<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by koren on Medium]]></title>
        <description><![CDATA[Stories by koren on Medium]]></description>
        <link>https://medium.com/@koren?source=rss-91ec3ba23e8d------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*xIWIteXhL4OajDya0MwGeQ.jpeg</url>
            <title>Stories by koren on Medium</title>
            <link>https://medium.com/@koren?source=rss-91ec3ba23e8d------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 03:52:04 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@koren/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[DVF : le prix réel des ventes immobilières, et ses cinq angles morts]]></title>
            <link>https://medium.com/@koren/dvf-le-prix-r%C3%A9el-des-ventes-immobili%C3%A8res-et-ses-cinq-angles-morts-3bc1e9a4ce0b?source=rss-91ec3ba23e8d------2</link>
            <guid isPermaLink="false">https://medium.com/p/3bc1e9a4ce0b</guid>
            <dc:creator><![CDATA[koren]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 12:03:16 GMT</pubDate>
            <atom:updated>2026-11-07T12:03:16.261Z</atom:updated>
            <content:encoded><![CDATA[<p>Depuis 2019, l’État publie le prix de vente de presque tous les biens immobiliers de France. Pas une estimation, pas un prix affiché dans une vitrine d’agence : le montant écrit dans l’acte notarié, avec la date, la parcelle cadastrale et la surface. C’est la base <strong>DVF</strong>, <em>Demandes de valeurs foncières</em>, et elle est gratuite, ouverte, téléchargeable par n’importe qui.</p><p>Nous l’avons ingérée en entier dans la base de La Grande Carte le 2026–11–07 : <strong>20 382 915 lignes</strong>, qui correspondent à <strong>7 319 609 ventes</strong> entre le 1ᵉʳ janvier 2021 et le 31 décembre 2025. Ce décalage entre le nombre de lignes et le nombre de ventes est, à lui seul, le piège dans lequel tombent la plupart des gens qui manipulent DVF. On y vient.</p><h3>Ce que c’est, exactement</h3><p>DVF est produite par la <strong>DGFiP</strong> à partir des actes de vente enregistrés par les notaires. Elle est diffusée sous <strong>Licence ouverte</strong> (réutilisation libre, y compris commerciale, avec attribution), deux fois par an — typiquement en avril et en octobre — sur data.gouv.fr.</p><p>Nous utilisons la version <strong>géolocalisée consolidée</strong> préparée par Etalab : le fichier brut de la DGFiP ne contient qu’une adresse en texte, Etalab y rattache l’identifiant de parcelle et des coordonnées. C’est ce qui permet de répondre à la question qui intéresse vraiment les gens : <em>qu’est-ce qui s’est vendu ici, sur cette parcelle, et pour combien ?</em></p><p>Deux caractéristiques structurent tout le reste :</p><ul><li><strong>La base est glissante sur 5 ans.</strong> Aujourd’hui elle commence au 1ᵉʳ janvier 2021. À la prochaine publication, l’année 2021 disparaîtra. Si vous avez besoin d’un historique plus long, c’est à vous de conserver les millésimes précédents — personne ne le fera pour vous.</li><li><strong>La maille, c’est la parcelle.</strong> 100 % de nos 20,4 millions de lignes portent un identifiant de parcelle. Sur les 94,6 millions de parcelles que compte le cadastre français, <strong>9 732 937 ont changé de main en cinq ans</strong>, soit à peu près une sur dix, réparties sur 33 383 communes.</li></ul><h3>Ce qu’on en tire</h3><p><strong>Le prix des voisins, pas le prix annoncé.</strong> Un bien affiché à 320 000 € dans une annonce peut s’être vendu 287 000 €. DVF donne le second chiffre, pour les biens d’à côté, avec leur surface et leur date.</p><p><strong>Le marché d’une commune, en volume.</strong> C’est souvent plus parlant que le prix. À l’échelle nationale, le nombre de ventes est passé de <strong>1 726 304 en 2021</strong> à <strong>1 229 371 en 2024</strong> : moins 29 % en trois ans. 2025 remonte légèrement, à <strong>1 314 051</strong> (+7 %). Un marché qui se contracte d’un tiers en volume ne se lit sur aucun prix médian.</p><p><strong>L’historique d’un bien précis.</strong> Une maison vendue trois fois en cinq ans, ou un terrain acheté puis redécoupé, ça se voit. Et ça se discute, au moment de négocier.