Summary
- InternetNZ prints DNS availability at 100% against a 100% target on its 2023–24, 2024–25 and 2025–26 report pages, and at 100.00% for two years in its audited statement; the notes under those rows expand acronyms and define WHOIS, then stop.
- Across the twelve documents read for this piece, no page prints the measurement window, the denominator, the monitoring instrument or the party computing the figure, and the maintenance carve-out printed twice in the same page set never sits on an annual output row.
A row of three cells carries the whole claim. On the Annual Report 2025–2026 page, under the heading “Registry performance”, the columns are “Target” and “Output”, singular. DNS availability [i] reads 100% against 100%. EPP availability [ii] reads 99.99% against 99.9%. WHOIS availability [iii] reads 99.99% against 99.9%. The three notes beneath the table do the work of definition, and the work they do is naming: [i] “Domain Name System”; [ii] “Extensible Provisioning Protocol”; [iii] “WHOIS — public database that is updated when a domain name is registered or when DNS settings are updated”. A reader finishes the table knowing what the acronyms stand for, and nothing about what was counted.
The same three rows have been typed four times. The 2024–25 page prints DNS 100% against 100%, EPP 99.997% against 99.9% and WHOIS 99.99% against 99.9%. The 2023–24 page prints 100%, 99.997% and 99.998% against 100%, 99.90% and 99.90%. The 2022–23 page prints DNS availability 100.000%, EPP availability 99.999%, WHOIS availability 99.975%. Four financial years, twelve printed percentages, one footnote vocabulary.
The precision moves and nothing explains it. DNS availability appears with three decimal places in 2022–23, none at all in the three later report pages, and two in the audited table. That is a count of printed places, not evidence that the measure changed — but it does tell you that someone chooses the resolution per audience, and that no page says which printing a reader should treat as the record.
The audited version is the statement of service performance for the year ended 31 March 2026. It prints “Table 2: .nz operational measures”, with columns “Activity”, “Outputs in 2025–26” and “Outputs in 2024–25”: DNS availability 100.00% in both years, EPP 99.99% in both, WHOIS 99.99% in both. The statement runs pages 3 to 13, the financial statements pages 14 to 35. Grant Thornton New Zealand Audit Limited signed on 16 July 2026 and opined that the performance report presents fairly the service performance for the year “in that the service performance information is appropriate and meaningful and prepared in accordance with the Group's measurement bases or evaluation methods”, under ISAs (NZ) and New Zealand Auditing Standard (NZ AS) 1 (Revised). Read that clause twice. The Group's own measurement bases are the standard the audit applies. They are not printed beside Table 2.
Twice in this page set a reader meets the phrase that decides what the numbers exclude, and twice it sits somewhere else. The .nz statistics page states “Our availability service level commitment are as follows”, then “DNS is 100%”, then the registry system “which includes Portal, EPP, APIs and WHOIS, the availability service level commitment is 99.9% excluding scheduled maintenance”. The product-strategy tables of the 2024–25 and 2023–24 pages carry the line “Registry services availability (excluding scheduled maintenance)” at a 99.9% target, “On track”. Neither is an annual output row. The row that prints 100.00% does not say whether maintenance is inside it, and a reader cannot move the carve-out twelve centimetres to the left on their own authority.
The DNS page is the richest document of the twelve. It says the .nz name service “has to be available 100% of the time”; that InternetNZ runs nameservers inside New Zealand and uses “two international providers of a global network of nameservers”; that anycast is used “on some nameservers” so several machines answer as one; and that the operator deliberately runs different hardware, different operating systems and different nameserver software across the whole provision, which the page calls genetic diversity. Then a table names them: ns1 unicast in New Zealand, ns2 to ns4 anycast in New Zealand, ns5 and ns6 anycast and multiple international on CIRA, ns7 anycast and multiple international on Netnod. “All our nameservers are locally and remotely monitored”, and the page points to the RIPE NCC DNS Monitoring Service for external results on the secondary name servers. Not one of those sentences says that this monitoring, or any part of it, produces the number in the annual report. The engineering page and the assurance page describe one service in two languages that never meet.
Two more design documents fill in the register side. The production environment page says normal operation runs two instances, in Auckland and Wellington, one active and one in hot standby, with possible swaps during maintenance. The platform page records that the 2022 replacement deprecated the proprietary Shared Registry System protocol and standardised on the Extensible Provisioning Protocol, built with the Canadian Internet Registration Authority; the registry system page says InternetNZ keeps the definitive .nz register and that the new system went live on 1 November 2022, replacing one first developed in 2002. Three documents about shape. No realised figure anywhere in them.
