The Verification Venue · your own machine is the apparatus

The Disk You Are Somewhere Inside

Press one button. Your browser times its own round trip to the nearest edge of the internet, and that time is a hard physical fact: light in glass could not have carried it further than a certain number of kilometres. You now know something true about where you are. You also know, exactly and permanently, how little that is.

A round trip time is a bound, and it points one way only. It says the far end is no further than X. It cannot say the far end is X, it cannot say it is nearly X, and no amount of cleverness turns it into a place. Everything below is built out of that one asymmetry: first the number, then the shape of what the number cannot do, then a published result reproduced on your line, then the same reduction run over real bursts from a machine whose position nobody here knows.

01 · the instrumentOne button, one hard number

Twenty small requests, sent one at a time to this site's own origin, and the fastest one kept. Not the average: queueing only ever adds, so the mean is the floor plus a tax you did not measure. Singla and colleagues, timing real handshakes between Akamai's servers and end users for a day, found the average round trip sits at 1.9× the minimum in the median case. The minimum is the only statistic here with a physical meaning.

The burst

idle

This browser's clock tick

not read yet

Measured over

—

Refusal floor (3 ticks)

—

Probe target —: not run yet.

02 · the asymmetryThe arrow only points one way

Take your measurement seriously for a moment and ask the obvious next question: so how far away is it? The honest answer is that your own measurement is equally happy with every answer between zero and the bound. Slide the assumed distance and watch: the round trip never changes, because whatever the glass does not spend, the queues and the switches and the server do.

Assume a distance, and see what is left over

Run the burst above first: this slider divides a number you have not measured yet.

03 · the floorThe part that is just glass being glass

Some of your delay is nobody's fault. Light in a silica fibre core travels at about two thirds of its vacuum speed, because that is what a refractive index is, and the entire internet is made of the stuff. Singla and colleagues put it plainly: a 1.5× inflation over the speed of light would happen even on a perfectly straight cable laid along the shortest path on Earth. Below, drag the core material from solid silica to air and watch which part of your own number moves.

Change the material, not the engineering

Core refractive index — 1.4682

Right is standard single-mode fibre, the cable actually in the ground. Left is a hollow core, which guides the light through air instead of glass. Nothing between the two ends is a deployed cable; the slider is continuous because the arithmetic is, not because the world is. This slider reaches nothing else on the page. The measurement above, the grade below and the replayed traces are all pinned to real silica, because a control that quietly re-scored a measurement would be exactly the kind of check that cannot fail.

solid silica core, the fibre in the ground

Signal speed

—

One-way delay

—

Cut against silica

—

20,000 km, there and back

—

The hollow-core end is not a thought experiment. In 2025 a Southampton and Microsoft group reported a fibre with a —, and, of the propagation itself, — against standard telecom fibre. The cut this slider shows at that end is computed here from the refractive index of air alone and comes out at —. Their figure is a rounded one and this is not a reproduction of it: it is the part of their claim that is pure material physics, arriving in the same place.

04 · the reproductionHand it a distance, and it will grade your connection

The page has one bound and no idea where you are. You have the other half. Type the three-letter code of the airport nearest you and it will do the division: how many times slower than light your own line is, against the median that four researchers read off their Figure 3, whose population is roughly a million connections from 400 or more research hosts. (Their six million is the whole-fetch curve two figures earlier, and it is a different number entirely.) If you type somewhere your measurement forbids, it will tell you so and refuse to print a number.

Name a place. Nothing is sent anywhere.

Airport code nearest you

— codes are shipped with this page, from OurAirports. You are telling the page where you are; it never worked that out, and the moment you close the tab it is gone.

Run the burst first: there is nothing to divide yet.

05 · the ledgerWhat this page load cost, counted in round trips

Now that you have measured your own round trip, it becomes a unit. The browser recorded how long each phase of fetching this very page took; divide each by your round trip and you get the only honest currency for a handshake. Where a phase reads zero because it did not happen on this request, the table says so rather than dividing zero and calling your network perfect.

Two things this table is not. The encryption line is wall time around the handshake as the browser reports it, so it also holds the certificate checking this machine did locally, which is processor work and not distance; on a busy laptop that part alone can exceed the whole round trip, and a decomposition that files it under "the internet" is measuring the laptop. And if this page arrived over HTTP/3 there is no separate transport handshake to report, because QUIC folds connection set-up into the encrypted exchange: the connection line and the encryption line under it are then two readings of one event rather than two separable costs. The protocol actually negotiated is printed beside the table so you can tell which case you are in.