</p><h3>Les cinq angles morts</h3><p>C’est la partie que les sites qui revendent de la donnée DVF ne détaillent jamais. Pourtant, chacun de ces points peut vous faire dire une bêtise avec assurance.</p><h3>1. Une vente n’est pas une ligne</h3><p>Une mutation — une vente — génère <strong>une ligne par lot vendu</strong>. Maison + garage + jardin cadastré à part = trois lignes, et <strong>le montant total de la vente est recopié sur chacune des trois</strong>.</p><p>Seules <strong>36,4 %</strong> de nos ventes tiennent sur une seule ligne. Les 63,6 % restantes en ont deux, dix, parfois plusieurs centaines. Additionner la colonne valeur_fonciere donne donc un chiffre qui n&#39;a aucun sens : on compte la même vente autant de fois qu&#39;elle a de lots.</p><p>La règle : <strong>dédoublonner par identifiant de mutation avant toute somme.</strong></p><h3>2. Diviser le prix par la surface ne donne pas un prix au m²</h3><p>Même piège, en plus sournois. Si vous prenez les lignes « Maison » et que vous divisez le montant par la surface bâtie, vous attribuez à la maison le prix de la maison <em>plus</em> celui du garage, du terrain et de tout ce qui partait dans la même vente.</p><p>Mesuré sur 2025 : <strong>38 % des lignes « Maison » appartiennent à une vente qui contenait au moins deux bâtiments.</strong> Sur ces lignes-là, le prix au m² calculé naïvement a un 9ᵉ décile à <strong>7 076 €/m²</strong>, contre <strong>4 423 €/m²</strong> pour les ventes à bâti unique. Soixante pour cent d’écart dans la queue de distribution.</p><p>La bonne nouvelle, parce qu’elle est mesurable elle aussi : la <strong>médiane</strong> nationale, elle, tient bon — 2 207 €/m² avec la méthode naïve, 2 184 €/m² avec une méthode propre, à 1 % près. DVF n’est donc pas « inutilisable », contrairement à ce qu’on lit. Elle est robuste en agrégat et trompeuse au cas par cas. Or c’est précisément au cas par cas qu’on s’en sert pour estimer <em>une</em> maison.</p><h3>3. Trois départements et un DOM n’y sont pas</h3><p>Le <strong>Bas-Rhin (67)</strong>, le <strong>Haut-Rhin (68)</strong> et la <strong>Moselle (57)</strong> relèvent du <em>livre foncier</em> d’Alsace-Moselle, un régime hérité du droit allemand, et non du cadastre de la DGFiP. <strong>Mayotte (976)</strong> est absente aussi. Vérifié dans notre base : zéro ligne pour ces quatre territoires, quand les Vosges voisines en comptent 49 968 et la Meurthe-et-Moselle 80 591. La Corse (2A, 2B) et les quatre autres DOM, eux, sont bien là.</p><p>Si votre outil affiche « aucune transaction » à Strasbourg, ce n’est pas un marché atone : c’est une base qui ne couvre pas le territoire.</p><h3>4. « Vente » ne veut pas dire « vente de marché »</h3><p>Sur nos 7,3 millions de mutations : 6 803 387 ventes ordinaires, mais aussi 405 153 <strong>VEFA</strong> (achats sur plan, dont le prix suit une logique de promotion et pas d’ancien), 62 338 <strong>échanges</strong>, 31 407 ventes de terrain à bâtir, 16 447 <strong>adjudications</strong> (enchères, souvent sous contrainte) et 1 004 <strong>expropriations</strong>.</p><p>Et le montant lui-même mérite un regard : <strong>54 889 mutations à 1 € ou moins</strong>, <strong>262 273 sous 1 000 €</strong> — donations déguisées, ventes intrafamiliales, régularisations — , <strong>20 755 sans aucune valeur</strong>. Soit environ 3,6 % de la base à écarter avant tout calcul.</p><h3>5. Ce que DVF ne dit pas</h3><p>Ni l’état du bien, ni sa performance énergétique, ni son étage, ni sa vue, ni s’il était loué. Deux maisons de 100 m² dans la même rue au même prix peuvent être une ruine et un bien rénové. <strong>8 189 148 de nos lignes n’ont même pas de type de local</strong> : ce sont des terrains et des parcelles non bâties. Et <strong>5 308 320 sont des « dépendances »</strong> — garages, caves, abris — qui, prises isolément, affichent des prix au m² absurdes.</p><p>DVF donne un prix et une surface. Tout le reste du raisonnement reste à faire, avec d’autres données : le DPE, le PLU, l’exposition aux risques, le bruit, les revenus du quartier.