Summary
- The November 2025 ISP Column prints China's IPv6 fragmentation drop rate as "a massive 76%"; the APNIC Blog version dated 14 November 2025 prints "6%" in the same sentence, while the Japan and India sentences agree word for word on both carriers.
- The APNIC Labs v6frag map prints a seven-day average with explicit start and end dates and a "Window (Days)" control, and prints no sample frame and no denominator stated as a count in the text read.
- No correction changing the China figure turned up in the inspected text of either carrier, and no archived November 2025 rendering of the map was obtained, so the three percentages cannot be resolved to the window and count that produced them.
One sentence, two numbers
Geoff Huston's measurement columns sit in an unusual place: one person runs the experiment and writes the sentence about it. That makes a small documentary question worth asking of any percentage in the prose. Which page produced it, over what period, out of how many measurements.
The November 2025 ISP Column, "How we Measure: IPv6", answers part of that. Its fragmentation paragraph reads: "the IPv6 fragmentation drop rate in China, where the IPv6 deployment is currently underway, has a massive 76% drop rate." ISP Column, November 2025
The APNIC Blog carries the same article under the same title, dated 14 November 2025, and the sentence arrives with a different number: "the IPv6 fragmentation drop rate in China, where the IPv6 deployment is currently underway, has a 6% drop rate." APNIC Blog, 14 November 2025
Everything around the figure is the same, including the sentences before it. Japan's IPv6 fragmentation packet drop rate is "an average of 20% of all users"; India's "has a far smaller 1% drop rate". My own count across the three carrier artifacts read: two print 76, the ISP Column HTML page and its PDF rendering ISP Column PDF, November 2025; one prints 6. My own subtraction: the pair sits 70 percentage points apart.
In the text read on 21 November 2026, neither carrier nor the PDF carried a correction, erratum, revision note or "this post was updated" line changing that figure. The absence is scoped to the text read: comments, footers and post-publication metadata were not read end to end.
What the map prints
Both versions name the fragmentation map as the source. On that map, the visible text prints "7 day average (14/11/2026 - 20/11/2026)" beside a control labelled "Window (Days)" set to 7. APNIC Labs v6frag map That is a printed scope: seven days, with a start and an end.
The same visible text carries the title, a map link, a remember-choice control and four metric labels: "Frag Drop", "Atomic Frag Drop", "HBH Header Drop", "DST Header Drop". It carries no prose methodology, no statement of which networks, resolvers or vantage points form the sample, and no denominator stated as a count. The page is dynamic, so that describes the text read, not tooltips or script-loaded content.
Then there is the second surface. The zoomable map prints the same title and the same metric with the same "Window (Days) 7" control, over "7 day average (08/11/2026 - 14/11/2026)". APNIC Labs v6frag-zoom map One metric, two surfaces, windows six days apart — my subtraction of the two printed dates.
The programme does have a general method page, and it stays general: "Our measurements are made by tasking a browsers to resolve DNS Names and fetch very small web objects, exercising a particular technlogy in so doing" (spelling as printed). APNIC Labs background It does not describe the fragmentation experiment, name a v6frag population or print a count.
What the prose carries, and what stays behind
The column describes the cited map as "current" and says it is "generated on a daily basis". It gives no start or end date for the 20%, the 1% or the 76%. It says the measurement system presents "between 25M to 30M ad impressions per day", and separately prints no count denominator for the fragmentation percentages. It names the map as the source and defers to it; its own text defines no population.
So "all users" stays undefined on both carriers — all sampled users, IPv6-capable users, or IPv6 samples. The phrase does the heavy lifting in the Japan sentence and no page read says which one it means.
The threshold arrives as an adjective rather than a line. India is "far smaller" at 1%, China is "massive" at 76%. A reader is told which proportion is a problem, and no page read prints the value at which a proportion becomes one.
The written method exists, and it is dated
APNIC does publish readable method prose for this experiment. A 2021 article describes the setup: end hosts recruited through an ad campaign; a front end that fragments outbound TCP packets above 1,200 octets at randomly chosen, 8-octet-spaced initial sizes; reassembly inferred from ACK sequence numbers. It carries a table of IPv6 samples by region, one row printed as "| XA | World | 6.68% | 23,209,746 |". APNIC Blog, 23 April 2021
A 2022 article adds that the measurement "uses scripts within an online ad campaign and executes some 4M different measurements per day, spanning a large proportion of the IPv6 client base", and describes 28 fragmentation subtests with initial sizes chosen at random between 1,200 and 1,416 bytes. APNIC Blog, 13 October 2022
That is what a method page looks like when it is readable, and it is also dated: those articles describe the experiment as configured then. Neither defines the current report's population, and neither is where the November 2025 column sends a reader.
Where the audit stops
Two documents would close this. One is an archived November 2025 rendering of the v6frag map carrying the window that produced the column's percentages. The other is a fragmentation-specific method page stating the sample population and a current denominator. Neither was obtained. The map as read serves windows in November 2026; the column is dated November 2025.
The consequence is narrow, and worth stating precisely. The arithmetic travels: same sentence, same unit, same comparison, intact on both carriers. The count does not travel, and neither does the population.
So here is what a stranger can compose from the pages read. The window is printed and the denominator is not. The map is named and its method sits elsewhere, and older. Two of the author's own surfaces print one metric over two different seven-day windows, with no reconciliation a reader can see.
Huston's corpus is auditable at all because he prints the page he measured from, and most measurement claims in this industry cite nothing a reader can open. The audit stops at the edge of the dashboard: at a seven-day window whose numerator reached the sentence and whose denominator stayed on the server.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






