Summary

  • Skeeve Stevens argued that an APNIC account holder with an IPv6 /32 could need a second allocation for an unconnected network because a smaller route from the first block might be filtered. APNIC 29 declined his more specific 2010 proposal after operators challenged the premise, the use of address space, and the proposed remedy.
  • Stevens returned at APNIC 31 with a broader test: the holder would need to announce its existing allocation and show APNIC a compelling reason for more. The meeting reached consensus in February 2011; APNIC's Executive Council endorsed the proposal in May and implemented it in August.
  • The exchange left a durable control question. APNIC could set allocation criteria and instruct hostmasters, while network operators controlled which routes their routers accepted. The adopted rule moved some judgment from a narrow written exception into APNIC's review of each request.

A /32 with nowhere to go

The comic scale of IPv6 creates a trap for the uninitiated: a /32 contains an enormous number of addresses, so why would its holder ask for another? Stevens's answer was topology. His 2010 proposal described an operator with one network in New Zealand and a separate one in Singapore. There was no private backbone between them. The second network had a local transit provider. The operator could carve a smaller prefix from its first /32 and announce it at the second site, but a filter configured to accept APNIC IPv6 routes only through /32 might discard the announcement.

That is how a very large allocation can become an awkward travel companion. The addresses exist in the registry. The remote branch cannot count on the world seeing its subprefix. Stevens proposed a second allocation for the disconnected network, so the new site could announce its own full sized route through a local provider.

Stevens framed this as a model network. The draft even showed a prefix-list rule that would accept routes from APNIC's 2400::/12 block through /32 and reject a /35 carved from one of those /32s. The missing fact was prevalence: APNIC's proposal record gave the filter configuration but no count of networks running it. That gap soon became the meeting's most interesting participant.

Stevens was well placed to bring a small-operator concern to the room. In his 2009 APNIC Policy SIG nominee statement, he said he had worked with Australian Internet providers since 1994 and built a business supporting small and medium ISP networks. The proposal reads like a practitioner looking at the forms a client would have to fill out after the physical network had already been built. A transit contract, a router and a usable prefix still leave the second country's traffic at the mercy of the routes other networks accept.

The old subsequent-allocation rule helps explain why this was a policy proposal at all. APNIC's IPv6 policy text from that era evaluated additional space primarily through utilization: how many /56 assignments an LIR had made, measured against an HD-ratio threshold of 0.94. That made sense for an operator whose first allocation was filling with customer assignments. A second isolated network could have a different problem. The first /32 might remain lightly used on paper even as the operator struggled to announce a useful slice of it from another geography. The utilization test would read abundant unused capacity where the engineer saw a blocked route.

The distinction between an allocation and an announcement is the hinge. APNIC can register a range to an account holder. The holder then chooses how to announce routes to its upstreams, and those upstreams and their peers apply routing policy of their own. A registry record tells the world who holds the numbers; it cannot make each transit network carry every more-specific prefix. Stevens's proposal tried to solve the applicant's dilemma on the side APNIC could administer. It offered an additional allocation that might be announced as a whole /32, avoiding a particular class of filters aimed at routes longer than /32.

That solution has an obvious limit. A second allocation may be no help if a network filters for another reason, and the proposal supplied no measurement of how often a separate /32 would turn a failed route into a successful one. Stevens had a mechanism and operating experience. The committee had to decide whether the mechanism was sufficiently credible to justify a new path through allocation policy. In the first round, it was not.

The first room was unconvinced

At APNIC 29 on 4 March 2010, Stevens presented a revised version of the first draft. APNIC had just debated removing an aggregation condition from its initial IPv6 allocation policy. He argued that policy freedom alone would not make a smaller route pass filters elsewhere. The new proposal asked APNIC to issue subsequent IPv6 space to an existing holder with an additional disconnected network.

