Summary

  • AFRINIC's minute for the policy meeting on 27 November 2009 records that IPv4 Soft Landing did not reach consensus and was returned to the Resource Policy Discussion mailing list. This was the proposal's second public-meeting failure, but the text carried towards Dakar was not the same text that had met the earlier result.
  • The proper comparison is the archived 13 May text against the archived 28 September text carried into the November meeting. The September revision changed definitions, qualification language, the allocation ceiling, the utilization test, and the size and release mechanism of the reserve, while retaining the /23 limit, IPv6 condition, eight-month period, and full regional-use rule.
  • The objections visible around Dakar concerned who would exercise judgment over a larger reserve, how allocation size should be evaluated, and whether a registry could condition an entire allocation on geographic use. Those are operational questions, but they are also questions about the scope of private coordination under scarcity.
  • Repeated no-consensus after substantive revision is therefore consistent with a durable constitutional disagreement over registry scope. It does not establish a vote, veto, obstruction, capture, bad faith, illegality, institutional collapse, or a single cause. Ordinary good-faith refinement remains a strong alternative explanation, and the surviving decision record is too incomplete to rank the explanations conclusively.

The second return is the event

The decisive archival act is concise. AFRINIC's public minute dates the policy meeting to 27 November 2009 in Dakar, Senegal, and says that the IPv4 Soft Landing proposal did not reach consensus. The proposal went back to the Resource Policy Discussion mailing list for further discussion. That return, rather than adoption, is the event that needs explaining. It was the second time a public meeting had ended that way for this proposal.

The word “second” matters only if it is handled with discipline. It establishes repetition, not the reason for repetition. The earlier result is a boundary comparator: the proposal had previously been returned without consensus. Reconstructing that earlier meeting, or narrating the proposal's full origin and later life, would obscure the narrower question raised by Dakar. By November, participants were not simply looking again at a static document. A new archived version dated 28 September stood between the two meetings. The analysis must therefore ask what changed, what did not, and what kinds of disagreement remained after the changes.

That approach prevents two opposite errors. One would treat repeated failure as self-explanatory evidence of obstruction. Nothing in the surviving materials warrants that charge. The other would treat the event as mere procedural delay without examining the substance that survived revision. That is equally unsatisfactory. The September text altered the distribution of eligibility, the evidence required from applicants, and the institutional location of discretion. Meeting-day messages then pressed concrete questions about how scarce address space could be used.

A second return after those revisions carries more analytical information than the bare phrase “no consensus,” even though it never reveals a decisive cause by itself.

The official chronology supports only a modest frame. The proposal archive records a 13 May update, an earlier failed-consensus event, a further update on 28 September, and the second failed-consensus event on 27 November. The older meeting minute is useful only to confirm the earlier result and return. Dakar owns the substantive inquiry here: the September text, the objections near the meeting, the later issue summary, and the continuing gaps in the decision record.

The comparator must be May against September

The operative textual comparison is explicit: the archived 13 May proposal is the baseline, and the archived 28 September proposal is the version carried into the November meeting. This is more precise than saying vaguely that “the proposal evolved,” and more honest than claiming certainty about the paper, screen, or file used in the room. The most recent archived text before Dakar is the September version. The surviving materials do not establish the exact copy displayed, circulated, or tested during the session.

The distinction is important because a policy discussion can become detached from its controlling wording. Participants may remember a prior clause, discuss an amendment orally, or respond to consequences rather than line edits. A post-meeting summary may compress several stages into one description. The archive allows a reliable comparison of two dated texts, but it does not allow an observer to reconstruct every document transition in the Dakar room.

The right claim is consequently limited: the September text was the latest archived version carried towards the November meeting; it is not proven to be the exact visual or working copy before every participant.

Within that limit, the changes were substantive. They affected who counted as a new or existing local Internet registry, whether an applicant merely requested an allocation or had to qualify, how the four-allocation ceiling was counted, which earlier allocations entered the utilization test, and how much address space would be reserved under whose release judgment. These were not cosmetic edits. Each changed either an applicant's evidentiary burden or the registry's room to decide.

Several important provisions did not move. Both texts retained a /23 as the minimum and maximum allocation during the exhaustion phase. Both linked the IPv4 allocation to an accompanying IPv6 allocation, used an eight-month allocation period, and directed AFRINIC-issued IPv4 space to be used entirely for operations within the AFRINIC service region. Stability matters as much as revision. It identifies which constraints persisted into Dakar and therefore which questions revision had not eliminated.