This page's own arrival

Run the burst above: without a round trip there is no unit to count in.

06 · the golden tracesWhere the answer is known, and where it still is not

A measurement from a stranger's laptop cannot be checked by anybody. So this page also ships real bursts, captured on one machine to twelve targets from Melbourne to Cape Town, and replays them through the reduction you just ran. The interesting constraint is that nobody here knows where that machine was. Its position was deliberately not recorded. Geometry alone still has plenty to say.

Take the tightest bound as an anchor: the machine is within that distance of the edge that answered. For any other target, the triangle inequality then forces a minimum distance, and the finite size of the Earth forces a maximum. Two constraints, no position, and a real bracket on how many times slower than light each of those paths ran.

Replay

Committed trace

Captured —, link condition: —. One TCP connect is one round trip: SYN out, SYN-ACK back, which is the same quantity the paper's Figure 3 calls the TCP handshake time.

Anchor

—

Pairs tested

—

Impossible pairs

—

Bounds bigger than the planet

—

target	fastest ms	no further than, km	distance must be in, km	times slower than light

Rows shaded warm have a bound larger than half the Earth's circumference, so the bound alone excludes nothing; their brackets come from the planet's size, not from the cable. In — rows the low end of the bracket is exactly the refractive index, because the upper end of the distance range is the bound itself. Those low ends are the fibre floor restated, not a measurement, and they are marked grey for that reason. Cape Town is the one that pays for the whole exercise: its bracket is —, which starts above the published median of 3.2 times, and nobody involved ever learned where the measuring machine was.

Three browser engines, one link, one edge

The same burst, fired by Chromium, Firefox and WebKit within minutes of each other on the same machine. Browsers deliberately blunt performance.now() to make timing side channels harder, and they do not agree on how much. That policy, not your connection, sets the resolution of everything above.

engine	clock tick	fastest probe	bound	protocol

Widest minus narrowest: —. Against a bare TCP socket on the same machine, the finest-grained engine reads —. A one-millisecond clock cannot report a value at all, only the interval below it, so the honest test is whether the fine clock's answer lives inside the coarse one's interval. That test only means anything between engines that crossed the same wire, and they do not always: which protocol each one negotiated is in the table, and it moves between captures.

—

Five traces built to fool it

If the reduction returns a confident number from noise, it is not an instrument, it is a random number generator with a citation. These are run live, in your browser, every time this page loads.

control	must	did

And the sixth, which nothing in the reduction can catch on its own: a real burst filed under the wrong city. Only the geometry notices.
Filed correctly → —
Filed as Cape Town → —

And the seventh, which is the instrument's price list rather than a pass or a fail. Take a real burst and answer some of its probes early, as a proxy or a caching box on your own side of the link would. Below the edge-sharpness rule's rank the instrument declines; at it and above, the rule is blind and the bound it prints is too tight, which is worse than too loose because a too-tight bound is a sentence that is false rather than merely unhelpful. Both halves are run live, on the committed Cape Town burst, right now.
—
—
—

The check, run in your browser, now

running…

Every figure above is recomputed from the same module the offline verifier imports. Rows marked RESTATEMENT only prove this page agrees with itself; rows marked CONTROL and PERTURBATION are the ones that could actually go red.

The apparatus: what is being run, what is assumed, and what this page does not claim

What is being run

One reduction, in one file. lightbound.mjs holds the constants, the refusal rules, the bound, the great-circle distance and the geometry, and it is imported without modification by this page, by the offline mirror (research/how-far-away-the-internet-is/06-verify.mjs) and by the headless page check (verify-how-far-away-the-internet-is.mjs). Nothing reimplements it. probe.mjs holds the part that touches the network, and the committed browser traces were produced by executing that exact file inside each engine.

The bound is min RTT × c / (2n). Nothing subtler is going on: a round trip covers the path twice, and the signal moves at c/n. Any real delay beyond propagation, and there is always some, only makes the bound looser. That is why it is a bound and not an estimate.