The room's objections did not form one neat camp. Lorenzo Colitti asked why issuing a second /32 was better than announcing a piece of the first one: two visible routes could consume the same routing-table entries while a second block spent more address space. Raphael Ho imagined a global operator with ten offices asking for ten /32s and proposed work on routing practice. Philip Smith liked the operational objective but disliked making address policy solve a problem created by other operators' filter settings.

Gaurab Raj Upadhaya argued that a separate network could use a separate APNIC membership and questioned whether the policy would make filters accept the result anyway.

Stevens pushed back on time. Networks had to be deployed while router configurations changed slowly. He said his clients had encountered old bogon lists that stayed in use after the address blocks they covered had become legitimate allocations. For an operator trying to turn up a distant site, a future document recommending better filters was an unsatisfying substitute for a route accepted today. He also said APNIC's hostmasters could apply judgment to requests instead of treating the policy as a vending machine for /32s.

Another policy in the same meeting made the argument more subtle. Prop-082 removed an aggregation requirement from initial IPv6 allocations. The community accepted the principle that a registry allocation rule should leave more room for an operator's routing design. Stevens told the room that permission to announce a smaller route would not make a distant network accept it. One proposal removed a written policy constraint; the other asked what to do when an unwritten operational habit remained. That is why the two could appear compatible to Stevens and contradictory to critics. They were working on different layers, but a fix at one layer could reward a problem at the other.

The concrete examples made that disagreement visible. Colitti's arithmetic counted route-table entries and address space: if two locations each announce a route, why consume two /32s when one /32 can be divided? Ho counted offices: a rule designed around one awkward new site could become a rule for ten. Stevens counted engineering hours and the time spent persuading someone else to retire a stale filter. These are all real costs, but the 2010 transcript supplies no common unit for comparing them. A ballot in a policy room cannot turn them into one spreadsheet column.

Smith's intervention cut through the easy version of both positions. He said the behavior of operators applying rigid /32 filters was regrettable, yet he was not objecting to the proposal merely to defend those filters. That matters because the room was not choosing between people who understood routing and people who did not. The dispute was over which institution should carry the operational mess while routing norms caught up. Stevens's proposal would let APNIC absorb some of it in new allocations. A best-practice document would ask network operators to absorb it in filter changes.

A second membership would make the applicant absorb it in administrative fees.

Stevens also exposed a price hidden in that last suggestion. If the same company creates a second APNIC account for a distinct network, it may get another /32 under initial-allocation rules. The addresses then exist either way. Requiring a second account changes who pays and how the request is described. That is the sort of difference a form can make while pretending to be about network geometry alone. The 2010 meeting did not settle the price question; it showed that "use the existing process" is itself a choice with costs.

That was a clash between two kinds of waste. The objectors saw the risk of spending address space and teaching every operator to request another block when a filter was badly set. Stevens saw idle deployment, administrative duplication and engineers spending their days persuading strangers to change routing filters. The registry could write a rule for its own decisions; neither side could order the entire Internet to accept a /35. A few participants opposed the proposal, and the chair recorded no consensus. The APNIC proposal register returned it to the author and mailing list.

The rewrite made the problem larger than geography

When Stevens returned at APNIC 31 in February 2011, the geographic story remained, but the proposed rule had changed. His written version 3 used an Australian network and a proposed Cambodian branch as its example. It also added transition technology such as 6RD and other reasons APNIC might accept. The operator had to hold and announce an existing IPv6 allocation and present a compelling reason for another. An unconnected second network became an example of the need, rather than the single doorway into the rule.

The chronology matters. Stevens spoke on 23 February about language he had revised during the meeting; the final written version is dated 24 February. The chair told the room the text would still have to go to the mailing list after any meeting consensus. APNIC did not turn a lunchtime revision into an instant binding rule. It exposed the new approach to the subsequent comment period.

The debate became more pointed. Stevens said roughly two dozen cases had been referred to him during the intervening year, often involving separate networks. The public transcript gives that as his testimony; it contains no verified list of those cases or measurement of their reachability. Randy Bush replied that he was announcing a /48 in three places without a reachability problem and wanted a serious measurement study. Philip Smith said he had announced a /35 across many locations and asked which providers had blocked Stevens's reported cases. Stevens did not have provider names to give the meeting.

