Summary
- The testbed publishes an allocation-only origination rule, a default announcement cap, a ban as its sanction, and the prefix filter an operator is meant to paste. Every obligation it prints is addressed to the experimenter.
- The proposal as approved — the prefixes, announcement shape and change frequency staff signed off — is not among the pages published on 19 November 2026, and the same publisher prints two announcement caps, two ASN counts and two prefix inventories without reconciling them.
An experiment called “SBAS Secure Backbone Prototype” was approved at Princeton University on 06/30/2020 2:10 p.m. and allocated 66.180.190.0/23 — a /23 of routable IPv4 space to announce on the public internet. That is what the testbed's own table prints, and it is the whole of the public record for that decision: a title, an institution, a timestamp, a prefix.
The document the decision rested on is not part of the record. Before an experimenter announces anything, PEERING asks them to describe how many prefixes they need, what their announcements will look like, how frequently those announcements will change, which PEERING servers will send them, and any modification to BGP attributes — plus whether traffic will be exchanged and, if so, its type, average and peak bandwidth and the remote endpoints (Experiments). That description was written. Staff read it, checked it against the acceptable-use policy and replied with design feedback. It is not on the site.
That absence is the story, and it is a documentary absence. Nothing here says an experiment misbehaved, that a route leaked, that any network was surprised by an announcement, or that a prefix was used for something other than research. Those claims would need BGP data I did not collect. What I can read, and what a receiving network can read, is a rulebook with a hole where the authorisation should be.
What the register prints, and what it does not
The control surface starts at a registry object. ARIN records AS47065 as PEERING-RESEARCH-TESTBED-USC-UFMG-AS47065, held by the USC / UFMG PEERING Research Testbed, registered 2009-06-16 and last updated 2019-04-23 (ARIN record). That object identifies a holder. It does not say what the holder announces, and it was not touched for six years while the project's prefix inventory changed shape.
The project's own claim is deliberately narrow, and worth quoting exactly: “The PEERING testbed will announce only prefixes allocated to the testbed or to researchers using the testbed. This condition ensures that the testbed can only affect traffic destined to PEERING IP addresses. The testbed filters announcement of any other prefix.” It adds that it will only announce BGP messages whose AS-PATHs start with AS numbers allocated to it or to its researchers (About).
That paragraph is the load-bearing claim of the entire control surface, and it is a claim about the testbed's own machinery. I ran no measurement that verifies it. A network that accepts a testbed route is trusting a sentence on a web page, plus its own inbound filters.
Three gates, all of them pointing inward
The acceptable-use policy is where the obligations live, and every one of them refers to a document the reader does not have. Experiments “MUST follow the proposal as approved by PEERING support, and any desired changes in experimental plans MUST be approved by PEERING support before being run.” An experimenter must “not announce any prefixes outside of those approved for your use by PEERING support in response to your experiment proposal,” and must “only make announcements as described in your proposal AND approved by PEERING support, and only make announcements for the time period for which PEERING support has explicitly approved your use” (AUP).
The same page forbids “ill-crafted or anomalous BGP announcements, including announcements that are larger/longer than announcements that occur regularly on the ‘normal’ Internet,” and caps the default rate unless support explicitly approves otherwise.
Now read the enforcement clause. The consequences are disabling the account, banning the institution and/or the researcher from PEERING, and informing the organisation's administration. Every sanction is addressed to the person who asked to announce. Not one of them creates a record a stranger can check. The rulebook is a discipline imposed on the experimenter, published in full; the authorisation it enforces is an internal artefact.
The approval is also the plumbing
It would be easy to read the proposal process as paperwork. It is not. The experiments page says approved proposers receive certificates for VPN tunnels to, and BGP sessions with, PEERING servers. The approval is the thing that provisions transport. Without the certificate there is no session, and without the session there is no announcement.
That makes the unpublished document more consequential, not less. The instrument that decides whether a prefix appears in the global table is the same instrument that hands out the credentials to place it there, and it is held in one place by the same staff who verify that the experiment follows policy. The authority to authorise and the authority to verify are not separated here; they are the same desk.
Two caps, two counts, no reconciliation
The publisher prints its own numbers twice and does not reconcile them. The About page limits each BGP peer to “at most 10 announcements per hour for each PEERING prefix.” The policy page says that, unless explicitly approved, “announcements of a particular prefix cannot be more frequent than once every 10 minutes (averaged over a day), and announcements cannot be made to a particular PEERING node more frequently than once every 10 minutes (averaged over a day).”
Those are different shapes: a per-peer hourly ceiling against a per-prefix and per-node ten-minute floor with a daily averaging window. An operator building a filter from the first sentence and an experimenter planning to the second are not working from one rule. I print both and adopt neither.
The counts diverge too. The About page says “In most cases, any of our 7 ASNs can originate announcements” and “We control 3 IPv4 prefixes, and 1 IPv6 prefix,” listing 184.164.224.0/19, 138.185.228.0/22, 204.9.168.0/22 and 2804:269c::/32. The peering page says “Any of our 8 ASNs can orignate announcements for the prefixes listed below” — the publisher's own spelling — and lists eight IPv4 prefixes and three IPv6 prefixes, including 66.180.190.0/23, 45.132.188.0/22, 147.28.2.0/23, 147.28.4.0/23, 147.28.8.0/21, 2604:4540::/48 and 2602:f83d::/40 (Peering). The seven additional ASNs are enumerated identically on both pages: 61574, 61575, 61576, 263842, 263843, 263844 and 33207.
