Summary
- Seattle Internet Exchange prints eighteen route-server filter tests in the order they run, and a per-ASN error file that names the rejected prefix, the peer, the AS path and the printed reason — an enforcement record detailed enough to be audited line by line.
- Three reviewed pages state that the error files listing dropped prefixes reset daily in UTC, so the detail is real and the history is not; the only dated aggregate is an approximate 5% prefix drop attributed to "Chris" in the 16 April 2020 minutes, printed without baseline, denominator, window or activation date.
The receipt that names your prefix
Start with the granularity, because it is better than most networks manage. The exchange's Route Server Drops page says, in a provider-reported excerpt of its text as rendered on 22 November 2026, 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." The tests are printed in order, and the page offers a focal view for a single ASN.
There are eighteen labels on that list. It opens with "Not peer AS", "Bad next hop" and "Default route", runs through "Path too long", "Bogon prefix", "Cymru fullbogons prefix" and "Too large", and then reaches the part that concerns us: "RPKI ROA_INVALID adjacent", "RPKI ROA_INVALID transit", "No IRR route object adjacent", and four rows that turn on an IRR as-set — whether the ASN is in it, whether the prefix is in it, whether the origin ASN matches the route objects, and whether a route object for that origin exists at all.
The pseudocode printed alongside emits its own strings, which do not match the table labels word for word: the adjacent-origin failure prints as "RPKI ROA_INVALID! Dropping.", the downstream one as "RPKI ROA_INVALID! Transit_dropping."
Labels and log strings are two vocabularies for the same decision, and the site keeps both. That is not a defect to score. It is the ordinary seam between a page written for humans and a daemon that writes lines, and it tells you which surface is authoritative for what: the table for the order, the file for the individual prefix.
The file itself is the good part. A sample per-ASN error file, retrieved as provider-reported text, reads:
2026-06-01 00:00:01 136557:1:1500:v4:Host: 212.81.40.0/24 from 206.81.80.172 (path 136557 209854): RPKI ROA_INVALID! Transit_dropping.
One line carries the timestamp to the second, the instance and AFI context, the organisation, the exact prefix, the peer address, the AS path, and the reason in the route server's own words. If you announce 212.81.40.0/24 as 136557 and a downstream 209854 lacks a valid origin, this line tells you so. No aggregate anywhere in the reviewed set is that useful to the network that lost the announcement. And the Participants page splits the accounting with unusual care: "'err' details are for errors the route server adjacent peer is responsible for, while 'xerr' details are for errors downstream (transited) ASNs are responsible for." A participant is told not merely that something failed but whose evidence failed.
Eighteen tests, one day, no series
Now the sentence that undoes it. The same participants page says: "The error files listing dropped prefixes are reset daily UTC." So does the Route Servers documentation page, in the identical four-word phrase. So, in its own wording, does the IRR Tutorial: "The error counts/details reset on a daily basis." Three reviewed pages, one counting rule, and the rule is that the record begins again every midnight.
The tutorial is the most revealing of the three, because it converts the reset into an instruction. Participants "should also routinely check the SIX participants page or your specific route server drops analysis page to see if your network has any errors." Check routinely — because if you check monthly, you will see the last day of the month and nothing else.
The drops page is candid about its own unit: "The counts in the table reflect the number of unique drops for a test since the start of the UTC day." That is the whole denominator. It is a window, not a population, and it is not comparable to itself across days unless someone saves it. Here is my own inference, and I label it mine rather than the exchange's: because a prefix is dropped at exactly one test and is not tested further, the eighteen per-test counts for one instance would partition that instance's dropped prefixes, so their sum would be the instance's distinct dropped-prefix total for the day.
The page prints no such total, and the numeric cells were blank in the text I could read, so I did not run the sum. I report the rule, not a figure.
That is the gap in one sentence: the exchange publishes a per-prefix ledger that can answer any individual's question today, and publishes nothing that answers the institutional question — how much was refused, of what, over what period. The record is granular and it is not a series.
The 5%, and the surname that is not there
There is one dated aggregate in the reviewed set, and it is not a table. The minutes of the 16 April 2020 annual meeting, as retrieved, record: "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."
Read what that sentence does and does not carry. It carries a speaker, an approximate figure, and a noun — "the prefix counts" — that is not defined. It does not carry the pre-activation count, the post-activation count, the population counted, the measurement window, the activation date the drop is measured from, or any list of affected prefixes. Five per cent of what, over how long, on which instance, is not answerable from the page. The same minutes record an operational line — "Route server now very strict - utilizing RPKI and as-set downstream ASN prefix filtering" — which is a policy statement, not a count.
Now the identity link, stated at its weakest rather than its strongest. The minutes print a given name, "Chris", and no surname in the passage. The join to Chris Caputo — the person the directory records as a route-server developer and operations contributor at this exchange — rests on that role description and on the operational context of the minutes, not on a printed surname anywhere in the sentence. None of the seven reviewed surfaces carries a byline; they are Seattle Internet Exchange pages, and the exchange publishes them. Only the minutes attach a figure to a person, and they attach it by first name.
There is one more date pair on that page worth naming, without reading anything into it. The retrieved header prints "Approved by Membership, April 1st, 2021" above minutes for a meeting held on 16 April 2020. An approval date later than the meeting it approves is odd on its face and may simply be how the header was templated. I cannot resolve it from the text, so I leave both dates where they are printed and draw no conclusion about either.
The General Announcements page dates the two policy steps without printing a number for either: strict filtering using Internet Routing Registry data from 2017-03-07, and "additional strict filtering using RPKI" from 2020-01-15. Those rows are the only place in the reviewed set where a policy change is dated and a person is not quoted. The 5% sits between them in time, in a document written for a members' meeting.
What the record cannot be asked
Put the pieces in one frame and the shape is clear. The exchange can show a named participant, on a named day, which of their own announcements was rejected, by which test, with which path and which reason. It cannot show a regulator, a researcher, or a peer network how many prefixes its filters refused last month, or whether the rate of refusal rose after either policy change. Answering that would require a daily archive of the per-ASN files, a monthly total with its denominator printed beside it, or a rule that the tables keep more than one day. None of the three appears in the reviewed pages.
The reviewed file is a live daily artefact, and the sample line is dated 1 June 2026: it evidences the format, not the contents of any particular day. One lead I could not close belongs here rather than in a claim — no post by the operator to a participants list stating per-ASN error counts was located. That is an unverified lead, not an absence.
What survives is a distinction this record makes unusually sharp. Publishing at a granularity that lets an individual check their own prefix is a real commitment, and SIX does it. Publishing in a form that lets anyone check the system over time is a different commitment, and this record does not make it. The first protects the participant who looks today. The second would protect the argument that the filter was worth its cost — and that argument is currently carried by one approximate percentage in a six-year-old meeting minute, offered by a man identified by his first name.
Every quotation above is a provider-reported excerpt from search-indexed text of the page, rendered 22 November 2026. No page was read as a first-party rendering; the drops table's numeric cells were blank in the retrieved text; and the review covers seven pages on one host on one date. Absence claims are scoped to those seven pages and that day. The file behind those per-instance counts is a plain text log, one line per rejected prefix. Its sample line, dated 1 June 2026 and therefore evidence of the format rather than of any particular day's traffic, carries a timestamp to the second, the instance, ASN, MTU and address family, an organization label, the exact prefix, the peer address, the AS path and the printed reason — in the sample, "RPKI ROA_INVALID! Transit_dropping." — per-ASN xerr log.
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.