The rule fixed before measuring

The minimum, never the mean. Queueing is one-sided noise. Fixed before the first capture.
A three-letter code names one point on Earth, namely the coordinates OurAirports gives that airport, and the same rule resolves Cloudflare edges and AWS regions alike so neither end gets a friendlier position.
Five refusal rules, all written down before any trace existed: at least 5 usable probes; a minimum of at least 3 clock ticks; a median no more than 20 times the minimum; the edge unchanged across the burst; nothing non-positive. Three of the five carry a numeric threshold and those three live in RULES; the other two are structural and have nothing to tune.
One rule was not pre-registered. The sixth, the sharpness of the left edge of the distribution, was added after the noise control was found to trip the median rule with only — of margin, and its threshold of 0.5 was placed in an observed gap rather than predicted. Recomputed in front of you from the committed bursts right now: —. A number chosen by looking at the answer is weaker than one chosen before it, and this page will not present the two as equals. —
And the sixth rule was read at the wrong rank until an adversarial review caught it. It was written as the tenth percentile, which on a twenty-probe burst is the second-smallest sample and nothing more, so two probes answered early by anything nearer than the edge went straight past it and the instrument printed a bound five times too tight with no complaint. The threshold did not move; the rank is now the lower quartile, which is what "the left edge of the distribution" meant in the first place. It costs an attacker —, and the seventh control under the refusal table runs both halves of that in front of you. No rule of this shape can do better: a fixed rank always has a number of probes that buys past it.

What this page does not claim

It does not know where you are, and it cannot find out. A round trip bounds a distance from above. One bound is a disk. Nothing here inverts it, and there is no setting that makes it try.
It does not claim the far end is at the bound. The bound is the furthest it could be.
It does not claim the glass bound holds when the path is not glass. There are two bounds here and only one of them is unconditional. Light in vacuum is a limit nothing beats, and the vacuum cell beside the headline is that limit. The glass figure is tighter by the index of refraction and buys that tightness with an assumption about the cable: a microwave link runs at almost the vacuum speed, a hollow core at about 99.97% of it, and free-space optics likewise, so a route carrying one of those hops can legitimately reach further than the kilometres printed above. This is not a corner case invented to be careful. The opening example of the paper cited below is the Chicago to New York microwave route, built precisely because it beats fibre. Nothing here can see which materials your packets crossed.
It does not claim that the bound holding is evidence about your network. Those kilometres were computed from the very round trip that reached the edge, and the edge is by construction the nearest one there is, so the only way the headline sentence could come out false is if the packet had outrun light in glass. A test whose sole failure mode is a violation of physics is a tautology with units on it. The falsifiable work on this page is elsewhere: twelve committed bursts that must stay geometrically consistent with each other, five broken traces that must be refused by name, a real burst filed under the wrong city that must become impossible, and a headless run against a server that answers after a delay it never tells the page.
It does not claim the edge's name for itself has been verified. The three-letter code arrives in a file the edge writes about itself, so the position the geometry is checked against is supplied by the party being checked. Nothing here can confirm it, which is why the impossible verdict in section 04 names it as one of the things that might be false rather than assuming it is true.
It does not claim your traffic follows a great circle. Cable routes are longer than straight lines, which is part of why your inflation exceeds the index of refraction.
It does not claim a Cloudflare edge sits at an airport. The code names the nearest major international airport; the building is somewhere in that metro, tens of kilometres out.
It does not claim the committed traces are representative of anything but one home connection in Australia on one evening.
It does not claim the noise rule is strong. A gentler kind of noise would get past it.
It does not claim the bound is safe against something on your own side of the link answering for the edge. A proxy, a service worker, a captive portal or a caching box that answers a large enough minority of the probes drags the minimum down, and a minimum that is too small makes the bound too tight, which is the one direction in which the headline sentence can be false rather than merely loose. The edge-sharpness rule prices that attack at — and cannot price it higher without refusing honest jittery links. Nothing here detects it, and the only reason the number is worth anything is that the probe is same-origin and cache-busted.
It does not claim the grade in section 04 is a like-for-like comparison at short range. Inflation divides by the distance, so it runs away as the distance shrinks: the last mile, the switch and the server cost what they cost whether the far end is 50 km away or 5,000. Against a published median of 3.2 times, no line at all can be graded better unless the place you name is further away than — of the bound you measured, whatever that bound turns out to be, and the edge that answered is by construction the nearest one there is. Section 04 prints that crossover in kilometres beside its verdict for exactly this reason.
It does not claim these figures are comparable to the paper's without qualification. The paper measured from well-connected research hosts in 2014, over TCP, and it excluded the few hundred sites that used SSL from the data it published, so it carries no encryption term at all; this measures one residential line in 2026 over HTTP/3, where a cold fetch pays for a handshake the paper never had in front of it. That is why the ledger in section 05 is denominated in your own round trip and never in theirs.