This mixed pattern—meaningful change around stable controls—is why the second result cannot responsibly be labelled either rejection of an unchanged draft or proof that revision was futile. Some ambiguities were narrowed. Some discretion was made more explicit. One reserve grew and shifted from an automatic release rule to a judgment exercised by the Board. A stringent geographic rule remained. Revision clarified the proposal's governing choices; it did not necessarily make those choices less contestable.

Definitions moved the scarcity boundary

The May text defined an existing local Internet registry as one that had already received IPv4 space from AFRINIC. A new local Internet registry was a recent member that had not yet received an allocation. The September version added the function of assigning address space to end users and tied new-registry status more specifically to having received no allocation before the exhaustion phase.

At first glance, definitions can look like housekeeping. Under scarcity, however, a definition assigns applicants to categories with different prospects. The relevant cut-off determines who arrives at exhaustion carrying an allocation history and who is treated as new. Adding the end-user assignment function also clarifies the operational role being classified. These changes do not create sovereignty or ownership; they sharpen a service rule. Yet that rule can affect which operator receives scarce address capacity and under what conditions.

Classification therefore belongs in the constitutional analysis, not because it is public law, but because private bookkeeping has distributive consequences.

The September definition appears responsive to ambiguity. It specifies the relationship between membership, previous allocation, and the onset of exhaustion. That is evidence for the benign view of the revision: policy authors were doing ordinary work to make an allocation rule more implementable. But implementability does not answer authority. A clearer category can still embody a contestable choice about whose prior position counts and how the registry's ledger history shapes future access.

No surviving material shows that definitions themselves were dispositive at Dakar. They should not be inflated into a hidden cause. Their analytical value is narrower. They show that the text had undergone meaningful refinement and that the proposal's object was not simply conservation in the abstract. It translated scarcity into classifications that staff would later have to apply to real applicants.

“Requests” became “qualifies”

In the May formulation, the exhaustion-phase /23 minimum and maximum would be allocated to a local Internet registry that requested it. In September, the operative verb became “qualifies.” One word changed the visible posture of the rule. A request describes an applicant's act. Qualification implies evaluation against conditions.

That distinction matters because it identifies an administrative decision point. If the proposal advanced and was implemented, AFRINIC staff would have to determine whether an applicant met the applicable requirements. The task would remain private registry administration: checking records and applying a service policy. It would not make staff a regulator, court, prosecutor, or owner of the applicant's network. Still, a qualification judgment could control access to scarce number resources, so the criteria, evidence, consistency, and route for challenging an error would matter to operators.

The change can be read as a drafting improvement. A scarcity policy cannot plausibly mean that every request automatically succeeds regardless of conditions. Making qualification explicit may simply align the verb with the rest of the rule. But it also exposes the practical power that a softer verb concealed. Someone has to decide whether the conditions are satisfied. When the pool is nearly exhausted and the permitted allocation is small, that service judgment can affect expansion, routing plans, customer onboarding, and capital already committed to infrastructure.

The 4 December outcome later described an issue involving a /24 and /23 minimum or maximum based on request evaluation. The summary is terse. It does not identify speakers, quantify support, or preserve final wording. It nevertheless confirms that allocation size and evaluation remained part of the unresolved issue set. The archive supports the existence of that issue, not a conclusion about which formulation a majority or any other defined constituency preferred.

The four-allocation ceiling was tightened in meaning

The May text said that no local Internet registry could receive more than four additional allocations during exhaustion. September removed “additional” and said no more than four allocations during that phase. The practical significance is a clearer count. Under the May wording, a reader could ask whether the first exhaustion allocation sat outside the four additional allocations. The September wording places all exhaustion-phase allocations inside the ceiling.

This change narrowed an ambiguity rather than enlarging discretion. It is among the strongest pieces of evidence that the revision process was responsive. A fixed ceiling is easier to predict than a ceiling whose starting point is unclear. Applicants can model a maximum number of allocations; staff can apply a simpler count; other participants can assess how quickly repeated applications might draw down the remaining pool.

