Summary

  • A mean or median completion time describes only a centre. It cannot show whether a minority of IPv4 transfers spend months in repeated evidence requests, bilateral handoffs, legal review or unexplained inactivity. The 90th and 95th percentiles, tail-to-median ratio and maximum subject to privacy protection are necessary measures of institutional risk.
  • Completed cases are a selected population. If difficult requests remain pending, are returned for resubmission, are cancelled after a deadline or are withdrawn because the transaction can no longer wait, a completion-only median can improve while real access deteriorates. Open-case age and non-completion outcomes must be published beside completed-case duration.
  • A transfer needs several clocks: first submission to accepted completeness, completeness to decision, approval to record change and total elapsed time. Each interval must identify its controller - source party, recipient party, source registry, recipient registry, joint scheduling, legal hold or external authority - and preserve every switch rather than assigning the whole case to one side.
  • Resubmission must not erase history. Registries should link related tickets into one transfer episode, disclose the count and reason for each return, and report both the duration of the current ticket and the cumulative age from the first materially complete attempt. Otherwise difficult cases can be made statistically young by closing and reopening them.
  • Reason categories should distinguish incomplete identity evidence, signatory authority, resource eligibility, policy qualification, dispute, sanctions or court hold, fee or agreement, technical cutover, counterpart coordination, party non-response, refusal, withdrawal and administrative closure. These are analytical categories, not accusations of misconduct.
  • Number Resource Society can define an open event vocabulary, let parties export a signed case history, publish comparable distributions and audit bilateral handoffs without deciding who deserves a transfer. Its value would be portability and measurement, not a new approval queue.

A two-day claim can coexist with a six-month problem

Institutional delay often arrives dressed as reassurance. Most cases are completed promptly. The normal request is uncomplicated. The average is within target. The median applicant has little to complain about.

Each statement may be accurate. None answers the question that matters to the operator whose financing, deployment or customer migration depends on the transfer that is not normal.

The RIPE NCC says that, in most cases, an intra-regional transfer is completed within one or two business days once the parties have supplied the requested material. ARIN says that after it receives a signed Registration Services Agreement and all applicable fees, the resources will be transferred within two business days. APNIC has historically described a one-to-two-working-day turnaround for responding to resource requests, while warning that clarification can extend the process. These are useful commitments. They are also clocks with different starting lines.

RIPE's description appears near the complete-document stage. ARIN's two-day statement begins after approval conditions, fees and an agreement have been satisfied. APNIC's older statement concerned a response, not necessarily final registration. None, by itself, measures the period from a party's first serious submission to final record change. A buyer can therefore experience a long transaction while every quoted service interval remains formally intact.

This is not a semantic objection. It is the central accountability problem. The longest cases are usually the ones in which the institution exercises the most consequential judgement: whether a historic holder is the same legal person, whether an officer can bind it, whether a resource is eligible, whether a recipient qualifies, whether a dispute is active, or whether two regional rule sets can be reconciled. Those are exactly the cases in which members need to know how authority is used.

A single median conceals them by design. The median is the point below which half the observations fall. It can remain two days if 51 cases take two days and 49 take half a year. It can improve if staff complete more easy cases while old ones wait. It can improve again if those old requests are closed and resubmitted under new ticket numbers. A statistic intended to reassure can become least informative where institutional risk is greatest.

The answer is not to discard the median. It is to stop asking it to represent the distribution. A credible transfer-time report should show the centre, the tail, the unfinished stock and the movement of control between them.

The denominator changes while the institution is speaking

There are at least four populations behind any claim about transfer speed. The first is every inquiry. The second is every request formally submitted. The third is every request accepted as complete enough for substantive review. The fourth is every request completed during the reporting period. Their durations and outcomes are not interchangeable.

Completion statistics usually describe the fourth population. That choice has a respectable operational purpose: a finished case has an observed end date. It also creates selection. A case open at quarter end has not yet contributed its eventual duration. The longer it remains open, the longer it remains absent from a completion-only calculation. Statisticians call such observations censored because the outcome is not yet fully observed. Administrators are more likely to call them pending. Either way, excluding them makes the reported population systematically easier than the live caseload.

Consider a desk that completes 90 routine transfers in ten days and carries ten disputed transfers into the next year. Its median completed time is ten days. Its 90th percentile may also be ten days, depending on the calculation convention. Yet one in ten of the period's materially live cases is already older than the published experience. If the same ten cases are still open next quarter, another fast median can be announced. The institution appears consistently efficient precisely because its hardest work never reaches the denominator.

