an event, and "overdue" is a statement about clocks. The ops twin is uttered in some retro every week: "we haven't had a major incident in eighteen months, we're due." You are not due. Incidents arrive when their causes do, not when the calendar feels heavy.
Now make it yours. The pager goes off, or the morning dashboard just looks wrong. P99 latency, the time your slowest one percent of requests take, has been climbing for forty minutes. Or a language model that has behaved for weeks produces a run of strange outputs. Or checkout conversion has sagged a point and a half since Thursday's deploy. Or the queue is draining slower than it fills.
Every one of these is the Laschamps question. Excursion or reversal?
Both answers are live. The latency climb might be a noisy neighbor on shared infrastructure, gone by morning, an excursion. Or your working set has just outgrown memory and the cache hit ratio has started its cliff-dive, a genuine regime change that will never fix itself, a reversal. The odd model outputs might be sampling noise on hard inputs, or the upstream provider shipped a model update and the behavior you tuned around is permanently different. The traffic bump might be a bot scrape, or the first morning of a baseline that a launch just doubled for good.
And because the early signal supports either reading, there are two symmetric ways to fail, and you have seen both. Over-reaction treats every excursion as a reversal: the 3 a.m. rollback of an innocent deploy, the emergency re-architecture around a spike that would have faded by Friday, thresholds re-tuned around noise until alerts mean nothing. Under-reaction treats a real reversal as an excursion: that's the postmortem where the regime change was visible for six weeks while everyone called it "that blip again", the slow boil dismissed daily until it was a crisis. Neither error is stupidity. Both are premature certainty applied in opposite directions.
It's fair to note where the analogy strains: a geophysicist can hold the question open for centuries; you have minutes and a duty to respond. But that objection sharpens the lesson, because the discipline that follows is precisely about what you may do immediately and what you must refuse to do early.
| Geomagnetism | Your system |
|---|---|
| Excursion: departs and returns (a blip) | Noisy neighbor, bot scrape, sampling noise on hard inputs |
| Reversal: flips and stays (a regime change) | Working set outgrew memory; upstream shipped a new model |
| "The field is gone in 2,000 years at this rate" | Linear extrapolation: a trend is not a trajectory |
| "We're overdue for a reversal" | Gambler's fallacy: a process with no clock owes you nothing |
| The inner core remembers; persistence converts it | Baselines and architecture convert only on persistence |
If the wait-and-see answer feels like an admission of failure, statistics says otherwise. The formal version of this problem is sequential change-point detection, deciding, observation by observation, whether a process has shifted regimes. E. S. Page's CUSUM procedure, published in Biometrika in 1954, is the classic tool, and its central lesson has not budged in seventy years: you tune a threshold that trades detection delay against false-alarm rate. Want to catch every real shift instantly? Accept a torrent of false alarms. Want silence unless it's real? Accept that you confirm the shift late. You can choose your point on that frontier. You cannot leave the frontier.
The ambiguity of an early deviation is not your team's ignorance. It is a mathematical property of early evidence from noisy systems. The people who study the actual magnetic field, armed with a dedicated satellite constellation and 180 years of observatory records, answer "we cannot tell yet" in public, about the literal planet. You are allowed the same answer about a dashboard.
What you are not allowed is to stop there. "Can't tell yet" earns its keep only when it comes with a plan for telling eventually. That plan has a shape.
Refuse the straight line. Ban "at this rate, X by Friday" from the incident channel unless the mechanism is genuinely linear. A disk filling at a constant write rate extrapolates honestly; error dynamics, latency under feedback, load under retries do not. When someone projects a nonlinear metric to its doom-date, the projection is theater.
Start from the base rate. Excursions outnumber reversals in the geological record, and your own history almost certainly agrees: pull up the last year of large deviations and count how many resolved without anyone doing anything fundamental. For most systems the overwhelming majority were blips. So the prior says blip, which is not a license to relax. The base rate buys you patience, not blindness; it tells you the cost of waiting a beat is usually small, and it reminds you that the rare true reversal is dangerous precisely because the prior votes against it every time.
Measure your excursion timescale, and gate the expensive response on persistence. This is the single most useful number you can extract from your own history: how long do your transients live? Pull your past anomalies and measure time-to-return-to-baseline; take a high percentile, that's your system's excursion timescale, the analog of the few-thousand-year window inside which the field's departures come home. A deviation younger than that window has not yet earned the regime-change response. Cheap, reversible moves are always allowed while you wait: shed load, scale out, roll back the suspect deploy; the geophysicists' equivalent is satellites powering down instruments for the anomaly pass. What waits behind the persistence gate is the expensive, slow-to-reverse commitments: the re-architecture, the re-baselined alerts, the rewritten capacity model, the declaration of a new normal. In core terms: let your slow layer (baselines, invariants, architecture) be converted only by persistence, never by panic. The fast layer does not get to rewrite the slow layer on its own say-so.
Hold the ambiguity out loud, with a date on it. "I can't tell yet" is a complete, professional status when it ships with its instruments: here's the deviation, here are both hypotheses, here's the persistence gate it has to outlive, here's the re-review time, and here are the tripwires, defined now, in advance, that would upgrade it early (a second independent metric confirming, a mechanism identified, the deviation accelerating instead of decaying). Write down today what evidence would change your mind, because after another week of slow burn you will be tempted to move the goalposts in whichever direction you were already leaning.
One more look at the inner core, because it answers a question this essay has been circling: what makes a regime change real? The planet's most violent field collapse in the recent record turned out to be a blip because something slow at the center never got the memo, and dragged the whole system home. The day the slow layer flips too, that's the day it was a reversal all along.
Your systems have a slow layer: the baselines, the capacity model, the architectural assumptions, the contracts you've made with the teams around you. Most deviations will die against it, the way every excursion for the last 780,000 years has died against the inner core. The rare one that doesn't, the one still burning past your excursion timescale, still pushing on the slow layer, has identified itself, and now the big response is justified by evidence rather than by adrenaline.
Until then, the strongest thing a careful operator can say while the needle is wandering is not yet, said not as a shrug, but as a measurement plan with a deadline. The amateurs commit early, in both directions. The discipline is the instrumented wait.
Sources: the Laschamps excursion record and excursion/reversal taxonomy (Laschamp event literature; Geophysical Journal International on the excursion/reversal distinction; "Wandering of the auroral oval 41,000 years ago," Science Advances, 2025); South Atlantic Anomaly assessments (GFZ German Research Centre for Geosciences; "The South Atlantic Anomaly," Frontiers in Earth Science, 2016; "Recurrent ancient geomagnetic field anomalies," PNAS); dipole decay and reversal chronology (GFZ secular-variation data; USGS on reversal non-periodicity; Mahgoub et al. 2023, JGR Solid Earth, on the Brunhes–Matuyama transition); change-point detection (E. S. Page, "Continuous Inspection Schemes," Biometrika 41, 1954).
Let your slow layer be converted only by persistence, never by panic.
An agent's slow layer is its trust substrate: the provenance of what it actually did and the reputation it earned over a track record, neither of which a single strange run should flip. The Agent Trust Stack is that slow layer made durable, a verifiable history and a persistence-weighted reputation, so one excursion stays an excursion and only a real, sustained regime change moves the baseline.
pip install agent-trust-stack · npm install agent-trust-stack
vibeagentmaking.com → · See it in action