The counterexamples exposed the evidentiary weakness of a universal filtering claim without proving every small route was reachable.

Read the exchange as a scene about evidence, not as a referendum on Stevens's honesty. Stevens described referrals from operators who needed a practical answer. Bush described his own successful announcements. Smith supplied another successful /35 and asked for the identity of networks that blocked the routes in Stevens's cases. Each speaker had a different sample, collected for a different purpose. APNIC's published transcript gives the debate and its disagreements; it does not give a regional routing measurement.

That missing measurement mattered because the proposed remedy was durable policy, while a badly configured filter might be transient—or stubborn enough to outlive several policy cycles.

The 2010 and 2011 texts also moved the case beyond route filtering. In version 3, 6RD appeared as another example of why an operator might require a separate IPv6 range. Stevens told the room the proposal should leave room for transition methods and new network structures that the old utilization test did not anticipate. That widened the policy's possible beneficiaries and reduced its dependence on proving every detail of the /32-filter story. A narrow fix for disconnected geography would have required the community to agree about that geography.

A broader alternative criterion let APNIC review several technical stories against one standard.

The breadth introduced its own risk. A catch-all reason can become an invitation to ask for more space with a rehearsed explanation. Stevens acknowledged at the meeting that the additional allocations could accelerate consumption; he considered the effect small against the available IPv6 pool. Critics cared less about running out of IPv6 bits tomorrow than about making a habit of issuing another block to compensate for avoidable routing behavior. The two sides were guarding different resources: Stevens guarded deployment time; opponents guarded allocation discipline and operational incentives.

Neither resource is infinite just because IPv6 has a long address.

Other participants offered a different route through the registry itself. Izumi Okutani described JPNIC's approach: two networks with separate global ASNs inside one organization could become two accounts and obtain two initial /32s. The meeting quickly reached the cost of that alternative. Two accounts meant something like two memberships. Stevens accepted that additional space would carry a price but questioned the point of making the applicant create a second account merely to achieve the same allocation outcome. The invoice was another way policy could decide network design.

Then Bush located the real edit. The earlier draft had tried to specify the qualifying network shape. The new one was leaving more of the test to APNIC hostmasters. Was that the change? Stevens said yes. The Secretariat said a passed proposal would need an operational guideline or checklist for evaluating requests. The meeting therefore was deciding how much discretion to hand the registry and how visible that discretion ought to be to applicants. "Compelling reason" sounds plain until a hostmaster must decide whether the reason actually compels.

The burden in that phrase has two sides. The applicant must supply a network plan worth believing. APNIC must turn a private assessment into a repeatable administrative practice. At the meeting the Secretariat said it could implement the proposal within months, but would need guidance on how to judge an additional request. A checklist could ask whether the applicant already announces its allocation, how the second network is connected, and what technical constraint the new range solves. It could also reveal whether a request is merely a desire for a cleaner spreadsheet.

The public proposal creates a standard; the hostmaster's work gives that standard its real edge.

This is where Stevens's position as an applicant advocate mattered. A person who helps small ISPs complete resource requests can see the point at which an apparently neutral rule demands the wrong evidence. A utilization threshold asks how much of the first block has been assigned. The disconnected-network case asks whether the existing block can be used in the second place under the available routing arrangement. The applicant must tell a different story to satisfy each test. Moving to a "compelling reason" path gave that second story somewhere to go, provided APNIC would listen and document its judgment.

Okutani's JPNIC example sharpened the governance choice. The same physical arrangement—two networks, two autonomous systems, one ultimate organization—could be represented as two accounts with separate initial allocations or one account asking for a subsequent allocation. The network packets would not learn which administrative route the operator had chosen. The operator's costs and application duties would. Stevens argued that forcing the first representation simply to obtain a second usable block added paperwork without improving the engineering. Others saw the account boundary as a useful discipline.

The argument was about where to put responsibility, not merely where to put a slash in a prefix.

