Summary

  • The earlier wording of section 6.5.8.2 sent a one-site case in two directions: its 75% next-nibble rule pointed to /44, while its first example began /44 at two sites and thereby implied /48 for one.
  • The Recommended Draft makes the intended baseline explicit and sets closed bands of 2–12 → /44, 13–192 → /40, 193–3,072 → /36, and 3,073–49,152 → /32. It remains a Recommended Draft Policy, not adopted or implemented policy.
  • The useful reform is not to declare /48 a law of IPv6. It is to publish one versioned decision surface in which operative text, examples, implementation tests and any correction all return the same answer.

One divided by one

The contradiction fits on one line:

one site / one available /48 = 100%.

The earlier sentence said that an organization qualified for the next larger nibble boundary when its sites exceeded 75% of the /48s available in a prefix. Applied to a /48, one site occupies the one /48 counted by that denominator. One hundred percent exceeds 75%. The literal result is the next larger nibble boundary, a /44.

Then the example changed the answer. Its first band said that more than one and no more than 12 justified sites received /44. A single site therefore remained at /48.

The independently hosted February revision notice states the conflict plainly. It also records the first structural repair: change “the initial assignment size” to “larger initial assignment sizes”, so the initial /48 stands before the formula instead of looking like an exception to it.

This is not a dispute about difficult mathematics. Both readings are easy. The defect is that they cannot both govern the same case.

The example had become executable policy

Examples are usually treated as explanation. Here the example carried a normative result that the operative sentence did not reliably produce. A reader who followed the equation obtained /44; a reader who followed the table obtained /48. Actual registry practice had to choose between them.

That choice may have been completely uniform. The Recommended Draft's Advisory Council assessment says the clarification matches actual current practice and introduces no operational change. The independent PPML archive of the June announcement preserves that statement together with the full current and proposed text.

But the statement and its proof are different objects. This source set contains no anonymized case sample, decision test, historical aggregate or implementation note from which an outside reader can reproduce the claim. There is no basis here to say that an applicant received the wrong size, suffered delay or was treated inconsistently. There is equally no basis to transform “no operational change” from an attributed institutional statement into an independently tested fact.

The correct response is not suspicion. It is a parity test.

What the Recommended Draft fixes

The 20 May text does more than append “unless one site”. It rewrites the section as a closed sequence.

First, organizations that meet an initial eligibility criterion are eligible for an initial /48. Larger sizes are then determined by justified site count. The bands are written without gaps at their named boundaries:

  • 2 to 12 sites receive /44;
  • 13 to 192 receive /40;
  • 193 to 3,072 receive /36;
  • 3,073 to 49,152 receive /32.

Second, assignment becomes allocation throughout the section. The Advisory Council says this matches current registry terminology because assignments are no longer issued in this context.

The NOG Alliance proposal overview lists ARIN-2025-7 as a Recommended Draft Policy and identifies 20 May as the last text change. That lifecycle label matters. Recommended means the text has advanced; it does not mean the Board has adopted it, the manual has changed or production systems have implemented it.

The present article therefore analyses a proposed repair. It does not announce a new entitlement.

/48 is a policy answer, not a protocol constant

It would be easy to mistake the neatness of the repaired table for technical necessity. RFC 6177 prevents that mistake. It withdrew the earlier one-size-fits-all recommendation that most end sites receive /48. It says the broad range of end sites requires more nuance, leaves the exact choice to the operational community and limits the IETF's role to architectural and operational guidance.

The RFC also warns against hard-coding a few prefix boundaries as though IPv6 had become classful. CIDR still applies to every bit of the prefix.

None of this makes ARIN's proposed one-site /48 wrong. It makes the nature of the decision clear. A registry community may choose a predictable baseline, aggregation rule and evidence requirement. That choice becomes legitimate through precise published policy, bounded authority and reviewable application—not because the number 48 arrived from the protocol as a commandment.

This distinction protects both sides. ARIN may apply the policy it actually publishes. An operator retains authority over topology, subnetting, service design and growth strategy within the resource decision. The registry does not acquire general design authority merely because it records the prefix.

Small ambiguity is expensive at the small end

IPv4 Global's ARIN 57 analysis calls the edit modest and argues that making the single-site outcome explicit can reduce back-and-forth for smaller end users. That is a commercial analyst's expected effect, not a measured performance result. The mechanism is nevertheless credible.

