Summary

  • RIPE NCC's due-diligence document ripe-863, stated date 20 July 2026, says the registry must be kept "up-to-date and correct". The RIPE Database Terms and Conditions, Article 6.1, say the RIPE NCC "does not guarantee the accuracy, or completeness or availability of the RIPE Database". One publisher, one object of description, two promises.
  • The audit rules that check data quality print their own ceiling — a third party who reports a problem learns nothing "unless it is proven that the reporting party has a valid claim" — and what the registry publishes instead is coverage: 98.7% of members verified within the last five years, and one non-match share of 3.4% whose base is never printed.

A European IPv4 sale closes with a signed agreement, a registry update, and a buyer who now holds a block registered to a new holder in the RIPE NCC service region. What the buyer acquired is a registration. Whether the record describing the seller was ever true, and who checked it, is a different question with its own documents — and the registry has published them in pieces, at different dates, in different genres.

Start with the document that names the subject. Due Diligence for the Quality of the RIPE NCC Registration Data, Document ID ripe-863, updating ripe-791, carries a stated date of 20 July 2026 and opens with the strongest accuracy sentence anywhere in the set:

"The RIPE NCC has a mandate from the RIPE community to keep an up-to-date and correct Internet number resource registry. In order to comply with this mandate, the RIPE NCC performs due diligence on organisations the RIPE NCC registers Internet number resources for."

Read what the sentence governs. It is a mandate about a registry, discharged by performing due diligence on organisations. It is not a statement about any record's truth on any given day, and the document does not present it as one. What follows in ripe-863 is the mechanics: the registry publishes the name and country code of the legally registered party, the resources that party holds, and — for legal persons — the identifier issued by the national authority where the party is registered.

Changes must come from a registered contact with delegated authority through the LIR Portal, or from an authorised person with general authority to act for the holder. The registry reserves the right to demand further documentation, verify it with third parties, or require notarisation. After registration it audits, and it maintains procedures so holders know which changes must be communicated: transfers, official name changes, sponsoring LIR changes, LIR closure and deregistration.

That is a serious procedural document. It is also silent on the thing a counterparty usually wants, which is a statement about the reliability of a specific record rather than about the seriousness of the process that produced it.

Two sentences, one object

Now open the terms that govern the database the due diligence populates. The RIPE Database Terms and Conditions settle the question that ripe-863 leaves to inference. Article 6.1: "The RIPE NCC does not guarantee the accuracy, or completeness or availability of the RIPE Database or of the data contained therein." Article 6.2 puts the duty elsewhere — the Maintainer must keep the data it maintains accurate and up to date, including correct contact details. Article 5.1 reserves a power to the registry: it may correct or delete data for a listed set of reasons, among them that "any data is inaccurate" and that data was entered by someone not authorised to do so. Articles 7.1 and 7.2 put use of the database at the user's own risk and exclude liability broadly.

So four provisions distribute one property. The registry claims a mandate for correctness; the contract disclaims a guarantee of accuracy; the maintainer owes the accuracy duty; the registry holds the correction power; and the user carries the risk. Set the first sentence against the second and neither wins. The instructed reading is not that one is false. It is that they govern different things. The procedure promises a process — due diligence performed on organisations — while the contract disclaims an outcome — accuracy of data.

A buyer who quotes the mandate as an assurance is quoting a procedural commitment against a contractual refusal of the same property, in documents issued by the same organisation.

The dates matter here and the documents do not resolve them. ripe-863 is dated 20 July 2026, later than every other instrument discussed below. The Terms and Conditions carried no visible revision date in the text obtained, so which version was in force on 22 November 2026 is unknown. A reader can therefore establish what the current procedure says and cannot establish, from these pages, which contractual text it sits beside.

The audit regime prints its own ceiling

The procedure's enforcement arm is described in RIPE NCC Audit Activity, Document ID ripe-694, dated 10 January 2018. It states the registry performs audits after resources are registered "to actively check the quality and validity of the data in the RIPE Registry", and names three kinds. The Assisted Registry Check reviews a holder's legal name, address, contact details, registered contact persons and the correctness of the resource registration. "Selected" audits begin on suspicion of inaccurate data or policy violation. "Reported" audits begin on a third-party complaint supported by sufficient evidence.