Withdrawal creates a second selection. A commercial agreement may expire, financing may lapse, a deployment may move elsewhere or a counterparty may lose patience. If the parties withdraw after 120 days, a completion-only series never records the burden. Labelling the case a private commercial failure does not remove the possibility that administrative uncertainty contributed to it. The registry need not decide causation. It must preserve duration and allow the party to choose a broad withdrawal reason.

Refusal creates a third selection. A decision reached after long review is an institutional outcome even though no transfer appears in the public completed-transfer record. Cancellation creates a fourth. APNIC's current guide says that a recipient must acknowledge a source-initiated transfer within 30 days or the request will be cancelled. That is a clear rule, but the resulting case belongs in the duration account: initiated, waiting for recipient, cancelled at day 30. Omitting it because no block moved would turn an explicit service event into statistical silence.

Intake screening creates a fifth. A request returned as incomplete may never enter the substantive clock. If requirements are clear and the omission obvious, that separation is defensible. If a reviewer asks for successive pieces of evidence over several cycles, treating each cycle as pre-clock activity can hide the institution's contribution. The report needs both first-submission age and accepted-complete age.

The denominator should therefore be stated, not assumed. Each published number must say whether it covers inquiries, submitted requests, complete requests, decisions, record changes or all closed cases. A quarterly reconciliation should show how cases move between those states. Without that bridge, a fast median is an assertion about a population selected by the producer.

The mean and the median tell different half-truths

The mean is often criticised because one extremely old case can pull it upward. The median is praised because it is robust to extremes. Both observations are statistically correct. The institutional conclusion commonly drawn from them is not.

If the purpose is to describe the experience of a typical completed case, the median is useful. If the purpose is to measure total waiting burden, the mean has information the median discards. Fifty transfers taking two days and one taking 300 days create a median of two, but the exceptional case accounts for more waiting time than all the routine cases combined. That burden is real even if it is not typical.

The NIST statistical handbook describes the median as the 50th percentile and notes that percentile estimates can vary slightly according to the interpolation method used, particularly in small samples. That technical detail matters in a thin directed inter-regional route. A registry should identify whether it uses, for example, the R7 or R8 convention, how it handles tied dates and whether a partial business day is rounded. Otherwise two institutions can publish nominally identical percentiles from the same small population and obtain different values.

The minimum distribution should include case count, the 25th, 50th, 75th, 90th and 95th percentiles, arithmetic mean and a privacy-protected maximum or oldest-case band. The 25th and 75th percentiles show the central spread. The 90th and 95th show the tail. The mean captures cumulative burden. The maximum, where a case cannot be identified, tests whether an ancient file has been left outside normal management.

Two derived measures are especially revealing. The first is the tail-to-median ratio: the 90th percentile divided by the median. A median of two days and a 90th percentile of 20 gives a ratio of ten. A later median of one day with the same 90th percentile gives a ratio of 20. Routine automation improved while tail governance did not. The second is excess tail days: the number of calendar days above a declared service threshold, summed across cases. It indicates how much waiting is concentrated outside the promise.

No composite score should merge these measures. A desk can have a good median and a poor tail. Another can have a slower median but a compressed, predictable distribution. For a buyer coordinating a financing deadline, predictability may matter more than the typical value. Members should see both rather than a managerial grade chosen by the institution.

The distribution also needs confidence about sample size. A 95th percentile based on 12 cases is unstable. Publication should show the count and, for thin cells, combine periods or display age bands instead of spurious precision. Suppression must be visible. A blank cell should mean too few cases under a stated rule, not an inconvenient result.

The tail is where discretion becomes a cost of entry

Delay is not distributed neutrally. Large repeat entities can maintain counsel, specialist staff and alternative inventory. A smaller operator may have one planned acquisition tied to a data-centre opening, customer contract or migration window. The same 90-day uncertainty imposes different economic damage.

That makes tail time a membership-accountability question, not merely a customer-service metric. If old cases cluster among first-time recipients, legacy holders, cross-border parties or organisations with uncommon legal forms, the distribution reveals which kinds of entity bear the institution's interpretive uncertainty. A global median can conceal that pattern because routine transfers by experienced repeat users dominate the centre.

Segmentation should be limited to factors that plausibly change the work: intra- or inter-RIR, source and recipient registry, transfer policy type, resource class, presence of pre-approval, first-time or repeat entity, ordinary or disputed identity, and enhanced legal review. It should not publish names or sensitive allegations. It should also avoid creating a discretionary category called complex that simply absorbs every delayed case.

Complexity needs an entry rule. A case might enter enhanced review when there is an active third-party claim, a court order, sanctions screening match, dissolved holder, contested succession, inconsistent public incorporation record or specified legal escalation. The event should be timestamped. If the issue clears, the case leaves the category. The published report should show how many cases entered, exited, completed and remained pending.

