Summary
- SeattleIX prints its route-server filter as eighteen named tests in a fixed sequence and states that a prefix failing any test "is dropped prior to further tests", so the reason a member sees is the first failed test, not the whole fault.
- The published pseudocode lets a valid ROA retire IRR route-object verification for an adjacent origin, while a downstream origin must still belong to the peer's IRR as-set before any RPKI shortcut runs; no cryptographic attestation substitutes for membership at that hop.
- The only cost figure in the reviewed record is a 2020 indication of "an approximate 5% drop in the prefix counts", carrying no baseline, denominator, window or activation date, inside minutes headed by Chris Caputo whose printed approval line is dated a year before the meeting it approves.
The order is the finding
Filter documentation usually answers one question: what gets me dropped. The drops page at SeattleIX answers a second, more awkward one: in what order. It opens with the sentence that turns a list into a machine —the drops page states that when a prefix is advertised, the route server "performs a series of tests to decide whether the prefix may be accepted. If a prefix fails to pass any test, it is dropped prior to further tests." Then it prints its reasons "in order of tests": Not peer AS, Bad next hop, Default route, Local subnet, Path too long, Bogon prefix, Cymru fullbogons prefix, Too large, Bogon ASN, PeeringDB ASN never via RS, Transit ASN, RPKI ROA_INVALID adjacent, RPKI ROA_INVALID transit, No IRR route object adjacent, ASN not in as-set, Prefix not in as-set, Prefix in as-set but origin ASN IRR mismatch route objects, Prefix in as-set but no IRR route object for origin ASN.
Read the sequence and a design intention shows through. The first eleven tests are about identity and topology: is this my peer, is the next hop sane, is the path absurd, is this address space a bogon, is the origin a known transit provider or an AS that has asked never to be routed via route servers. Only then does the machinery reach into the two registry systems — RPKI, then IRR — and, inside the IRR block, into as-set membership and origin consistency. The cryptographic check sits above the registry check it is meant to make redundant.
This is not a taxonomy. It is an execution path, and it matters that the exchange printed it, because most operators publish the fact of filtering and not its ordering. Two dated announcements supply the timeline: an entry dated 2017-03-07 says the route servers "now perform strict filtering using Internet Routing Registry data", and an entry dated 2020-01-15 says they "now perform additional strict filtering using RPKI" — each date belonging to its own entry on a page that carries no byline and no single publication date (general announcements).
A rank is not a diagnosis
Once a prefix is dropped at its first failure, the label a member receives is positional. A prefix killed at "Path too long" never learns whether it would also have failed the as-set test four rows later, and a prefix killed at "ASN not in as-set" never learns whether a valid ROA would have carried it past the IRR check had the run reached that point. The published reason is the earliest disqualifying condition, not the set of disqualifying conditions. That follows from two printed sentences rather than from anything the exchange says about its motives.
The counting rule on the same page widens the gap between what a member sees and what the table measures: "The counts in the table reflect the number of unique drops for a test since the start of the UTC day" (drops page). The unit is a drop per test per day, which is neither a prefix census nor an outage measure — a prefix re-announced into the same failure is a second event, and a prefix failing at row fifteen is invisible at rows sixteen through eighteen. On the copy of the page consulted for this article the dynamic counts were blank, so no drop volume is quoted here. What remains solid is the shape: the exchange chose to count events per test rather than to publish the population it filters.
Where a valid ROA buys a skip, and where nothing does
The route-server documentation states the substitution plainly. RPKI data updates hourly; if an ROA exists for a prefix and the route servers receive an announcement in violation, "it will be dropped"; and "At the SIX, an RPKI ROA replaces the need for an IRR route/route6 object. An IRR as-set remains needed for downstream ASNs" (route servers).
The pseudocode shows exactly how far that replacement travels. On a valid result the code sets rpki_roa_valid = 1 with the comment # IRR route object verification is not needed when RPKI ROA is valid. For an origin adjacent to the route server, the IRR peer-prefix check is guarded by if 0 = rpki_roa_valid, so a valid ROA skips it. For an origin further down the path the order is reversed: the code first tests whether the last non-aggregated ASN sits in the peer's as-set and drops with not member of IRR as-set object! before any RPKI shortcut applies; only the later as-set prefix test and the IRR origin check are guarded by if 0 = rpki_roa_valid (drops page).
That is a hop-limited substitution, and it is the most interesting line of code in the record. A cryptographically signed assertion can retire a registry object one hop away from the route server. It cannot retire a membership question one hop further out, because "is this ASN in my peer's as-set" is not an evidence-ranking problem — it is a question about which networks the peer has declared it will transit for, and only the peer's own as-set answers it.
Caputo's own priority statement, written in April 2022, sits precisely on the evidence-ranking half of that split. In a message to the ARIN public policy mailing list he writes that he speaks "in my opinion as the developer of the SeattleIX/SIX's strict route server filtering technology", that "At the SIX, we prioritize RPKI information over IRR and ARIN OriginAS data", and that "RPKI is secure and thus preferred over OriginAS whois" (ARIN-PPML message). The statement is a ranking of evidence sources. The pseudocode honours it at the adjacent hop and stops there, because at the downstream hop the question changed from evidence to membership.