The chair called for support and opposition to the revised approach. With no strong objection, the Policy SIG found consensus on 23 February. The register records the final eight-week comment period from 1 March to 26 April 2011, APNIC Executive Council endorsement on 6 May and implementation on 9 August. APNIC's implementation notice confirms that prop-083 entered its active policy documents that August. Stevens had not persuaded everyone that filters were the central cause. He had persuaded the process that APNIC needed a route to consider requests whose need did not fit the old utilization test.

The intervening steps earned their place in the story. The APNIC 31 Members Meeting transcript records the membership reaffirming the Policy SIG's consensus. The end-of-comment notice says the final eight-week period produced no objections, so the chairs considered the consensus maintained. The Executive Council then endorsed the decision. These were different bodies and different opportunities to object; counting the Policy SIG show of hands alone would miss the rest of the mandate. It was a public process with a sequence, and Stevens's revised text survived it.

The passage also illustrates the limits of consensus as evidence. The meeting's support proves the community accepted a policy route. It does not quantify how many /35s failed in the wild, how many extra /32s APNIC later issued under this criterion, or whether route filters changed because of the debate. Those questions need other records. The vote authorized the Secretariat to judge applications under the new rule. It did not certify Stevens's operational theory as a measured law of the Internet.

Who got to decide what a network needed

There is a precise institutional move inside this story. The 2010 draft named disconnected networks as the condition. That made the rule relatively easy to state but rigid at the edge: a transition deployment or another unusual topology might have the same operational problem and still miss the test. The 2011 text opened the door to several cases and asked APNIC to decide whether the applicant's reason was compelling. That gave a hostmaster room to match the rule to actual deployments. It also made the hostmaster's judgment more consequential.

Stevens's achievement was therefore neither simply "more IPv6" nor the defeat of a bad router filter. The policy was an interface between two systems controlled by different people. APNIC controlled allocations and the evidence it would ask for. Transit networks controlled route acceptance. Applicants controlled their network design only within the limits those two systems imposed. Prop-083 gave the applicant another registry path when the first allocation was operationally awkward. It could not reach into another operator's router and remove a filter.

The revision also avoided pretending the room had proved a global rule about /35 reachability. The transcript contains opposing experiences, a request for measurements and Stevens's own examples. The adopted solution required a reason on each application rather than a theorem that every network filtered the same way. That trade made the proposal politically possible. It placed the disputed evidence where the Secretariat would have to examine it case by case.

The current APNIC number resource policy shows the structure that survived. Its subsequent-allocation section retains the HD-ratio utilization route and separately permits another allocation when an ISP or LIR demonstrates a valid reason. It directs readers to APNIC's request guidelines. In those current guidelines, special circumstances include a discrete network while the existing allocation is being announced, technical transition, a need larger than remaining space, and another reason APNIC accepts. The 2026 documents have been revised since 2011, so their wording is today's rule; the recognizable architecture is the coexistence of a numerical threshold and a reasoned exception.

That coexistence is important. The ordinary utilization path treats the first allocation as a container being filled. The alternative path treats the allocation as one piece of a network design. Each is a reasonable question for a different case. Confusing them would either let every unused address block block a legitimate new topology, or let every assertion of inconvenience bypass evidence about actual use. The policy's stability depends on keeping both questions available and letting the hostmaster say which one the application answers.

APNIC's current text on the size of subsequent allocations also recognizes a special case for separate disaggregated ranges requested for multiple discrete networks. That is a trace of the issue Stevens carried into the room. It does not tell us which applicants have used it, or how often; the document is a rule, not an award ledger. What it does show is that the registry still has to describe the shape of multiple networks inside a single allocation system. The address pool was never the only geometry in the debate.

The APNIC 29 room had asked whether registry policy should spend addresses to route around operators' old habits. The APNIC 31 room answered with controlled discretion: sometimes an extra allocation could be justified, and APNIC would have to decide when. Stevens wrote the request and stood through the argument. The routers kept their votes. The registry acquired a new one.