The distinction between complexity and queueing is crucial. A difficult legal question can justify more work. It does not justify an unmeasured interval in which nobody acts. Controller clocks can show that a case required 40 days of party evidence collection, seven days of registry analysis and 30 days awaiting a counterpart registry. That is a more credible explanation than saying complex cases take longer.

Nor should careful review be treated as failure. A fast but careless transfer can create conflicting records, routing-security disruption or fraud risk. The objective is bounded, explainable time, not indiscriminate speed. Tail reporting makes careful review defensible because the institution can demonstrate what happened, who had the next action and whether similar cases received similar treatment.

The opposite is also true. If two legally comparable cases take radically different times without a recorded reason, the distribution points toward inconsistent administration. Members can ask for process repair without seeing confidential documents. Equality of treatment is not proved by identical outcomes; it is tested by comparable stages, reasons and clocks.

One transfer contains several clocks

End-to-end time is indispensable, but it is too coarse to diagnose a delay. A transfer should be represented as a sequence of common events even when regional procedures differ.

The first event is initial submission: the earliest request that identifies the source, intended recipient and resource set well enough to show a real proposed transfer. The second is intake response: acknowledgement, rejection as malformed or a specific request for missing material. The third is accepted completeness: the desk confirms that substantive assessment can begin, without implying approval.

The fourth group covers source review: holder identity, signatory authority, resource status, minimum size, holding period, dispute and any applicable agreement. The fifth covers recipient review: identity, account status, policy qualification, pre-approval and other applicable conditions. The sixth is institutional decision: approved, refused, returned for a material amendment or held under a stated external constraint.

The seventh is bilateral handoff for an inter-RIR case: notice sent, notice accepted, counterpart questions issued, counterpart decision returned and joint record-change date agreed. The eighth is closing conditions: fees, agreements, acknowledgements and cutover material. The ninth is record change: recognised registration updated and completion notice issued. The tenth is technical confirmation: any expected record, routing-security or reverse-DNS actions checked, while making clear which are party responsibilities rather than prerequisites to recognised transfer.

These events produce four headline durations. Intake time runs from initial submission to accepted completeness. Decision time runs from completeness to institutional decision. Closing time runs from approval to recognised record change. End-to-end time runs from initial submission to record change or non-completion. Publishing only closing time would let the long review disappear; publishing only end-to-end time would make diagnosis impossible.

ARIN's public transfer instructions illustrate why boundaries matter. Source and recipient submit separate requests for a specified-recipient transfer. Staff link the tickets, invoice a processing fee, evaluate the requests independently, then issue other fees and an agreement where applicable. The final two-business-day commitment begins only after ARIN has the signed agreement and applicable fees. A reader who sees only that final commitment cannot infer how long the linked review took.

RIPE's inter-RIR instructions provide another example. After one registry approves, the other evaluates its side; both then update records on an agreed date. RIPE says inter-RIR transfers require more work because policies differ, time zones intervene and registry updates must be synchronised. That account identifies plausible stages. It still needs observed distributions for each stage if members are to know whether the extra time lies in substantive review, handoff or scheduling.

A common event vocabulary would not force every RIR into the same internal procedure. Each can map its events to the shared milestones and publish exceptions. Comparability requires common meanings, not identical offices.

Every day should have a controller

The phrase processing time implies that the institution is processing continuously. Often it is not. The file may be waiting for the source, the recipient, the other registry, a payment, a legal authority or a coordinated cutover. Fair measurement must recognise those changes.

Each case should have exactly one primary controller at any moment and may have secondary observers. The primary categories are source party, recipient party, source RIR, recipient RIR, both RIRs jointly, both transaction parties jointly, external legal authority and scheduled technical transition. A separate system-hold category can cover a documented outage, but it should not become a convenient residual.

Control changes when a specific action request is sent or satisfied. A request for documents should state what is needed, who owes it and the deadline. The clock passes to that party when the notice is delivered. It returns to the registry when the response arrives. If the response is partial, the registry should record whether the original request was unsatisfied or whether it has introduced a new requirement.

That distinction addresses a common source of resentment. From the institution's view, a case may have spent 60 days waiting for the applicant. From the applicant's view, it may have answered five successive questions within a day each while the desk took a week to formulate every next question. Both can honestly describe the other as the source of delay. An event history resolves the dispute: party-controlled hours, registry-controlled hours and the number of question cycles.

Parallel work needs a rule. If source review and recipient review can proceed simultaneously, calendar time should not be double counted. Publish elapsed calendar time once, plus controller-attributed effort intervals that reconcile to it. Where two actors can act independently, use a joint-open state and identify which required actions remain. The public aggregate can show joint days without disclosing documents.