Two sentences in the same document cap what that machinery can deliver to anyone outside the relationship.

"Unless it is proven that the reporting party has a valid claim on the Internet number resources under audit, the RIPE NCC will not report its findings to the reporting party due to confidentiality reasons."

And: it "should not be expected that the RIPE NCC will conduct a full audit of all registration data held by a Resource Holder" where that is unrelated to the reason the audit was initiated. The consequence clause is real — incorrect or fraudulent information may lead to termination of the agreement, deregistration and a referral to the authorities, and hijacking cases let the registry reverse or update data — but consequences are not disclosures.

Trace what that means for the one party with money at stake. A buyer who suspects the seller's registered legal name or address is wrong can report it. The registry may run a Reported audit. And unless the reporting party proves a valid claim on the resources themselves, it will be told nothing about what the audit found. The instrument that could answer the buyer's question exists, is staffed, and its output is withheld by rule. That is a distribution of information, chosen in writing: findings belong to the holder relationship, not to the market watching the record.

Whether that is the right policy for a registry with commercial members is an argument. The informational consequence is not an argument; it is printed.

What gets published instead: coverage

The registry does publish, abundantly. Three layers of it, and each measures something slightly different from accuracy.

Layer one is coverage. The Annual Report 2024, ripe-840, reports 2,445 Assisted Registry Checks against a committed target of 2,400, producing 5,500 corrective actions. Under a commitment still marked "In progress", it states that "98.7% of all members and 99.8% of all independent resource holders were verified within the last five years". A separate commitment — "Conduct an external audit of our Registry accuracy and implement any necessary improvements" — is marked "Completed". The same report explains what an ARC actually inspects: legal name, address and contact information, IPv4 and IPv6 registration details, and whether BGP announcements align with the Routing Registry.

Read the unit before the number. "Verified within the last five years" counts coverage of a five-year clock. It is not the share of records found to be correct. A record checked in 2021 and quietly wrong since 2024 sits inside the 98.7%, and nothing in the report claims otherwise. The figure is frequently repeated as though it were an error rate. It is a measure of how recently someone looked.

The planning document one step upstream is blunter about that. The Activity Plan and Budget 2025, ripe-835, dated December 2024, records that by the end of June 2023 94.3% of members and 92.6% of independent resource holders had been verified at least once within five years; sets a 100% objective for 2024; proposes compressing re-verification from five years to two; notes that an external audit report on the integrity of the registry was received in 2024, with some suggestions already implemented and completion expected in 2025; plans to extend automated monitoring of company-registration details to roughly 20,000 independent resource holders; and then writes the sentence that does the load-bearing work in this article:

"We will also refine our current metrics to better measure the accuracy of Registry data."

That is the publisher naming its own measurement gap in a budget document. Once the body that owns the metric says the metric needs refining in order to measure accuracy, the absence of a per-record accuracy measure stops being a reporter's inference and becomes a dated admission. It is also specific in scope: a plan published in December 2024 describes the state of the metrics at that moment. It does not tell a reader what was refined afterwards, and the later documents obtained here do not report a successor metric.

Layer two is the verification clock and its triggers. A Registry Update presented at RIPE 88 on 22 May 2024 gives the most granular public breakdown available: verify all member and End User data every five years, complete 2,400 ARCs a year, check against official sources covering legal name and address, registration or identification number and status, monitor member data against Dun & Bradstreet, log a last-verified timestamp, and bucket the population by recency. As of the second quarter of 2024 that bucketing put 66.3% of members inside two years, 31.2% between two and five, and 2.5% beyond five; for End Users the split was 53%, 46.6% and 0.5%. The same slide set counts 16,378 member verification events in the previous twelve months by trigger: transfers 4,770; organisational changes 3,615; Assisted Registry Checks 2,098; registry accuracy monitoring 1,737; LIR account activations 979; other 3,179.