The one number the record leaves standing
The 16 April 2020 annual meeting minutes are the only document in the set that puts a cost on the change. They record, in the operational update attributed on the page to Chris Caputo, that the "Route server now very strict - utilizing RPKI and as-set downstream ASN prefix filtering", and in the question section that "Peter Helmenstine asked about the impact of using RPKI on the route servers. Chris indicated there was an approximate 5% drop in the prefix counts when it was activated" (2020 minutes).
Everything needed to audit that figure is absent from the same document. No pre-activation count, no post-activation count, no denominator for "the prefix counts", no measurement window, no activation date and no list of affected prefixes. The RPKI announcement entry is dated 2020-01-15 and the meeting is 16 April 2020, so by my subtraction of those two printed dates 92 days separate them — mine, not the exchange's, and it tells us only that an activation could have preceded the meeting. Between the 2017-03-07 IRR entry and the 2020-01-15 RPKI entry, again by my own subtraction, 1,044 days pass.
The header of the minutes is worth reading as carefully as the body, because it carries the authorship and a date that does not sit still. As retrieved, it reads: "Seattle Internet Exchange Annual Meeting April 16th, 2020 Teleconference Minutes by Chris Caputo Approved by Membership, April 1st, 2021." The approval date printed on the page precedes the meeting date printed on the same page by 350 days, my subtraction again.
Later in the same document the previous year's minutes are recorded as moved, seconded and passed "by acclamation and without objection", so a recorded approval is not unusual there; what is unusual is one dated before the meeting whose record it approves. I note the discrepancy as printed. It is not evidence of anything beyond itself, and the arithmetic defect in a document is never an accusation about the hand that wrote it.
Authorship: the minutes and the opinion
The two documents that carry Caputo's name do different work. The minutes are headed with his byline and are an institutional record approved by the membership; the 2022 mailing-list message is one individual's stated opinion on a public policy list, not an exchange policy document. Neither is a signed policy instrument, and the two operational pages that describe the filter carry no byline and no date at all.
So the person-level story here is not a story about an office. It is about a written act: the same hand that wrote the members' record of the strict filter also names itself, two years later, as the developer of that filtering technology and states its evidence priority in public, and points readers to the drops page for the details. Katrin — that is the chain, and it runs through documents a stranger can open.
Who carries the error, and where a member takes it
The participant table splits the consequence by column, and the split is worth understanding before arguing with a drop. As retrieved, it explains that the err columns "are for errors the route server adjacent peer is responsible for, while xerr details are for errors downstream (transited) ASNs are responsible for", with per-instance counts and links into a drops view focused on a single ASN (participants). That page was retrieved as page text and was not opened in full here; what it demonstrates is that the exchange assigns responsibility by hop, not by member.
Where a member goes to disagree is a shorter answer than the mechanism deserves. Across the FAQ and rules pages reviewed on 22 November 2026, no procedure for contesting a route-server drop is printed; the rules page carries the peering sentence — "Peering is bilateral with the exception of route servers. There is no Multi-Lateral Peering Agreement (MLPA)" — and a suspension provision aimed at participant unresponsiveness, not at filter disputes (rules). The route-server page does invite mail if an as-set is not being recognised, "for special handling" (route servers), which is a practical channel and not a printed appeal. Those two pages were read as page text rather than opened in full, so the absence is claimed for the text reviewed, not for the whole site.
Optionality, and what the exit actually exits
The exchange describes the service as optional — "We offer optional route servers" (about) — and honours the PeeringDB option by which a network asks never to be reached via route servers. Both are real exits. Neither is an exemption inside the service. No page reviewed here prints a way to keep using the route servers while opting out of the filter, and the distinction matters: the choice on offer is whether to take the filtered path, not whether to be filtered while on it.
What these eight pages do not attest
- The live drop counts on the drops page were not populated in the text retrieved, so this article cites no drop volume.
- The participants and rules pages were retrieved as page text and were not opened in full, which scopes every absence claim about them.
- No reviewed page prints a pre- or post-activation prefix count, so the 5% indication cannot be reproduced from this document set.
- No reviewed page prints a procedure for contesting a drop attributed by
errorxerr. - The 2022 mailing-list message is one individual's opinion and priority order, not an exchange policy document.
The bill for publishing the mechanism
Publishing an execution path is a serious act of accountability, and the Seattle Internet Exchange did it: eighteen tests in sequence, the pseudocode that applies them, the counting rule, and a per-ASN view that names which hop carries the blame. Very few exchanges put that much of their reasoning on a public page, and the reasoning is honest about its own asymmetry — the cryptographic proof is preferred where it applies and the registry membership requirement is kept where it is still the only witness.
What the record does not carry is the arithmetic. One sentence in a members' meeting is doing all the work of justifying the cost of the strictest filter in the published material, and it is a sentence no member can check: no baseline, no denominator, no window, no date of activation. The exchange published its reasoning down to the last if. The sentence its members are likeliest to repeat is the one with no numbers behind it — and there is still no printed place to take a drop you believe is wrong.
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.