This is not pedantry about a stale page. The same peering page publishes the Cisco/Quagga-style prefix filter an operator is meant to paste, and that filter permits 66.180.190.0/23 up to /32. Which inventory is current is not stated anywhere I read. An operator who builds to the smaller list drops traffic the testbed considers its own; one who builds to the larger list accepts more than the About page admits.
The one instance you can check
The peering page says the maintained prefix list is a route-set called RS-PEERING-TESTBED in RIPE's IRR. Query that name through a public IRR mirror and you get several objects rather than one, sourced from ARIN, RADB and RIPE, all described Columbia/UFMG PEERING Research Testbed and all pointing back at the project site. Their member lists are not identical: the ARIN-sourced object renders 2804:269c::/42 where others render /32, 147.28.10.0/23 appears in some members and not in the ARIN-sourced listing, and 66.180.190.0/23, 2604:4540::/48 and 2602:f83d:80::/43 sit under a “PEERING user prefixes” remark (route-set).
Now the prefix that carries the only approved experiment I could read. Query 66.180.190.0/23 and you get route objects naming three different origins: AS47065, AS88 and AS400065. The AS47065 object is described “Princeton University”; one ARIN-sourced copy was created 2020-07-02 and one RADB copy was changed 2020-06-12 by a princeton.edu contact, both citing Princeton OIT ticket REQ0166317 (prefix). This is a route-object listing, not a live BGP measurement. I did not query a looking glass, a collector or an RPKI repository, so which origin is announced today is not established here — and that uncertainty is exactly the point. The published approval surface and the published routing surface do not line up one-to-one, and a stranger cannot close the gap from documents alone.
The last layer is self-reported. A profile page for AS47065 lists the network as Columbia/UFMG PEERING Research Testbed, also known as Transit Portal and BGPMUX, with an open peering policy, network type Educational/Research, and 100 IPv4 and 300 IPv6 prefixes, last updated 2026-06-24. It files RS-PEERING-TESTBED under an “IRR as-set” field, though the project's own page and the IRR view both render it as a route-set (profile). Self-reported, not authoritative, and inconsistent with the project's own vocabulary.
Who is actually on the other end
Here the publisher's own text concedes the shape of the problem. The About page asks experiments, where possible, to offer remote networks an opt-out option, naming internet operator forums such as the NANOG mailing list as the route.
That sentence is worth slowing down for. A peer can refuse the session, or filter the prefix list, and it chose to accept the testbed's routes in the first place. A third party whose traffic depends on an address inside 184.164.224.0/19 or 66.180.190.0/23 made no such choice and holds no session to withdraw. The published opt-out channel is a mailing list, which is a real channel but not a contract, and the fact that the guidance is addressed to “remote networks” generally rather than to peers is an admission that the affected population is wider than the set that agreed to anything.
The reasonable reading is not that PEERING is hiding something. A research testbed that published every proposal would chill the experiments it exists to enable, and the sanction regime is genuinely strict by the standards of this industry. The reasonable reading is that the disclosure perimeter stops at the experimenter's obligations and never reaches the recipient's need. A network seeing an unfamiliar prefix from AS47065 can check that it falls inside a published allocation, and that is where the paper trail ends.
What would close it
One paragraph would do most of the work: for each approved experiment, the prefix, the approved announcement window, the maximum change frequency and the originating ASN — not the whole plan, just the four fields the policy already treats as conditions. An operator could then compare what it sees against what was approved, and a researcher would lose nothing commercially sensitive.
The second-most useful thing is boring: pick one frequency rule. A per-peer hourly ceiling and a per-prefix ten-minute floor with a daily average cannot both be the operative cap, and whichever governs should be the one printed on the page an operator reads.
Until then, the honest description of this control surface is that it publishes the discipline and withholds the licence. The discipline is real, detailed and enforced. The licence is a certificate on a staff machine.
Boundaries
- Everything here is a reading of pages and records retrieved on 19 November 2026. A retrieval limit is a fact about access, not a statement about the publisher.
- The four first-party pages carry no version number or last-updated date in what I read. I cannot date them and cannot say which is current.
- Registry and IRR material came through a registry REST rendering and a third-party IRR mirror, not from querying the registry or an IRR directly. One route-set name returns several objects whose members differ.
- Where the same publisher prints two formulations of a cap, two ASN counts and two prefix inventories, I print both and adopt neither, and make no finding about which governs.
- The approved proposal, its announcement plan, the certificates and the BGP sessions are not among the pages I read. I have not seen any approved proposal text and cannot describe what one contains. That is an absence from the pages retrieved, not proof that no such record exists or that publication was ever intended.
- For 66.180.190.0/23 the IRR view lists three origins. That listing is not a live BGP measurement, and which origin is announced today is not established here.
- Nothing here asserts that any experiment leaked, harmed, surprised or withdrew routes from any network, or that any resource was misused. No incident is reported and none is implied.
- The experiments page's tables were truncated in the retrieved text, so the approved experiment I quote is one row, not a census.
- The publisher's claim that its rule means it “can only affect traffic destined to PEERING IP addresses” is quoted as the publisher's claim. I ran no measurement and do not verify it.
- The self-reported profile is not authoritative for prefix authorisation, and the registry record identifies a holder without stating what is announced.
- A literal search of this publication's archive for the subject returned no indexed rows on 19 November 2026. That is a retrieval result about one search string, not a claim that nothing has been published about the testbed anywhere.
- The prefix counts in the two inventory printings are reported as printed. No arithmetic of mine is offered as the publisher's total.
Member Briefing
Read the full trend briefing
Members read the full briefing with the source notes behind every finding.
Strategic Circle
A daily briefing in your language, built from changes we have verified.
Join for USD 20/monthCancel anytime.
Members also get the full research library, a daily briefing on the networks they follow and the private forum.