Add those six triggers and they land exactly on the printed total: 4,770 + 3,615 + 2,098 + 1,737 + 979 + 3,179 = 16,378. That is my addition, and it is worth performing, because it means the one ledger a reader can rebuild here rebuilds cleanly. Now read the two largest triggers. Transfers and organisational changes together account for 8,385 of the 16,378 events — again my subtraction. Half of the verification activity is generated by the two events that most often immediately precede a change of control of address space.

Both readings are correct, and they pull opposite ways. Reassuring: the record is refreshed by the very transaction that makes it matter, so a buyer buying through a transfer is buying a record that was checked at the moment of transfer. Uncomfortable: the check ran because the seller's event triggered it, on the holder's side of the relationship, and nobody was checking on the buyer's behalf. Note also what the unit is. These are verification events, not distinct members and certainly not distinct records; one record can be verified more than once in a year and one member can generate several events.

The slide does not print a de-duplicated count, so 16,378 cannot be converted into a proportion of anything.

The same presentation states that an "Independent external review of the Registry conducted every two years", and names a February 2024 transfers integrity audit with Ernst & Young that concluded "No major venerabilities identified", with recommendations added to registry optimisation plans. The audit report itself was not obtained in this review. A quoted conclusion inside a vendor's own slide is a summary of a report by the party the report examined, which is not the same document as the report.

The one error-shaped number, and its missing base

Layer three is the closest thing in the reviewed set to an error rate, and it arrives without its denominator. The Registry Highlights presented at RIPE 90 on 14 May 2025 describes matching organisational members against national business registers through a third-party provider with worldwide coverage and automatic links to national registers. The result is printed as "Organisational members 3.4% = 687 members not matched", of which 196 are organisations. The same material reports the alert ledger: 2,389 Registry Accuracy Alerts, 1,430 updates made, 956 non-essential updates.

Take the percentage first. Three point four per cent of what? The base is not printed. If 687 is 3.4% of the organisational member population, that population is about 20,205 — an arithmetic implication of the printed pair, mine, not a printed figure, and not checkable against any other total in this set. Two denominators are conceivable: the population counted for the business-registry matching exercise, or organisational members as counted somewhere else.

A reader holding one number and one percentage cannot tell which population was divided, and the difference between those populations is exactly the difference between a small problem and a bigger one.

Then the nested count. If 687 members were not matched and 196 of them are organisations, then "organisations" is a subset of the unmatched and not the whole of it. The slide does not define the pairing, so a reader cannot tell whether the remaining 491 are a different legal form, a different register, or a different matching outcome entirely. A count and its subset are two facts; the relationship between them is a third, and it is not printed.

The alert ledger does not close either. 1,430 updates plus 956 non-essential updates is 2,386 against 2,389 alerts — three short, with no cause stated in the text obtained. That is my addition. Three rows in an operational summary is not a scandal and I am not calling it one. It is a reminder about genre: these are service-highlight slides, not reconciliations, and the same slide that misses three rows also prints a percentage with no base at all.

The measurement that did score records, once

The only instrument in this set that ever scored individual records is also the oldest. Registry Data Quality Assessment — Phase 2, describing an evaluation begun in 2009, states:

"Out of more than 3.4 million entries examined, 96.09% scored extremely high with an average score of 4.5/5. The remaining 3.91% scored an average of 3/5."

The partition closes: 96.09 + 3.91 = 100.00, my addition. The method cross-checked IPv4 space against the IANA registry, the RIPE NCC's own statistics files, other registries' statistics files, the Routing Information Service, the internal membership database and the RIPE Database, tagging each data set with a confidence score from 0 to 5. The lowest-scoring /8s were predominantly legacy space, including "orphaned" space with no current LIR listed — the segment a buyer of old address space is most likely to encounter.