</p><h3>Comment on l’affiche</h3><p>Sur <a href="https://lagrandecarte.fr/?src=medium">La Grande Carte</a>, cliquez une parcelle : la section <strong>Transactions</strong> liste les ventes de la parcelle elle-même, puis celles des environs — <strong>dédoublonnées par mutation</strong>, donc une ligne par vente et non une par lot.</p><p>Quand une estimation indicative est affichée, elle est construite sur un échantillon filtré selon exactement les règles de cet article : ventes et VEFA uniquement, montant ≥ 5 000 €, surface ≥ 10 m², prix au m² ≥ 200 €, comparables de taille voisine, rayon élargi de 500 m à 5 km jusqu’à obtenir au moins huit ventes. En dessous, <strong>rien ne s’affiche</strong> : pas d’estimation inventée sur trois ventes — ce qui est aussi, mécaniquement, le cas en Alsace-Moselle et à Mayotte.</p><p>Pour les développeurs, la même donnée sort de l’API publique (/api/v1, documentation sur <a href="https://lagrandecarte.fr/api-doc?src=medium">/api-doc</a>) et du serveur MCP, qui permet d&#39;interroger la base depuis un assistant IA.</p><p>Le site n’a ni publicité, ni tracking, ni inscription obligatoire.</p><p><strong>Source</strong> : Demandes de valeurs foncières (DGFiP), version géolocalisée consolidée par Etalab — data.gouv.fr, Licence ouverte. Données du 1ᵉʳ janvier 2021 au 31 décembre 2025, telles qu&#39;ingérées le 7 novembre 2026. Tous les chiffres de cet article sont issus de requêtes sur cette base ; elles sont reproductibles à partir du jeu de données public.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3bc1e9a4ce0b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Our spam blocklists was disabled for weeks, and how we noticed]]></title>
            <link>https://medium.com/@koren/our-spam-blocklists-was-disabled-for-weeks-and-how-we-noticed-925ce516fd49?source=rss-91ec3ba23e8d------2</link>
            <guid isPermaLink="false">https://medium.com/p/925ce516fd49</guid>
            <category><![CDATA[infrastructure]]></category>
            <category><![CDATA[spam]]></category>
            <category><![CDATA[mail]]></category>
            <category><![CDATA[rspamd]]></category>
            <dc:creator><![CDATA[koren]]></dc:creator>
            <pubDate>Sat, 07 Nov 2026 12:01:43 GMT</pubDate>
            <atom:updated>2026-11-07T12:01:43.600Z</atom:updated>
            <content:encoded><![CDATA[<p>Here is a line from a real spam scan on our own infrastructure:</p><pre>DBL_BLOCKED_OPENRESOLVER(0.00)</pre><p>That symbol does not mean “this message is clean”. It means the Spamhaus Domain Block List refused to answer the query, rspamd noticed, and the check contributed exactly nothing to the score. The message still got a score. The score still looked fine. Nobody was told.</p><p>We run a multi-tenant mail platform: a Postfix per customer, a shared rspamd scanning everything on submission, dedicated IPs, the full SPF/DKIM/DMARC/MTA-STS stack. Spam filtering is not a side feature for us, it is the product’s reputation. And for about five months our outbound scanner ran with <strong>every single DNS blocklist disabled</strong>, while logging the reason roughly four thousand times a week into a file nobody read.</p><p>If you run rspamd, SpamAssassin, or anything else that queries DNSBLs, there is a good chance you are in the same state right now. This post is how it happens, the three commands that tell you, and the fix — including two traps in the fix that nearly cost us customer mail.</p><h3>Two reasonable decisions that multiply badly</h3><p><strong>Decision one, by the blocklist operators.</strong> Spamhaus, DNSWL, SURBL, URIBL and SenderScore all refuse queries that arrive through large public resolvers. This is not hostility, it is capacity: if every query from behind 1.1.1.1 looked identical to them, they could neither rate-limit abuse nor tell a hobbyist from a company that should be paying. So instead of answering the question, they answer a sentinel:</p><pre># through a big public resolver, every query gets the same non-answer<br>zen.spamhaus.org       → 127.255.255.254   &quot;open resolver / excessive use&quot;<br>list.dnswl.org         → 127.0.0.255       &quot;query refused&quot;<br>score.senderscore.com  → 127.255.255.255<br>*.surbl.org            → (nothing at all)</pre><p>Crucially, that is a <strong>successful DNS response</strong>. Not NXDOMAIN, not SERVFAIL, not a timeout. A perfectly well-formed A record, carrying a value that means “go away”.