Summary

  • RFC 1797 temporarily designated network 39 for a bounded 1995 experiment: could independently operated networks use subnets of former Class A space without old classful assumptions breaking reachability?
  • RFC 1879 recorded real services, real query traffic and concrete routing failures; its “success” was conditional, not evidence that every path or implementation was ready.
  • Keeping the established F-root address beside the temporary Net 39 interface made the most consequential test reversible: new traffic could prove the path while stale clients retained an exit.

Two addresses, one service, one escape route

The F root name server did not simply move into Net 39. During the experiment it acquired a temporary interface at 39.13.229.241 while keeping 192.5.5.241. RFC 1879 explains why the older address remained: many users would not update their root.cache files, and removing it risked creating a black hole.

That detail changes the meaning of the trial. A clean laboratory cutover could have produced a sharper before-and-after chart, but it would also have mixed two questions. If clients failed, was subnetted Class A space unreachable, or had the experiment merely stranded software carrying an old bootstrap address? Parallel service paths allowed the network to learn about the new address without demanding universal coordination on day one.

The traffic was substantial. A report quoted by RFC 1879 said the server handled roughly 400 million queries during 38 days and that well over half arrived at the temporary address. This is a powerful operational receipt: the new path was not merely configured or advertised; it carried useful application traffic at scale. It is still a bounded receipt. The fragment does not provide a complete experiment-wide census, enumerate every source network or prove that every router and host could reach every Net 39 subnet.

The old address was therefore not evidence of cowardice or a failed migration. It was the mechanism that kept measurement from becoming hostage-taking. The trial could attract production-like load, reveal incompatible paths and still leave a known route for clients outside the experiment's adoption boundary.

A whole class became a temporary laboratory

RFC 1797 began from a problem that classless routing had already exposed. IPv4 addresses were historically interpreted through Class A, B and C boundaries. CIDR replaced that fixed geometry with explicit prefixes, helping slow address exhaustion and routing-table growth. Yet new allocation arithmetic did not instantly remove class assumptions embedded in routing protocols, router software and operator configurations.

The RFC temporarily designated network 39, a former Class A block, for an experiment running from 1 May to 1 December 1995. The space was not offered as a permanent new allocation regime. The document warned that IANA could reassign it after the experiment. That expiry was part of the technical design: participation could generate evidence without converting the trial into an irreversible property claim.

Two allocation cases were proposed. In the first, the high bit after the classful network number was zero, the next 15 bits encoded the low 15 bits of a participant's assigned autonomous-system number, and the final octet remained locally usable. It allowed rapid participation without a separate address-registration step, though the RFC noted a possible incentive to obtain an AS number for address space. In the second case, the high bit was one and IANA would assign variable-length prefixes from the remaining space.

Only case one delegations were actually performed. That boundary matters. RFC 1879 could report on the exercised AS-derived pattern, DNS arrangements and routing behaviour around it. It could not convert an unrun second case into evidence. Nor did either experimental layout become a recommendation for the permanent allocation of former Class A networks.

The test needed services, not ping alone

RFC 1797 called for popular web and FTP resources because a network that carries only synthetic probes may hide the conditions that make compatibility economically important. The result report lists Finger, HTTP, Telnet, FTP, Gopher, Kerberos, line-printer, X and DNS services. The temporary F-root interface was especially useful because DNS root queries provided a steady, meaningful workload.

This was an early expression of a principle that is easy to flatten into a slogan: running code is primary because it exposes dependencies that a diagram cannot. A successful ping can show that one packet returned. A root-server query exercises addressing, routing, a listening service, a request and a useful response. Neither proves ownership, universal reachability or every application's outcome, but the latter closes a longer evidence chain.

DNS itself required coordination. Forward records and 39.IN-ADDR.ARPA reverse delegation had to make experimental addresses intelligible. RFC 1797 also encouraged participants to exchange provider connectivity so portability and aggregation behaviour could be tested. The address format was only one layer; names, route policy, live announcements, forwarding and services each had to agree enough for traffic to arrive.

One intended /24 became an announced /8

The failures were not incidental footnotes. They showed exactly where the old model still controlled reality.

RFC 1879 records a classic configuration in which an operator intended to announce 39.1.28.0/24, but classful behaviour promoted the route to the entire 39/8. RIPv1 presented the same problem: when told about the subnet, it emitted the classful network route. The result was not a malicious hijack. It was a legacy inference made by software that lacked the prefix information needed to preserve the operator's narrower intent.

The distinction among layers is crucial. A routing-registry object could describe the intended policy. A router configuration could contain a local network statement. A live update could nevertheless announce a different prefix. A neighboring router could accept that update, and a packet could then follow it. None of these records should be substituted for another.

RFC 1879 names historical configuration controls that prevented automatic summarization and classful forwarding assumptions. Those examples belong to their period; they are not current operational advice. Their evidentiary role is more durable: they show that compatibility depended on installed behavior, not merely on whether a standards document recognized arbitrary prefix lengths.

Manual success was not scalable success

Participants could work around several problems manually. They could install explicit routes, adjust aggregation, correct registry data and coordinate who was authoritative to announce each block. That was enough to keep a bounded experiment moving. It was not a viable general operating model for many exits and many independently changing networks.

RFC 1879 is candid about this. It calls inter-provider cooperation a serious weakness and says manual solutions become unsustainable, especially when a network has multiple exits. A workaround proves that a failure has a local escape. It does not prove that the Internet can absorb the mechanism without continuous exceptional labor.

This is why the report's conclusion must retain its qualifier. The experiment “appears to have been” a success, and former Class A space could be carved up if it was delegated under normal Internet Registry conventions. The result did not say every old router was fixed, every classful protocol had disappeared or every path was safe. It converted an absolute fear—Class A space cannot be subnetted—into a conditional engineering proposition with known requirements and known seams.

Temporary authority left a visible boundary

The present IANA IPv4 registry lists 039/8 as allocated to APNIC from January 2011. That modern entry does not continue the 1995 experiment and says nothing about its participants. Its relevance is narrower and stronger: the experimental designation really was temporary. The address block could return to the shared allocation process and later acquire an unrelated production status.

That separation protects historical interpretation. A route observed today is not evidence about Net 39 in 1995. A present operator in 39/8 did not inherit responsibility for an old experiment. Reuse of the number does not merge institutional epochs.

Heng Lu's Minimum Initial Specification and Voluntary Adoption framework offers a contemporary lens on why the trial worked as governance. IANA supplied a bounded common resource and a closing date. Participants retained local choices about software, providers and configurations. Useful traffic tested voluntary interoperability. The outcome could inform a wider decision without pretending that a central sentence had already changed every router.

The Reality Layers framework adds a discipline for reading the receipts. Registry designation, routing policy, installed configuration, live announcement, packet delivery and application response were related but not interchangeable. Net 39 mattered because the experiment deliberately crossed those layers and recorded where they diverged.

This is a retrospective interpretation, not evidence that the RFC authors used Heng Lu's vocabulary. It does, however, explain why the old F-root address is the defining object in the story. Reversibility was not outside the test. It allowed the test to become real enough to matter. The new address could earn traffic before the old one was destroyed, failures could remain observable without becoming terminal, and a conditional result could replace a speculative dispute.

Sources