The status host is the only surface with a clock attached, and the clock is ninety days. It prints “Uptime over the past 90 days” and monitors EPP, TBR, Portal, WHOIS, Zone Publish, ns1 through ns7, .nz DNSSEC and a .nz NTP network whose nodes it lists in Auckland, Wellington and Christchurch. A rolling window measured to the day cannot be reconciled with a financial year by anyone reading in November, and the page does not pretend otherwise.
Then there is the year that printed 100%. The 2023–24 financial year contains 29–30 May 2023. The operator's own external review of that incident records that inconsistent DNSSEC information was visible for up to 19.5 hours for ac.nz and up to 14.75 hours for other .nz domains, that .nz DNS information was considered bogus until it expired or was flushed from third-party resolver caches, and that the root cause lay in a misalignment between the signing configuration and the DS-record TTL produced by the zone-build process. The same document states that over 85% of New Zealand Internet users reach the Internet through resolvers that rely on consistent .nz DNSSEC records. The 2023–24 page prints DNS availability 100% for that year. My arithmetic, and only mine: 19.5 hours is 0.22% of a 365-day year, and a figure printed as 100.00% cannot conceal more than 0.005%. Whatever that 100% measures, it does not count those hours as DNS unavailability. That is not a mistake in the arithmetic. It is the definition of the noun, and the noun is the thing nobody printed.
Two comparatives also disagree with themselves. EPP availability for 2024–25 reads 99.997% on the 2024–25 page and 99.99% in the 2024–25 column of the audited Table 2. WHOIS availability for 2023–24 reads 99.998% on that year's page and 99.99% in later printings. This reading found no restatement note in either document. Print both, adopt neither. The size of the first gap is 0.007 percentage points, which is my subtraction and says nothing about its cause.
So what does a registrar do with 100.00%? It can put the figure in its own marketing, which is what a rate is for. It cannot compute against it, cannot compare it with last year's figure taken from a different printing, and cannot set a customer expectation that will survive an incident the metric does not count. An enterprise resolver operator meets the same wall from the other side: the only instrument the operator names is external, and the link between what that instrument sees and what the annual statement prints is unstated. The audit does not close the gap either.
It certifies that the number conforms to a basis the operator chose and did not publish — a statement about consistency, not about legibility.
That is the finding, and it is the structural, unglamorous one. The duty to define what a hundred per cent counts sits with whoever prints it. This operator has discharged the duty to name the nouns and left the measurement where no reader can reach it. Four years of round numbers is not four years of nothing. It is four years of asking a reader to take the roundest possible number on trust.
What this reading did not find, counted over the twelve documents above, dated 22 November 2026:
- No reviewed page prints the measurement window for any annual DNS, EPP or WHOIS output. The 31 March year end is printed as the period of the statement, not as the window of the measurement.
- No reviewed page prints the denominator for any of the three outputs — no count of queries, requests, minutes or incidents.
- No reviewed page states who or what computes the figure: internal monitoring, an external probe network, the auditor's own procedures, or the external monitoring service the DNS page points to.
- No annual output row carries the maintenance carve-out. It is printed twice in this page set — against the 99.9% registry-system commitment, and against “Registry services availability (excluding scheduled maintenance)” in two product-strategy tables — and a reader cannot tell from the output rows whether it applies to them.
- No restatement or comparative-change note explains the 2024–25 EPP comparative at 99.997% on one document and 99.99% on another, or the 2023–24 WHOIS comparative at 99.998% and later 99.99%.
- The audited statement certifies conformity with the Group's own measurement bases and does not print those bases in the text of Table 2 or around it.
- No reviewed page says whether the DNS output measures authoritative nameserver response, resolution through a third-party resolver, end-to-end DNSSEC validation, or something else again.
- No reviewed page joins the architecture to the number: the nameservers, their operators and their anycast or unicast character sit on one page, the percentage on others, and no sentence connects them.
- No reviewed page prints a financial-year uptime figure from the status host, which offers a rolling ninety-day view instead, so a reader cannot reconcile the two.
- The full rendered body of the report and documentation pages, and the full text layer of the two PDFs, were not available to this reading. A definition could sit in material not read here, and where a printed absence carries the argument, that limit travels with it.
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.