</p><p><strong>Decision two, by rspamd.</strong> rspamd periodically sends each configured list a pair of monitoring probes: one name that must be listed, one that must not. If the “must not” probe comes back listed, the only sane conclusion is that something between rspamd and the list is lying: a hijacked resolver, a captive portal, an ISP wildcarding NXDOMAIN into ad pages. Trusting a lying list is worse than having no list, because a poisoned RBL can be used to block arbitrary mail. So rspamd logs something to the effect of <em>“DNS spoofing detected, disable object”</em> and stops querying that zone.</p><p>Each decision is correct. Put them together and the sentinel that means “you’re using a public resolver” is read as “your DNS is compromised”, and your scanner quietly unplugs its own reputation inputs.</p><p>Now add the third ingredient, which is the one that makes this a five-month bug instead of a five-minute one: <strong>Docker</strong>. Our containers resolved through Docker’s embedded DNS at 127.0.0.11, which looks private and local and fine — and which forwards to whatever is in /etc/docker/daemon.json. Ours was 1.1.1.1 and 8.8.8.8. From Spamhaus&#39; perspective we <em>were</em> a big public resolver.</p><h3>Why nothing alerted</h3><p>Three reasons, and all three are worth checking in your own setup.</p><p><strong>It degrades, it doesn’t fail.</strong> Mail kept flowing. Scores kept being produced. Content rules, Bayes, fuzzy hashes, SPF and DKIM verification all still worked, so scores were plausible, just reputation-blind. There is no error rate to graph, no 5xx, no queue backup. The only signal is the <em>absence</em> of symbols that nobody has an alert on.</p><p><strong>The evidence was in the logs, which is to say nowhere.</strong> About four thousand lines a week on the outbound scanner, about seven thousand on our mail-tester. Logs at that volume are wallpaper. We found it during a routine weekly review, not from monitoring.</p><p><strong>Config does not equal behaviour.</strong> This is the part that still bothers me. Our force_actions.conf had reject rules armed and committed:</p><pre>dbl_serious {<br>  action     = &quot;reject&quot;;<br>  expression = &quot;DBL_PHISH | DBL_MALWARE | DBL_BOTNET&quot;;<br>  message    = &quot;Message references a Spamhaus-listed (phish/malware/botnet) domain.&quot;;<br>}</pre><p>Read the repo and we block phishing, malware and botnet domains fleet-wide. Read production and that rule had never once been able to fire, because the symbols it depends on could never be raised. A code review, a config review, and a rspamadm configtest all pass on a rule that is dead. The only thing that catches it is asking the live system a question.</p><h3>Three commands to check your own stack</h3><p><strong>1. Ask a blocklist its test point, through the resolver your scanner actually uses.</strong> Every major list publishes one name that must come back listed. If you get the sentinel instead of the expected answer, you are blocked:</p><pre># expected from a resolver they answer: 127.0.0.2 (Spamhaus test point)<br># sentinel: 127.255.255.254  → your resolver is refused<br>dig +short 2.0.0.127.zen.spamhaus.org</pre><pre># expected: 127.0.1.2      sentinel: 127.255.255.254<br>dig +short dbltest.com.dbl.spamhaus.org</pre><pre># expected: 127.0.x.y      sentinel: 127.0.0.255<br>dig +short 2.0.0.127.list.dnswl.org</pre><p>Try it right now on the machine you’re reading this on. I just did, on a normal consumer connection, and got 127.255.255.254, 127.255.255.254, 127.0.0.255 ; three sentinels, three refusals. That is the default state of the internet, which is why this bug is so easy to ship.</p><p>Then run it <em>from inside the container</em>, not from your laptop: the question is which resolver the scanner uses, and in a container that is almost never the one you think. 127.0.0.11 is not a local resolver, it is a forwarder to whatever your daemon config says.</p><p><strong>2. Grep the scanner’s log for the monitor giving up.