A large address team can compare policy versions, consult specialists and absorb an extra request cycle. A one-site enterprise, institution or small operator may encounter the rule only when preparing its first direct IPv6 request. If the sentence says /44 and the example says /48, it must either over-document, ask for clarification or trust unwritten practice.

The cost does not come from the four bits alone. It comes from uncertainty about which public object has authority. Is the formula controlling? Is the table controlling? Does staff practice override both? Would a different reviewer choose differently? Can the applicant point to a stable version if the answer changes?

A clean rule eliminates those questions before they become a private conversation.

The policy mirror at its smallest scale

Lu Heng's Policy Mirror argues that a policy manual reveals the institution's idea of itself. A narrow registry protects uniqueness, record accuracy, evidence of control, security assertions and continuity. It crosses the line when administrative practice silently becomes wider authority over how operators may use scarce network resources.

ARIN-2025-7 is not a grand exercise of capital control. It is more useful as a microscopic case. A legitimate registry-sizing rule had two public executions. The example supplied the intended answer while the sentence supplied another. The resulting gap gave actual practice the final word without making that practice independently visible.

Repairing the sentence is therefore not cosmetic. It moves authority back into the public rule. The lesson also constrains the repair: publish the minimum common sizing decision and leave the organization's internal architecture local. A registry needs the facts that select a prefix band; it does not need to design the network merely because the application mentions sites.

A policy–practice parity table

The durable artifact can be smaller than a case file. For each material policy version, publish or preserve a public-safe parity row:

Field Minimum content
Text identity Policy identifier, version date, digest and lifecycle state.
Input class Eligible end user, ordinary-site count, extra-large-site need, correction or reconsideration.
Literal output Prefix returned by the operative sentence without consulting examples.
Example output Prefix returned by the published example or table.
Practice output Prefix class the registry says it applies, expressed as a test result or aggregate rather than an applicant file.
Disposition No difference, text corrected, example corrected, practice changed or case reopened.
Bounded reason Single-site baseline, multi-site threshold, extra-large-site evidence or another published code.
Correction path Authority, date, superseding version and method for challenging a mismatch.

This is my evidence model, not an ARIN schema. It does not ask ARIN to expose applicants' topology or commercial plans. It asks the public policy system to show that its three execution surfaces agree.

Four vectors catch the defect

A long policy discussion can miss what four deterministic tests expose:

  1. one ordinary site returns /48;
  2. two ordinary sites return /44;
  3. twelve ordinary sites return /44;
  4. thirteen ordinary sites return /40.

Run each vector against the proposed text, the example table and the implementation rule. Record the text digest and result. If all three surfaces agree, the boundary is portable. If they disagree, publication should stop until one named object changes.

The vectors are not instructions for designing those networks. They test the registry's classification. Extra-large sites and evidence-based exceptions belong in separate named vectors so that an unusual subnet need does not blur the ordinary site-count rule.

This method would also make future edits cheaper. A terminology change from assignment to allocation could be tested for semantic neutrality. A changed boundary would produce an intentional output diff. An editorial rewrite that unexpectedly changes a prefix would fail before reaching a live request.

The repair deserves completion, not exaggeration

The public record supports a restrained conclusion. ARIN found a real contradiction and advanced text that resolves it. The repair appears technically straightforward. The current evidence does not show applicant harm, misallocation or operational change. It also does not independently demonstrate historical uniformity.

The next useful act is therefore simple: carry the Recommended Draft through its remaining public process with the final text fingerprint and boundary vectors attached. If adopted, publish the effective version and the same tests. If practice truly does not change, the before-and-after outputs will show that. If a difference appears, the policy process will have a bounded object to correct.

One site should no longer have to read a sentence and an example as competing sources of law. The table has done its temporary work. The rule must now say the same thing.

Sources

Evidence limits

Verified: the preserved earlier sentence and example conflict for one site; the recommended text makes the single-site /48 and multi-site bands explicit; the Advisory Council characterized the edit as matching practice and causing no operational change; the independent tracker lists Recommended Draft Policy; RFC 6177 treats end-site size as an operational choice.

Not established: adoption, implementation, a real inconsistent decision, applicant delay or harm, the population of affected requests, the historical distribution of allocated sizes, or independent reproduction of the no operational change claim.