Weekends and holidays should remain in the end-to-end clock because commercial commitments do not pause. Business-day measures can be published as an operational supplement. Each registry should publish its holiday calendar and cut-off rule. Inter-RIR cases should use UTC event times and report calendar days globally, avoiding arguments over whose business day applied.

Controller attribution should never decide legal blame automatically. A court hold is not the court's fault; it is a factual state. A party-response interval may reflect an unusually broad request. Registry time may reflect necessary verification. The purpose is to locate the interval so the reason can be assessed, not to manufacture a league table of culpability.

Resubmission must not make an old case young

Administrative systems are built around tickets. Transactions are not. A buyer and seller may pursue the same transfer through several tickets because a request was returned, a pre-approval changed, a document expired, a legal entity was corrected or the inter-RIR route required separate submissions.

If every new ticket resets the clock, performance can improve through clerical multiplication. The institution closes an old request as incomplete; the party opens a replacement; the completed ticket later shows a short duration. The user experienced one long transfer episode. The report describes two unrelated files, one perhaps excluded as an administrative closure and one celebrated as fast.

The durable unit should be a transfer episode. It begins with the first materially identifiable attempt by the same source, recipient and resource set. It ends with completion, final refusal, voluntary abandonment or a material change of parties or resources. Tickets, pre-approvals and bilateral references sit beneath it.

A material change needs a published test. Correcting a company suffix, replacing an expired certificate, dividing a resource set while preserving the transaction or supplying a clearer signature should not start a new episode. Replacing the recipient, substituting unrelated resources or entering a new commercial agreement after final abandonment may. When uncertain, the system should preserve the link and mark the point at which analytic treatment changes.

The report should show ticket count per episode, requests for further information, resubmission count and cumulative age. It should also show the duration of the successful ticket for operational debugging. Both views are legitimate; only the cumulative view reflects the entity's exposure.

Repeated requests need reason codes. The first return may concern missing corporate registration, the second signatory authority, the third resource eligibility. If the desk could reasonably have asked for all three at once, serial review is an institutional design problem. If new facts emerged, the history will show that. Publishing aggregate cycle counts creates an incentive to front-load clear requirements without pressuring analysts to approve weak evidence.

Document expiry deserves special treatment. If a registry asks for a recent corporate record and its own delay causes the record to age out, the replacement interval should not be attributed wholly to the party. The event history can mark registry-induced renewal. This is a small rule with large fairness value, particularly where obtaining official documents is slow or costly.

No party should be able to erase a prior refusal by opening a nominally new case, either. Episode linkage protects institutions as well as applicants. It reveals repeated attempts, policy changes between attempts and whether a later approval followed new evidence or a different interpretation.

Refusal, cancellation and withdrawal are time outcomes

Transfers that do not complete still consume institutional capacity and entity time. A proper duration report gives them status rather than treating them as noise.

The top-level outcomes should be approved and registered, approved but not closed, refused, withdrawn by source, withdrawn by recipient, mutually withdrawn, cancelled for non-response, administratively replaced, stayed by an external order and pending. Every closed outcome should retain the end-to-end age and controller history.

Reason categories need enough detail to support reform but not enough to identify a party. A workable list includes: source identity not established; source authority not established; recipient identity not established; resource registration conflict; resource policy restriction; recipient policy qualification; active ownership or control dispute; sanctions or legal restriction; incomplete agreement or fee; counterpart RIR incompatibility; counterpart RIR awaiting action; technical transition issue; source non-response; recipient non-response; mutually requested pause; and commercial withdrawal with no further reason.

These labels should describe the decisive administrative condition, not allege moral fault. Identity not established does not mean fraud. Non-response does not mean bad faith. A resource conflict does not decide ownership. Commercial withdrawal does not prove that the institution caused the deal to fail. The public value lies in counting recurring conditions and their ages.

Secondary tags can preserve complexity. A case may be refused because resource eligibility was not established after a dissolved holder's succession documents proved limited public evidence. The primary reason is resource eligibility; secondary tags are historic succession and document insufficiency. Public cells should be suppressed where a rare combination would reveal the party.

Appeal or reconsideration must remain linked. Report time from original submission to original decision, decision to review request, review duration and final outcome. Do not replace the first decision with the later one. A reversal after 200 days is evidence that both the initial interpretation and review mechanism matter.

APNIC's 30-day acknowledgement rule offers a clean example of why reason matters. A cancelled request at day 30 should appear under recipient non-acknowledgement, not registry refusal. ARIN's separate linked tickets create another: one side may qualify while the other remains incomplete. The episode outcome and each party-ticket status should both be retained.