</strong></p><pre>docker logs rspamd 2&gt;&amp;1 | grep -iE &#39;spoof|disable object&#39; | sort | uniq -c</pre><p>Zero lines is the goal. Thousands of lines means the list named in them has been off since whenever those lines started. Ours went back as far as our log retention did, which is how we learned the answer to “since when” is “we don’t know, at least five months”.</p><p><strong>3. Scan a message and look for </strong><strong>_BLOCKED symbols.</strong> Any symbol ending in _BLOCKED ,DBL_BLOCKED_OPENRESOLVER, SURBL_BLOCKED, URIBL_BLOCKED ; is the list telling you it declined to answer. It is not a verdict about the message. Which brings us to the expensive trap.</p><h3>The trap: *_BLOCKED is not a listing</h3><p>When we fixed the resolver, we also went to arm a second reject rule that had been dead just as long. As written, it rejected on this:</p><pre># what we nearly shipped<br>expression = &quot;SURBL_BLOCKED | URIBL_RED | ...&quot;;</pre><p>Both halves are landmines.</p><p>SURBL_BLOCKED means <em>&quot;the list refused or throttled our query&quot;</em>. With the blocklists dead it had never fired. With a working resolver it fires the first time we cross a free-tier rate limit ; and that rule would then have rejected <strong>every outbound message containing a URL</strong>, platform-wide, for as long as the throttle lasted. A fix to one bug arming an outage in another.</p><p>URIBL_RED is subtler. It is not a spam verdict; it lists young domains and whois-privacy domains. Our customers buy fresh domains for their mail; that is literally the product. And the URL blocklist modules check not only links in the body but the sender&#39;s own From, Reply-To, DKIM and HELO domains. That rule would have blocked new customers&#39; mail on day one, with a message claiming their domain was on a spam list.</p><p>What shipped instead rejects only on confirmed-spam lists:</p><pre>spam_url_listed {<br>  action     = &quot;reject&quot;;<br>  expression = &quot;URIBL_BLACK | ABUSE_SURBL | PH_SURBL_MULTI | MW_SURBL_MULTI&quot;;<br>  message    = &quot;Message references a domain on a public spam block list (URIBL/SURBL).&quot;;<br>}</pre><p>Before deploying it we looked up every customer domain on those four lists through the new resolver. None was listed. That check takes two minutes and is the difference between a feature and an incident.</p><p><strong>Rule of thumb: never put a </strong><strong>_BLOCKED symbol in a reject expression.</strong> It describes your infrastructure, not the message.</p><h3>The fix: recurse yourself</h3><p>The lists are happy to answer a normal resolver querying on its own behalf. So we stopped borrowing one. A small unbound container does full recursion from the host’s own IP, and only the scanner can talk to it:</p><pre>server:<br>  interface: 0.0.0.0<br>  port: 5353<br>  access-control: 0.0.0.0/0 refuse<br>  access-control: 10.253.53.0/29 allow      # just the scanner and us</pre><pre>  username: &quot;&quot;            # already unprivileged in the container<br>  chroot: &quot;&quot;<br>  trust-anchor-file: &quot;/usr/share/dnssec-root/trusted-key.key&quot;</pre><pre>  # Blocklist zones answer NXDOMAIN at intermediate labels, which a<br>  # minimising resolver can read as &quot;nothing below here&quot; and turn a<br>  # listing into a miss.<br>  qname-minimisation: no</pre><pre>  prefetch: yes<br>  hide-identity: yes<br>  hide-version: yes</pre><p>That qname-minimisation: no is not optional. It is a privacy feature that is on by default in modern unbound, and it quietly breaks DNSBL lookups, which are the one query type where leaking the full name to the authoritative server <em>is the entire point</em>.</p><p>Four more things we learned wiring it up, each of which cost an hour:</p><ul><li><strong>rspamd resolves nameserver names exactly once, at startup.</strong> Point it at a container by name and the day that container is recreated on a new IP, your scanner silently falls back and stays there until restarted. Pin the resolver’s IP on a dedicated bridge. Pin the scanner’s too, or they race for the first address in the subnet and whoever loses doesn’t start.