The one thing on this page that is not ours

—

This file requests nothing but its own origin, stores nothing, and sends your timings nowhere. It also cannot promise you a clean network tab, because the site's Cloudflare account has Web Analytics enabled and the edge injects a beacon script into every page here after the file leaves the repository. The check above counts what actually loaded on the copy you are reading, whichever copy that is, and prints it whether it is flattering or not.

Predictions, and how they did

P1 half held, and the half that did not is not scored as held. The prediction had two parts: the three browser engines would disagree about the distance purely because of clock coarsening, and the coarse engines' intervals would still contain the fine engine's answer. The first part came out: —. Whether the second part could be tested at all depends on which wire each engine negotiated, which is in the table above and moves between captures, so the sentence that follows is written by the arithmetic rather than by whoever last edited this list. — A containment test between two listeners that crossed different wires is not a test of clock coarsening, so where that is what happened the page reports the numbers and declines to call the prediction confirmed. An earlier version of this bullet said "both parts came out" while the computed sentence forty lines above it said the opposite; that is the failure this whole layer is about, and it survived because the two check rows guarding it restated their own branch condition.
P2 held, but for the wrong reason. Loading the link would leave the minimum alone. It did, on —. But the mean barely moved either, because two bulk streams did not congest this uplink, so the prediction was never really tested.
P3 died. The only target whose minimum moved materially between the two captures was Mumbai, and it got faster, by —. That is a route change between captures, not congestion, and it is the opposite of what a load test predicts. It is left in the committed trace rather than recaptured.
P5 held, for the wrong reason. Served from a loopback address, where the round trip is a fraction of a millisecond, the page should decline. It does, but not because of the clock floor written for exactly that case. Loopback times jitter so much relative to their own floor that the distribution has no sharp left edge at all, and the rule that fires is the sixth one, the one added after the fact. The instrument refuses to measure the machine it is running on, and the reason it gives is not the one it was designed to give.
P4 died. I expected the bound to be useful at long range. It is not: on this line, — are larger than half the circumference of the Earth and therefore exclude nowhere at all. The internet is slow enough that the speed of light stops being a constraint somewhere past Singapore.

An error this page made and kept

The first version of the bracket printed a lower end for every target and treated it as a result. It is not. Where the bound is smaller than the planet, the largest distance the target can be at is the bound itself, so the low end comes out as exactly the index of refraction, every time, for every connection on Earth. It was the refractive index restated as though it were a measurement. The rows where that is true are still shown, still with their 1.47×, and marked grey and named in the caption, because deleting them would hide how easy that mistake was to make.

Sources

read from the typeset PDF	page

—

The footnote a text extractor destroys. Footnote 5 on page 3 sets two stacked fractions, and pdftotext renders them as bare digits, which is how a plausible wrong number gets into circulation. Read off a rendering of the page, it says: —

The fibre. —

The hollow core. —

The edge code. Cloudflare's own support documentation: —

Places. OurAirports, released into the public domain; the committed rows are the survivors of a filter to airports typed large or medium that carry a three-letter code. Targets. AWS publishes the city each region is in; the regional service endpoints used are not anycast and resolve into the region they name.

Reproducing

node research/how-far-away-the-internet-is/01-airports.mjs
node research/how-far-away-the-internet-is/02-capture-tcp.mjs idle
node research/how-far-away-the-internet-is/02-capture-tcp.mjs loaded
node research/how-far-away-the-internet-is/03-capture-browser.mjs
node research/how-far-away-the-internet-is/04-negative-controls.mjs
node research/how-far-away-the-internet-is/05-build-page-data.mjs
node research/how-far-away-the-internet-is/06-verify.mjs
node verify-how-far-away-the-internet-is.mjs

Steps 02 and 03 talk to the network and will produce your traces, not these ones. Steps 04 to 06, and the headless check under them, are offline and reproduce every number on this page from the committed files.