A small share of cases will resist classification. Publish an other category, review it quarterly and disclose the leading themes once privacy permits. If other becomes a large bucket, the taxonomy has failed. If no case ever appears there, analysts may be forcing diverse facts into categories designed to flatter the process.

Pending age is the unpaid bill

Completed-case percentiles look backward. Pending age shows the obligation still outstanding. A transfer desk that reports one without the other is presenting paid invoices while leaving the unpaid stack off the balance sheet.

At each reporting date, the institution should publish the number of open episodes in age bands: under 3 calendar days, 3-7, 8-14, 15-30, 31-60, 61-90, 91-180, 181-365 and over 365. It should show the oldest pending age where privacy permits, the current controller and the principal pending reason. Thin categories can be combined across quarters.

The bands should be based on first submission and, separately, accepted completeness. A case can then be 70 days old but only ten days into substantive review, revealing 60 days of intake churn. That is not necessarily registry delay, but it is part of the user's experience and a candidate for clearer guidance.

Pending age should be a daily snapshot, not reconstructed only from quarter-end survivors after clean-up. Publish the quarter's average and maximum open stock, entries, exits and age transitions. A case moving from 31-60 to 61-90 should be visible even if it completes before the next quarter end. Otherwise management can time closures around reporting dates.

The RIPE NCC already demonstrates the principle in a different IPv4 service. Its waiting-list page displays the number of LIRs in queue and the days the first LIR has been waiting, updated every three hours. A waiting list is not a transfer desk, and its queue discipline is different. The point is narrower: an Internet registry can present current stock and oldest age publicly when it considers those facts relevant.

Old cases need a management protocol. At 30, 60, 90 and 180 days, an independent reviewer should confirm the controller, reason, last substantive action and next due date. That review should not force approval. It should prevent a file from remaining open without ownership. Aggregate results - cases reviewed, reasons corrected, stale holds removed - belong in the quarterly report.

Parties should receive the same age statement visible to management. It should show episode start, current stage, current controller, outstanding action and cumulative days by controller. A user should not need to telephone to learn whether a transfer is waiting for its registry, the counterpart or itself.

Pending stock is also a capacity signal. If arrivals remain stable while old cases accumulate, the desk has either limited public evidence capacity or a policy bottleneck. If stock rises because a legal event pauses a class of cases, the reason distribution will show it. Headcount and policy debate can then follow evidence rather than anecdotes.

Inter-RIR delay needs a paired clock

Cross-regional transfers are particularly vulnerable to statistical evasion because each institution sees only part of the case. The source RIR can say it approved promptly and sent the request onward. The recipient RIR can say it acted promptly once a complete referral arrived. Weeks between those claims can belong to nobody.

Every inter-RIR episode needs a shared pseudonymous reference. It should link the source decision, handoff package, receipt acknowledgement, recipient questions, response, recipient decision and agreed update. The reference need not be visible in the named transfer record. It must be available for joint aggregation and audit.

The handoff clock starts when the source RIR sends the complete bilateral package under the agreed protocol. The recipient RIR should acknowledge receipt automatically. If it considers the package incomplete, it should respond with a reason code within a stated interval. The clock then moves to the actor that owes the missing material. Silent handoff is not a legitimate state.

Direction matters. ARIN-to-RIPE may have different evidence requirements from RIPE-to-ARIN. RIPE's guidance notes that, for transfers from ARIN, a recipient may have to supply a plan to use at least 50% of the resources within five years. ARIN applies its own recipient requirements to incoming cases and requires reciprocal compatible policy for its paths. A global inter-RIR median would erase these directional conditions.

Publish distributions by directed RIR pair where case counts permit. When cells are too small, combine several quarters or disclose broad bands. Never fold a thin route into a global number without naming the combination. Entities deciding whether a particular path is viable need the path, not the global institution's typical experience.

The two RIRs should sign a quarterly reconciliation: opening shared pending episodes, new handoffs, completions, refusals, withdrawals and closing pending episodes. Unmatched references should be counted and aged. A disagreement over status should appear as unreconciled, not be silently assigned to whichever institution publishes first.

Joint scheduling time deserves its own category. Synchronised updates reduce conflicting registration and are operationally sensible. If the parties choose a future cutover for network reasons, that period should not look like desk inactivity. If the RIRs can update only on limited days because of internal staffing, members should see the constraint.

Comparability will expose real policy differences rather than pretending they are mere speed differences. That is useful. A route may be slower because one side conducts a needs assessment, another because legal documents require translation, another because the bilateral channel is manual. The report should connect time to reason while refusing to turn speed alone into virtue.