</li><li><strong>Keep a fallback, and make it the safe one.</strong> Our config is master-slave:10.253.53.2:5353,127.0.0.11. If the resolver dies, lookups go back to Docker&#39;s DNS: the blocklists degrade, mail keeps moving. It returns to the private resolver within about 30 seconds of it answering again. Failing <em>open</em> on a reputation input is right; failing closed means an unbound restart stops customer mail.</li><li><strong>“Free” lists are free for non-commercial or low-volume use.</strong> Recursing yourself makes them technically answer you; it does not make a commercial product’s queries licensed. We kept Spamhaus on their keyed Data Query Service for exactly this reason, and the same question is still open on the others. Read the terms before you depend on a list in production.</li><li><strong>One zone cannot be fixed this way.</strong> SenderScore answers the probe that must be NXDOMAIN with a listing, from any resolver, so rspamd disables it everywhere and logs about it forever. Known, harmless, permanently noisy ; and a good argument for alerting on <em>which</em> zone is disabled rather than on the count of lines.</li></ul><h3>The part we haven’t fixed</h3><p>Honesty, since this is a post about things being quietly broken: the second rspamd we run — the one behind our free mail-tester, is, as of the day I’m writing this, <strong>still in exactly the state described above</strong>. It resolves through public resolvers and every list it knows about is disabled. We found it a month after fixing the outbound side, during an unrelated audit, by seeing DBL_BLOCKED_OPENRESOLVER(0.00) in a real visitor&#39;s scan. Thirty days of logs say the same thing every day before that.</p><p>It is not as bad as it sounds, and the reason why is the interesting part: the tester doesn’t rely on the scanner for reputation. The app runs its own blocklist probes in its own code path, against Spamhaus’ keyed service and a handful of open lists, and, the detail that matters, it treats the 127.255.255.x range as <strong>indeterminate</strong>, never as &quot;not listed&quot;. So its verdicts are honest. What&#39;s lost is the scanner&#39;s own list symbols, and what leaks is a _BLOCKED symbol into a user-visible breakdown.</p><p>That accident is the whole lesson, though. One code path was written by someone who knew the sentinels exist and handled them explicitly; the other trusted a library’s defaults. Only one of them stayed correct.</p><p>Which generalises to something uncomfortable about every free deliverability scorer, ours included: <strong>a reputation checker whose reputation sources are dead still gives you a number.</strong> It will usually be a <em>better</em> number than the truth. If a tester tells you 10/10 and never tells you which checks it couldn’t perform, you have learned nothing about where your mail lands.</p><p>So the real takeaway isn’t “configure a resolver”. It is: <strong>monitor the monitor.</strong> One alert on disable object in your scanner&#39;s log, scoped to zone names, would have caught five months of blind scanning in the first hour; and would have caught the mail-tester the same day instead of a month later. It is the one piece of this we still don&#39;t have, which is why the second instance went on being broken after we understood exactly why the first one was. Fixing the cause is satisfying. Alerting on the symptom is what actually protects you.</p><p><em>We run </em><a href="https://adbtd.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=blocklists-silently-disabled"><em>adbtd</em></a><em>: dedicated IPs and mailboxes on your own domain, with the full auth stack and the spam scoring above. If you want to know whether your own mail lands, </em><a href="https://check.adbtd.com/?utm_source=medium&amp;utm_medium=article&amp;utm_campaign=blocklists-silently-disabled"><em>check.adbtd.com</em></a><em> scores any message in 30 seconds, no signup, its own blocklist probes query Spamhaus’ keyed service directly and treat a refusal as unknown, never as clean. The scanner behind it is next.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=925ce516fd49" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>