Even so, clarity does not remove distributional consequence. A uniform limit treats operators of different scales and growth patterns alike on one dimension. That can spread remaining access more broadly, but it can also constrain a network whose demonstrated demand is larger. The proposal paired the ceiling with a small /23 allocation size and a utilization threshold, creating a rationing architecture rather than a mere drafting convention. Whether that architecture is wise cannot be inferred from the no-consensus result. What can be said is that the architecture remained contested enough for the proposal to return to the list.

The record does not disclose whether the count change was accepted without controversy, whether it was discussed at all in Dakar, or whether it influenced the final assessment. It belongs in the delta because it proves substantive revision. It must not be turned into an imagined room debate.

The utilization test reached further back

For an existing local Internet registry seeking another allocation, the May text referred to ninety-percent use of previous allocations received during the exhaustion phase. The September version required ninety-percent use of all previous allocations. The change broadened the evidence base. An applicant could no longer satisfy the wording by showing intensive use only of exhaustion-era space while leaving earlier allocations outside the stated test.

From a conservation perspective, the revision has a coherent benign rationale. If the objective is to preserve the final supply while serving demonstrated need, examining all earlier allocations is more complete than examining only the newest ones. It could deter an applicant from treating older underused space as irrelevant while seeking another scarce block.

From an operator perspective, the broader test also increases the registry's evaluative surface. Historical allocation records can reflect different technologies, customer commitments, topology, fragmentation, or deployment timing. The archive does not supply an implementation-impact study explaining how those situations would be assessed. Nor does it show the evidentiary materials available to the Dakar meeting. The constitutional point is not that utilization review is illegitimate.

It is that a private ledger service that conditions scarce access on a broader historical judgment owes criteria and evidence proportionate to the operational consequence.

Again, the archive does not identify this revision as the reason consensus failed. Its role is to rebut the idea that nothing serious changed between meetings. The September proposal strengthened a condition that staff would apply to real address histories. That is a material change even if no surviving objection singles it out.

The reserve became larger and discretionary

The most institutionally significant delta concerned the reserve. In May, the proposal set aside a /16 and provided for its automatic return to the general pool after the rest of the final /8 had been allocated. In September, the reserve became a /12. The Board could replenish the exhaustion pool from it when the remaining pool was empty or no longer contained enough contiguous space, taking account of demand and other factors described in terms of the community's best interest.

Two things moved at once. The amount withheld increased, and release shifted from a mechanical trigger to a discretionary judgment. A larger reserve could offer more room to respond to future scarcity. Board judgment could also adapt to conditions that a fixed rule failed to anticipate. Those are serious arguments for flexibility. A rigid release rule may perform badly when demand, fragmentation, or operational needs depart from the assumptions embedded in the text.

Flexibility, however, requires an authority and accountability story. Who is included in “the community”? What evidence would the Board examine? How would applicants know that comparable cases received comparable treatment? What reasons would be recorded? What review or correction would exist if a release decision imposed avoidable costs? The text's reference to demand and other factors did not answer those questions fully.

A 30 September mailing-list objection made the problem concrete. The participant questioned whether the Board's conception of the community's best interest would be shared and compared Board-triggered release with automatic release and an indefinite reserve. That message is evidence of one person's position, not a vote, veto, or collective verdict. Its importance lies in the issue it identifies: the September revision relocated a consequential scarcity decision inside a private corporate board.

Under the controlling authority boundary, AFRINIC is a bookkeeper, ledger service, and coordinator. The Board may have corporate responsibilities and may be assigned functions by a private policy. Those facts do not turn it into a sovereign allocator of public rights. The reserve clause would have given the Board practical control over when withheld address capacity became available to applicants. Practical influence is real, but it remains bounded service power. The stronger the operational consequence, the stronger the need for criteria, reasons, evidence, and a path to challenge error.

This is where the second failure becomes suggestive. Revision did not merely polish the earlier reserve. It made a larger pool subject to a more explicit institutional judgment, and an objection promptly questioned the identity between the Board's interest assessment and the interests of those affected. The record cannot prove that this concern blocked consensus. It can show that a dispute about discretion survived in visible form.

Geography converted conservation into a deployment condition

Both the May and September texts retained a stringent geographic-use rule: AFRINIC-issued IPv4 space was to be used entirely for operations within the AFRINIC service region. Around the Dakar meeting, the mailing list preserved a direct clash over that condition.