Case mix must be visible but cannot become an excuse

No serious comparison would place a routine intra-regional transfer beside a contested historic succession and call the faster registry better. Equally, no serious institution should label every old case complex and declare its headline service healthy.

The answer is stratification with stable rules. At minimum, report intra-RIR specified transfer, inter-RIR outgoing, inter-RIR incoming, merger or reorganisation, legacy or historic-resource transfer, ASN transfer and other resource classes separately. Within each, show ordinary, enhanced identity review, active dispute, legal or sanctions hold and technical closing.

Pre-approval should be disclosed because it moves work earlier. APNIC explicitly presents transfer pre-approval as a way to avoid unexpected delay; once the source is found and the amount is within approval, need justification does not have to be supplied again. A fast post-match transfer with pre-approval is not directly comparable with a case in which recipient qualification begins after the parties are ready. The report should show pre-approval time, transfer time and combined elapsed time where the approval was obtained for a specific intended acquisition.

Likewise, ARIN's final two-business-day period after agreements and fees should remain a closing measure. It should not be compared with RIPE's statement that most complete transfers finish in one or two business days unless the starting conditions are aligned. Metric labels should contain the event names, not just the word processing.

Volume weighting should be resisted. A /16 does not necessarily require 256 times the administrative work of a /24, and address count says little about legal complexity. Publish case-weighted duration. Block size can be a segment if evidence shows a process difference, but it should not dominate the service statistic.

Entity identity needs privacy. First-time versus repeat is useful because repeat users may understand evidence demands better. Publicly naming slow entities is unnecessary and could deter legitimate review. The institution can publish aggregate gaps and improve guidance. Independent auditors can test whether legal form, geography or entity size predicts delay after controlling for case type, without releasing names.

Every segmentation rule should be set before the quarter closes and retained in revision history. If a definition changes, publish both old and new series for an overlap period. Otherwise a desk can improve its headline by moving old cases into a newly excluded class.

The objective is not a perfectly adjusted ranking. It is a faithful map of where time goes. Institutions can disagree about policy while sharing the duty to measure its administration.

Public institutions already know how to show a tail

There is nothing exotic about publishing response-time distributions. Other bodies handling sensitive requests do it without releasing case files.

ICANN's August 2024 Registration Data Request Service metrics reported response-time distributions by outcome, including the 10th, 30th, 50th, 70th and 90th percentiles and the average. In that report, approved requests had a median response time of two days and a 90th percentile of 20; partially approved requests had a median of four and a 90th percentile of 26. The figures are not evidence about IPv4 transfers. They demonstrate the difference between the typical case and the tail, and they show that Internet-coordination institutions can publish both by status.

The UK Planning Inspectorate's February 2026 experimental statistics use the 25th, 50th, 75th and 90th percentiles for decision times. Its stated ambition is not merely to reduce the median, but to have the 90th percentile fall faster so that the gap contracts. The report also discloses counts, maxima and known data exclusions. Again, planning appeals are not number-resource transfers. The institutional lesson is that the longest cases can be an explicit performance entity rather than an embarrassment omitted from the headline.

These examples improve on a single target in three ways. They preserve the distribution, divide outcomes and identify data limitations. A transfer report should do the same while adding controller and pending-age information specific to a bilateral transaction.

Service standards should be framed as a distributional promise. For example: 50% of complete ordinary intra-RIR cases decided within two calendar days, 75% within five, 90% within ten and 95% within 20; no complete case without a recorded controller and next action; every case over 30 days independently reviewed. The numbers here are illustrative, not recommended global thresholds. Each institution should propose targets from observed capacity and member need, then expose performance.

Averages remain useful for staffing, but should never be the sole public measure. Satisfaction scores are useful for sentiment, but respondents can be selected and a fast routine case can dominate. Ticket counts are useful for workload, but a closed ticket may represent approval, refusal, cancellation or replacement. Each measure answers a narrower question than transfer access.

The strongest report is one that makes bad news interpretable. A rising 90th percentile with a stable median may indicate an emerging legal backlog. More recipient-controlled days may indicate unclear instructions. More bilateral days may indicate a failing interface. Publication turns a complaint into a repairable process.

Privacy does not require statistical darkness

Transfer files contain corporate documents, signatures, contracts, legal disputes and security-relevant records. None should be published merely to prove that a case was slow. Aggregate accountability can preserve confidentiality.

The first safeguard is minimum cells. Do not publish a percentile for a directed route and reason class with fewer than a stated number of episodes. Five may be adequate for broad age bands; ten or more is safer for percentiles. Combine adjacent quarters before combining materially different case types. State when a value is suppressed.

