Some things here are checked automatically. None of those checks can remove anything: only a person removes, and every removal is recorded and appealable. A pitch with no specific detail in it may stop being shown to search engines until somebody looks — it stays published and stays at its address. Endorsements from a group of accounts that arrived together, acted once and went quiet may stop counting toward a public total; the endorsements themselves are kept and a later review can restore them. The launch board only ranks endorsements from accounts that existed before the launch, which is why a brand-new account's endorsement counts on a listing and does not move a board position.
Deployed · version 4c164997
The checks
Four, and each says what it can do to what it finds. Everything here raises a flag for a person to read; the two that do anything else do the least reversible thing available.
This pitch may not say anythingmay withhold from search engines
The text is long enough to be judged and contains no specific detail — no version, no name, no dependency, no number — while leaning heavily on words that carry no information. This is a measure of the TEXT, not of who wrote it: we do not attempt to detect whether a person or a model produced it, and a human writing boilerplate scores the same way.
A group of accounts that arrived togethermay stop counting toward a total
Several accounts first seen within a short window of one another acted on the same thing, and did nothing else in the days afterwards. A launch shared in a newsletter looks the same on the day, which is why this is only assessed once those days have passed.
Accounts endorsing substantially the same setmay stop counting toward a total
A group whose endorsing overlaps almost completely and which has little independent activity. The endorsements are discounted from public totals and remain recorded.
Most of a launch's endorsements came from accounts younger than the launchdoes nothing on its own
Nothing has been done about this, and nothing needs to be: the launch board only ranks endorsements from accounts that existed before the window opened, so these were never counted toward its position. The flag is here so a person can see the pattern.
How often the pitch check is wrong
Only one of the four reads anybody's writing, and it is the only one that could be wrong about a person rather than about a pattern. So it is the one we measured.
0 of 20 genuine pitches were wrongly flagged. 6 of 6 empty ones were caught.
This rate is measured against a corpus assembled for the purpose, not a sample of real listings — running an experiment on identifiable makers' words and publishing the results is a thing we would be doing to them. The corpus is deliberately loaded with the hardest honest cases we could think of: pitches written in a second language, products with no numbers or code in them at all, and makers writing in exactly the enthusiastic register the check looks for. It is a floor on what we know, not a claim about what happens in production.
When a person reviewing a flag decides the check was wrong, that is recorded as a dismissal and counted. Those counts are what tell us whether the number above still means anything once real listings are involved, and they appear in the transparency record alongside everything else.
What we do not do
The more important half of this page. What a system catches is visible from outside; what it declines to attempt is not, and it is where the harm would have been.
Guessing whether a person or a model wrote something
Detectors of machine-written text score people writing in a second language as machine-written far more often than native speakers. On a directory of open-source products that would quietly penalise exactly the contributors we most want. We check whether text contains any specific detail, which is a property of the text and the thing a reader actually cares about.
Removing anything automatically
A false positive here would erase somebody's work from a public page with no notice. Automation may flag, discount and stop amplifying; the database itself refuses to record a removal that no person made.
Telling anyone which of their endorsements was discounted
It would be an accusation made by a machine to somebody with no way to answer it, and it would tell a group of accounts exactly which of them we noticed. We publish the totals instead.
Acting on a burst of new accounts on the day it happens
A launch shared somewhere popular looks exactly like a purchased one for the first few days. We wait until the difference is visible, and until then the launch board's own rule keeps the ranking honest without anyone being accused of anything.
What computes the numbers
Seven phases, on a schedule, with no person in the loop. Each says what it computes, how often it is asked to, and — read live from the same document the operators read — when it last did, and whether it is failing. A phase that has never run says so rather than rendering a zero.
The catalogueEvery ten minutes, after the upvote integrity pass and before the search index is built from it.
Writes the one table the front page and the API read — every listed product with its latest observations and its pictures, in the published order — so a page is one indexed read rather than a join over everything. Not a cache: the page says when it was computed.
Visits each listed product's own address in a browser and keeps a dated picture of it, shown in the observed frame and withdrawn past its window rather than frozen.
Not configured in this environment: no address is visited, no capture is shown, and no listing is marked as missing one. Up to 5 captures a run, each retaken after 7 days and withdrawn from the page after 30.
What we promise
Three sentences, each one a reader of the product could check, written before anything alerts on them. The number in each is read from the constant beside it, so the sentence cannot drift from the code.
1.A published listing is in search within 10 minutes — one materialisation interval — and the search index is reported unhealthy on /health after 30 minutes.
Where a breach is found (Workers Logs): msg = "sweep_phase_failed" AND phase = "search_rebuild"
2.Every count on the front page is younger than 3 hours — three integrity intervals — or the integrity phase is reported unhealthy on /health.
Read from SWEEP_STALE_AFTER_SECONDS.integrity = 3 hours
Where a breach is found (Workers Logs): msg = "sweep_phase_failed" AND phase = "integrity"
3.An anonymous front-page read is a colo cache hit 9 times in 10.
Read from CACHE_HIT_TARGET = 0.9 ratio
Where a breach is found (cache analytics): cf-cache-status = HIT over requests to the apex `/` with no cookie, as a share of all such requests, over the last hour
If something here happened to a listing you maintain and you think it was wrong, every act is appealable — including the automatic ones, which is why each of them leaves a record with a reason attached.