On 27 November, Mark Elkins proposed an allowance under which up to ten percent of an allocation could be used outside the region when it connected back to the region. This was one participant's proposed amendment, not a measured preference of the room or list. The same day, Andrew Alston objected more broadly. His position was that the restriction could burden African companies expanding outside the region, encourage route deaggregation, and force operators to maintain relationships with multiple regional Internet registries. That, too, was one participant's argument, not a community position.

After the meeting, the proposal author defended the restriction as protection against organizations establishing an African presence merely to obtain scarce regional space for use elsewhere. The rationale framed the policy as preservation of the region's last resources. It deserves to be stated fairly. A finite pool creates a risk that eligibility tied to a regional presence could be used to draw scarce capacity away from the operators and end users whose needs motivated the reserve. The archive records this as the author's rationale.

It does not adjudicate the empirical scale of the risk or establish that the restriction was the only way to address it.

The operational objection is also substantial. An African operator can expand abroad without ceasing to be African. A network's topology may cross regional boundaries for customers, resilience, interconnection, or service continuity. Requiring every unit of an allocation to remain within a registry region may make administrative geography override network engineering. If the workaround is to split announcements or seek relationships with multiple registries, conservation at the ledger level can create complexity in routing and operations.

The constitutional question is therefore sharper than “regional interest versus external use.” It is whether regional bookkeeping should remain a thin eligibility and recording service, or whether the ledger service may condition every address in an allocation on where an operator deploys it. AFRINIC did not own the operator's routers, customers, or network. A policy condition could exert powerful leverage because access to address space was scarce, but leverage is not sovereignty. It does not give the registry police, confiscatory, punitive, or adjudicatory authority.

The 4 December outcome recorded the geography issue as a proposed movement from full regional use to a ten-percent out-of-region allowance tied to connectivity back to the region. That confirms the issue survived into the result summary. It does not reveal whether the full restriction, the exception, enforcement practicality, or some other detail was dispositive. Nor does it establish that attendees agreed on the premise that a private registry should monitor deployment geography at all.

The geographic debate supplies the clearest example of revision exposing a scope dispute rather than settling one. The stringent clause did not change between May and September, but its consequences became more visible as participants tested it against cross-border operations. Policy development can be working properly when operational objections emerge. At the same time, the substance of those objections concerns the reach of coordination into network decisions. Both readings can be true.

The outcome names issues but not decision mechanics

The public minute establishes no consensus and return to the list. The PDP-MG's 4 December message adds three clusters: reserve size, a minimum-or-maximum allocation formulation involving /24 and /23 according to request evaluation, and the possible ten-percent exception to full regional use. Taken together with the dated texts and mailing-list messages, this makes the AFRINIC-11 event more reconstructible than the earlier sparse result.

More reconstructible does not mean complete. Neither record supplies a support count, opposition count, abstention count, eligible denominator, or separate denominator for people in the room, remote participants, and mailing-list contributors. There is no surviving transcript or full speaker list for this item. The materials do not state the exact consensus-assessment method, threshold, or reasoning. They do not show whether each issue was tested independently or the proposal was assessed only as a whole. No objection is identified as dispositive.

These omissions matter because “no consensus” is a conclusion about a coordination process, not an explanation of the underlying distribution of views. Without a denominator, it is impossible to translate list messages or room participation into a quantified constituency. Without reasons, it is impossible to know whether assessors considered one objection fatal, judged the issue mix collectively unresolved, or observed some other pattern. Without a transcript, silence cannot be interpreted as assent, opposition, absence, or uncertainty.

Named participants must remain named participants. Their messages are valuable because they state operational concerns and proposed wording. They are not ballots. They do not prove that a majority, minority, member body, operator class, or general public held the same view. The proposal author's defense is equally bounded: it establishes a rationale offered by the author, not the institutional truth of the policy or the intent of everyone who supported it.

Open participation strengthens the evidence base when it brings operational knowledge and criticism into view. It does not manufacture a demos. Meeting attendance, list activity, expertise, or a chair's assessment cannot authorize absent operators and end users unless an express authority chain does so. Consensus is an engineering coordination practice, useful for discovering workable terms and detecting unresolved risk. It is not a grant of sovereign jurisdiction.

That distinction works in both directions. A declaration of consensus would not have made AFRINIC a government, and a declaration of no consensus did not constitute a court judgment against the proposal. The event says that the private policy effort did not advance at that meeting and returned to discussion. Its meaning must come from the issues that survive around it, subject to the limits of the record.