The second is delayed reason publication for rare events. A court-held case may be publicly identifiable. Its age can be included in the global pending distribution while the detailed reason is reported later or grouped under external legal hold. Privacy protects the party; it should not remove the days from the total.

The third is pseudonymous audit. A public release can include a digest of the underlying event set and a confidential reviewer can reproduce the percentiles, controller totals and reconciliation. Parties should receive their own event history and may challenge misattribution. The reviewer publishes error rates and corrections, not documents.

The fourth is purpose limitation. Timestamps collected for service accountability should not become a public ranking of applicants or a new evidentiary demand. An institution does not need a party's commercial rationale to record that it answered on Tuesday. Nor should a entity have to waive confidentiality to contest a clock.

The fifth is retention discipline. Event metadata should survive long enough for trend and appeal analysis, while copies of sensitive evidence follow the applicable retention rule. The public statistical series should remain available even after case documents are lawfully destroyed. Aggregate history belongs to members.

Small registries or thin routes face genuine re-identification risk. They can publish global distributions, rolling four-quarter windows and coarse pending bands, while an independent body receives the detailed events. Complete silence is the least defensible option because it makes the smallest and potentially most fragile services impossible to evaluate.

Targets can be gamed unless the event history is preserved

What gets measured gets managed; it also gets rearranged. A transfer-time standard should anticipate predictable evasions.

One is delayed admission. Staff leave a request in intake until it is easy, then start the substantive clock. The countermeasure is dual publication from first submission and completeness, plus intake cycle counts. Another is premature closure. A ticket is cancelled shortly before breaching a target and reopened later. Episode linkage and cumulative age prevent the reset.

A third is controller dumping. A vague request for information passes the clock to the party while the registry remains uncertain about what it needs. Action requests should be specific, versioned and auditable. An independent reviewer can sample whether controller changes correspond to actionable obligations.

A fourth is class migration. Old cases are moved into special review after they become slow. Entry into an enhanced class should be based on a timestamped triggering condition, and the report should show performance before and after entry. The original ordinary-class age should not vanish.

A fifth is easy-case prioritisation. Staff improve the median by completing new simple cases while the old tail grows. Pending-age bands, oldest age and excess tail days reveal the trade. A sensible operating rule should reserve capacity for old cases without forcing analysts to decide them incorrectly.

A sixth is calendar manipulation. The published period excludes holidays or begins after fee payment even though the entity was ready earlier. End-to-end calendar days provide the stable public measure; business days can diagnose staffing. Event definitions should not move without a versioned change.

A seventh is denominator fragmentation. Source and recipient tickets are counted as two successful closures, or a bilateral case is counted by both RIRs. Transfer episodes and shared references keep the case count coherent. Ticket statistics remain available for workload but must not masquerade as transactions.

The final defence is revision history. If an event is corrected, preserve the old value, new value, reason, date and effect on published metrics. A transfer-time series without revisions can be silently polished. A service confident in its integrity should show how its data improve.

A minimum quarterly transfer-time statement

The reporting standard can be demanding without being bloated. One quarterly statement, accompanied by machine-readable aggregates, should contain six panels.

Panel one: population reconciliation. Opening pending episodes, new first submissions, accepted-complete episodes, registered transfers, refusals, withdrawals, cancellations, replacements and closing pending episodes. Show episode counts separately from tickets and resource records. Explain unmatched changes.

Panel two: completed and closed distributions. For each major case class and outcome, show count, mean, 25th, 50th, 75th, 90th and 95th percentiles for intake, decision, closing and end-to-end calendar time. Identify the percentile method, rounding, period and suppression threshold.

Panel three: pending stock. Show first-submission and complete-case age bands, oldest protected age, current controller and pending reason. Include entries into and exits from each band during the quarter, not just a closing snapshot.

Panel four: controller time. Show calendar days under source, recipient, source RIR, recipient RIR, joint RIR, joint party, external legal and scheduled transition control. Report the number of controller switches and further-information cycles. Reconcile totals to end-to-end duration.

Panel five: reasons and reviews. Show intake returns, refusals, cancellations, withdrawals, enhanced-review entries, appeals, reversals and old-case management reviews by broad reason. Link administrative reason categories to the relevant public rule or service step.

Panel six: bilateral integrity. For every directed inter-RIR pair above the privacy threshold, show handoffs sent, acknowledged, returned incomplete, approved, refused, unmatched and pending, with handoff and joint-scheduling distributions. Both institutions should publish the same reconciled totals.

Each release needs definitions, extraction time, known gaps, prior-release digest, corrections and a contact for challenges. The narrative should explain large movements without selecting anecdotes as proof. Data tables should remain downloadable in dated versions.