The companion account, Ensuring Registry Data Quality, records the repair volume — 1,097 IPv4 inconsistencies fixed, 600 of them in legacy space; 2,126 ASN inconsistencies; 1,698 objects created, 428 holder names updated, 135 out-of-date objects deleted — and two disclosures that decide how much the 96.09% can carry. Phase one excluded the RIPE Database from the comparison. And the confidence scoring leaned on proxies: whether a record had been updated recently, and whether the registry had been in recent contact with the address holder. It also states that "An independent third party also audits the RIPE NCC's Registration Services Department regularly to help ensure data accuracy" — an assertion of recurring external scrutiny, with the auditor unnamed and no report published in the text obtained.

So the 96.09% is a score built partly from maintenance activity, produced by a project that began in 2009, in a post undated in the retrieved text. It is a dated measurement, not a standing status, and it is not attached to any named record. Nothing obtained since attaches a score, a date or a finding to a specific entry.

The derivation a stranger can attempt

Put the question a counterparty actually asks — was the record I am relying on verified, and what was found — and walk it through what is published.

Step one: find the record. The RIPE Database is public, and ripe-863 states what will appear in it: name, country code, resources, and a national-authority identifier for legal persons. This step works.

Step two: find the verification. The RIPE 88 material says a last-verified timestamp is logged and the population is bucketed by recency. That describes an internal reporting capability. Whether that timestamp is exposed on the record a stranger can query is not established anywhere in the sources reviewed — a statement about this review's sources, not about the registry's whole publication estate.

Step three: ask what the audit found. ripe-694's confidentiality sentence answers for every third-party route: without a proven claim on the resources, the findings stay inside the relationship.

The derivation therefore breaks at step two for want of a printed field, and at step three by written rule. What survives is aggregate — coverage percentages, activity counts, one non-match share without a base, one undated score from a project begun in 2009, and a conclusion quoted from an external audit whose report was not obtained. This is the finding, and it is about where reliance can be placed, not about competence. The documents describe a live programme: a 2,400-ARC target exceeded at 2,445, 5,500 corrective actions in one year, verification triggered by real events, and a plan to halve the re-verification cycle.

They also describe a registry whose own budget says its accuracy metric needed refining.

What a buyer can do with that

In a RIPE-region transfer the registry record is evidence of registration. It records who the registry has been told holds what, and that is a real and necessary thing to be able to prove. It is not evidence of accuracy, and the registry says so in Article 6.1 while its procedure says the checking happens inside a holder relationship the buyer is not party to.

So construct the assurance yourself. Obtain the holder's corporate documents and check the legal name against the national register the registry itself uses as its reference. Get the sponsoring LIR's written confirmation of the holder and of the resources. Put the accuracy warranty in the agreement, because the registry will not supply one, and put the correction obligation there too, because the registry's power to correct inaccurate data under Article 5.1 is a power it holds, not a remedy it owes to you.

Treat the registry's coverage figures as evidence that a process runs, which they are, and not as a probability that your record is right, which they are not.

What would change this reading

Four developments would each convert a process account into a checkable one. A per-record verification date, or verification status, visible in the registry's public output would answer step two. The report of the external registry-accuracy audit that the 2024 Annual Report marks "Completed" and the 2025 plan says was received would let a reader test what the external reviewer found rather than how its conclusion was summarised. A dated successor to ripe-694, restating or amending the confidentiality bar, would tell a complainant what the 2018 rule still governs in 2026.

And an accuracy series with its base printed would make the 3.4% computable.

Until any of them appears, the honest summary is a two-clocks one. ripe-863 sets out a 2026 procedure. ripe-694, dated 10 January 2018, predates the entire verification-metric programme described above, and neither text obtained says whether the older document still governs the newer procedure's complaint route. A reader should not assume the 2018 confidentiality sentence has been replaced, and should not assume it still binds. The successor, if there is one, was not obtained in this review.

Sources

All nine documents were reached as provider-reported excerpts in a single retrieval on 22 November 2026; none was opened end to end, and the two external audit reports they refer to were not obtained. The BTW directory entry for the due-diligence document is Due Diligence for the Quality of the RIPE NCC Registration Data.