The /16-to-/12 anomaly must remain unresolved

One version-integrity anomaly prevents a perfectly smooth narrative. The 4 December outcome described a change from a /16 reserve to a /12 as one of the issues. Yet the archived 28 September proposal already stated a /12. If the September text was the latest archived version before Dakar, why did the post-meeting outcome describe the reserve change in that form?

Several possibilities can be imagined. The room might have considered an older copy; the outcome might have summarized the delta from May; an amendment sequence might have been compressed; or some other document-alignment circumstance might explain the wording. None is established. Choosing among them would turn a gap into a claim.

The correct treatment is to preserve the mismatch. It weakens confidence in any sentence asserting exactly what copy was tested in the room. It does not prove that the wrong document was used, that participants were misled, or that the result was defective. It also does not erase the dated September archive. The archive still establishes that a /12 and Board-release discretion appeared before the meeting.

Version integrity is not a clerical footnote when policy governs scarce operational inputs. Participants need to know which text they are evaluating; later readers need to trace the decision against that text. A mismatch between a dated proposal and an outcome summary creates uncertainty about document alignment. The responsible conclusion is modest: the decision record is incomplete at a point that could matter, and no speculation can repair it.

The strongest benign case is genuine policy refinement

The strongest contrary explanation deserves more than a passing concession. A second no-consensus result can be ordinary, good-faith policy development. The September text shows authors responding to ambiguity. It clarified category definitions, changed “requests” to “qualifies,” removed uncertainty in the four-allocation count, broadened the utilization test, and designed a larger reserve for difficult conditions. The return to the list may have protected implementation quality and minority concerns rather than exposed any constitutional fracture.

On this reading, the process acted as a safety valve. Scarcity rules are technically consequential. It can be prudent to delay adoption when questions remain about address-block size, reserve release, applicant evaluation, and geographic deployment. Visible objections and an author's response demonstrate engagement rather than bad faith. A policy process that refuses premature agreement can be healthier than one that converts momentum into adoption.

Other benign variables are possible but unproven. Meeting time, travel, language, remote connectivity, participation, and facilitation can affect an in-person event and the relationship between a room and a mailing list. The surviving sources do not quantify or establish any of those influences for this decision. They belong among unknowns, not findings. The issue mix may also have shifted as revision exposed new consequences; the archive cannot show when every participant formed a view.

Document alignment offers another benign possibility. The reserve mismatch in the outcome may reflect nothing more than a summary written against an earlier baseline. It could have complicated discussion, but the record does not establish that it did. The anomaly therefore cautions against certainty without supporting any accusation.

The benign case constrains the thesis in an important way. Repetition alone is weak evidence. Responsive drafting plus return to discussion can be a sign that open coordination is doing its job. The analysis becomes stronger only when it asks what survived revision. Here, the visible issues were not confined to grammar or an isolated prefix length. They involved where discretion sat, how rationing would be evaluated, and how far a regional-use condition could reach into an operator's deployment. Those are questions of institutional scope as well as implementation.

The two explanations can coexist. A proposal may remain technically immature because it has not resolved a constitutional question. A safety valve may work precisely because participants lack agreement about the legitimate reach of the coordinator. The archive cannot rank drafting difficulty against scope disagreement or allocate shares of causation. It supports a limited inference of consistency, not a definitive diagnosis.

Scarcity turned ledger entries into capital allocation

IPv4 exhaustion gave administrative wording unusual economic weight. A registry's ledger records allocations and coordinates uniqueness. Under abundant supply, many such decisions may feel routine. Near exhaustion, a definition, utilization test, maximum size, or reserve release can decide whether an operator can obtain capacity on expected terms. Bookkeeping then has the practical effect of capital allocation.

That effect does not transform AFRINIC into a sovereign. The registry remains a private service and coordinator. It does not own members' networks, rule a territory, police deployment, prosecute conduct, confiscate infrastructure, punish operators, or adjudicate public rights. Its influence arises from control of a scarce coordination point and from the operational costs of disagreement or exit. Calling that influence “regulatory” in a loose descriptive sense can conceal the absence of public-law authority.