No price, valuation or commercial term is required for this standard. It measures the administration of recognised transfer requests. That narrower focus is a strength: registries already generate the events as they communicate with parties, assess cases and update records. The principal barrier is willingness to make the tail visible.

Number Resource Society should make the clock portable

Comparable measurement need not give one organisation power over transfer decisions. Number Resource Society can support a thin accountability layer whose rules are public and replaceable.

Its first contribution should be an open event dictionary: episode, submission, completeness, action request, controller change, decision, bilateral handoff, closing condition, record change, refusal, withdrawal, cancellation and review. Each event needs a timestamp, actor class, reason and link to the prior event. Any registry or independent service should be able to implement it.

Its second contribution should be entity portability. A buyer or seller should be able to export a signed history of its own case, including every request, response acknowledgement, controller change and decision. Confidential evidence can remain encrypted or referenced by digest. If the entity changes adviser or seeks review, it does not depend on screenshots from one ticket portal.

Its third contribution should be comparative publication. NRS can receive signed aggregate returns, test arithmetic, reconcile bilateral episodes and publish distributions under the same definitions. Where a registry does not supply events, it can mark the gap rather than estimate a flattering number. Where parties voluntarily supply histories, it can report the sample and coverage without calling it complete.

Its fourth contribution should be audit. A rotating independent reviewer can sample event attribution, resubmission linkage and privacy controls. Findings should be public; case documents should not. Corrections should be append-only so no institution, including NRS, can rewrite a poor quarter.

The boundary must be explicit. NRS should not set a mandatory global duration target, decide whether evidence is sufficient, accelerate members who buy a premium service, require its certificate for registration or direct parties to preferred facilitators. It should not turn a measurement vocabulary into another queue.

Portability is the safeguard against that outcome. Definitions should be openly licensed. Parties should be able to take their event history to another auditor. Registries should be able to publish directly. If NRS stops serving users, the records and methods should remain usable elsewhere.

The positive institutional proposition is modest: a common clock, a verifiable history and a public comparison. That is enough to make monopoly administration contestable without pretending that measurement itself transfers the resource.

The first honest report can be issued before the perfect one

Registries need not wait for a global agreement. Each can begin with events it already records.

In the first quarter, publish opening and closing pending counts, registered, refused, withdrawn and cancelled outcomes, plus first-submission and complete-to-close medians, 75th and 90th percentiles. Add pending-age bands and oldest protected age. State every missing field.

In the second, introduce controller categories and resubmission linkage. Give each active party a case-age statement. Audit all cases older than 90 days and publish the reasons in protected groups.

In the third, reconcile directed inter-RIR handoffs with each counterpart. Publish unmatched cases and joint scheduling. Do not wait until every pair agrees; mark which pairs can and cannot reconcile.

In the fourth, commission independent validation of one full year. Test event completeness, clock resets, controller accuracy, privacy leakage and consistency between ticket, episode and completed-transfer counts. Publish management responses and revised data.

Targets should follow the baseline, not precede it. An institution that sets an easy median target before measuring its tail risks optimising the wrong thing. Members should ask first for a year of faithful distribution, then debate acceptable service levels by case class.

The initial release may look worse than the present claim. That is not evidence that service deteriorated. It is evidence that the unobserved queue has entered view. Improvement begins when an old case can no longer disappear behind a typical one.

Governing the tail

The transfer median is not false. It is incomplete in a direction that consistently favours the institution producing it.

It favours completed over unfinished cases, new tickets over old episodes, routine users over uncommon legal situations and one registry's clock over the interval between two. It turns withdrawal into absence, cancellation into clerical closure and repeated evidence demands into applicant time. Most importantly, it treats the cases involving the greatest discretion as statistical exceptions rather than the place where accountability is most necessary.

The reform is a distributional record. Publish the centre and the tail. Count the pending stock. Preserve first-submission age. Link resubmissions. Assign every day to a controller. Categorise returns, refusals and withdrawals without accusing the parties. Reconcile bilateral handoffs. Keep corrections visible.

Then ask the right performance question. Not: how quickly does the typical successful case finish? Ask: what is the probability that a materially complete transfer will remain unresolved after 10, 30, 90 or 180 days, who controls those days, and what reason explains the state?

That question matches the risk faced by an operator. It also gives a careful registry a fair chance to show its work. Necessary legal review can be distinguished from idle queueing. Party delay can be distinguished from serial institutional questions. Coordinated cutover can be distinguished from an unowned handoff.

The public does not need confidential documents to see whether a transfer desk is predictable. It needs the shape of time.

Sources