The proper response to practical power is a bounded authority chain. Criteria should be tied to the coordination service. Evidence should be relevant to allocation integrity and operational need. Reasons should be reviewable. Decision makers should not convert a general phrase such as “community's best interest” into permission to substitute their preferences for those of absent principals. Appeal and exit matter because an operator can bear real costs even when the decision maker is private.

This framework clarifies why the Dakar issues have constitutional content without becoming a claim of sovereign conflict. Reserve release asked who could decide when withheld capacity returned to applicants. Qualification and utilization asked how deeply the ledger service could examine and rank demand. Geographic use asked whether access to addresses could be leveraged into control over deployment location. Each issue concerned the perimeter of private coordination.

The term “constitutional” here describes allocation of authority inside an infrastructure arrangement. It does not suggest a state constitution, public regulator, or judicial dispute. It asks the prior question: what has the bookkeeper been authorized to do, by whom, for whose benefit, under what evidence, and with what remedy when the answer is contested?

The power chain ends with those who bear the cost

The upstream scarcity context began with the global allocation mechanism and the final-/8 trigger. That mechanism supplied a condition to which regional policy makers had to respond. It did not confer sovereignty on AFRINIC.

The proposal author drafted and revised the rule, then defended the regional-use rationale. Authorship is essential to policy development, but proposing words does not bind operators. Mailing-list participants contributed objections, alternatives, and operational evidence. Their participation enriched the record; it did not authorize them to act for people who were absent.

The PDP-MG and meeting assessors managed discussion and recorded no consensus and return to the list. That process determination had practical effect within private policy development. It was neither regulation nor adjudication. Because the surviving record omits the assessment method and reasons, outsiders cannot reconstruct how the conclusion was reached beyond the stated result.

Had a rule advanced and been implemented, AFRINIC staff would have interpreted qualification, utilization, and allocation conditions. Staff judgment would have been administrative performance of a ledger service. It would not have supplied punitive or police power. The September proposal also assigned the Board possible discretion over reserve release. A corporate board can exercise functions inside a private institution, but it cannot manufacture public sovereignty or ownership of operator infrastructure.

At the end of the chain sit applicants, local Internet registries, network operators, and end users. They would bear the consequences: scarcity, delayed expansion, routing fragmentation, multiple registry relationships, compliance work, mobility constraints, and continuity risk. These actors are not an audience below the process. They are principals in the infrastructure. Any legitimate coordination practice should remain answerable to their operational reality.

That final link is often lost when “community” is treated as a self-defining source of authority. A meeting can contain knowledgeable people and still exclude many who bear costs. A mailing list can be open and still have no meaningful denominator. Expertise and participation deserve weight as evidence, but neither supplies an unlimited mandate. The authority chain must be demonstrated, not inferred from attendance.

What the second failure can and cannot mean

Four propositions survive close scrutiny. First, the proposal received a second no-consensus result after substantive revision. Second, the issues visible around that result included institutional discretion, rationing criteria, and geographic control. Third, these issues reach beyond grammar into the scope of private registry coordination. Fourth, the richer surviving record makes the second result more interpretable than the earlier sparse result, even though the basis of decision remains incomplete.

Together, those propositions support a bounded inference: repeated no-consensus can be read as evidence consistent with a durable constitutional disagreement about private-registry scope, rather than slow drafting alone. “Can be read,” “consistent with,” and “rather than alone” do essential work. They preserve the difference between an analytical explanation and an archival fact.

The inference does not establish that every participant understood the conflict in constitutional terms. It does not identify a person, staff member, assessor, chair, or Board member as the cause. It does not prove capture, fraud, corruption, obstruction, bad faith, illegality, punishment, a veto, or institutional collapse. It does not quantify any majority or minority. It does not show that the same dispute persisted unchanged after 2009.

Nor can it answer counterfactuals. The archive does not show whether a thinner coordination-only proposal would have succeeded. It does not reveal which wording changes were accepted before, during, or after the outcome. It does not preserve the implementation-impact analysis available to participants. It does not tell us who “the community” was meant to include or authorize.

The second return matters because revision made the proposal's choices clearer while material scope objections remained. Its meaning lies in the shape of those surviving choices, not in a mythology of failure. The Dakar result is neither proof of dysfunction nor a blank procedural pause. It is evidence of unresolved coordination under scarcity, bounded by a record that lets us see the questions more clearly than it lets us see the decision.