When something published here is wrong, it gets corrected in the open rather than quietly edited. This page is the record. It is append-only: nothing is removed from it, including the errors, because a record you can delete from is not a record.
Each entry links back to the page it belongs to. The correction is also marked on that page, where the claim was, so a reader meeting the old version in a search result or a shared link finds the correction next to it rather than having to come here.
How long a notice stays where the claim was
A correction whose subject is still on the page stays in place permanently. The clearest case is the ETL showcase: the example dashboard's figures are invented, they are still displayed, and the marker saying so is what makes them honest. That is a standing disclosure, not history, and it never collapses at any age.
A correction whose subject was removed stays where the claim was for 90 days, then collapses to a dated one-line link into this page, with the full text moving here word for word. Ninety days covers the window in which someone can still meet the old claim in a cached result, a shared link, or a model that crawled the page before the fix — long enough that they land on the correction, short enough that a page is not permanently explaining an absence to people who never saw the presence.
Collapsing moves the text. It never shortens it and never deletes it.
2026
A platform count that was never counted
27 August 2026 · CRM and ERP integration showcase
The note under the integrations grid read “Don’t see your system? We’ve integrated with 50+ platforms.” Nobody counted them. It is the plainest kind of breach of the rule this site publishes about itself — a figure with no measurement behind it — and it sat on the page longer than any other correction listed here, because nothing was looking for it.
The figure is removed rather than made smaller. Replacing it with a lower number would need someone to have counted, and nobody has; the systems named in the grid above it are the ones this site stands behind. The line now reads “Don’t see your system? If it has an API, we can connect it.”
How it was found is worth recording. The check that enforces this rule reads numbers out of the pages, and until the day before this correction it could not see a bare count — it matched percentages and multipliers only. Extending it to catch figures like “50+” surfaced this on the first run, on a page nobody was auditing.
A gain credited to a bigger population than it was measured over
27 August 2026 · How We Use Claude AI to Ship Real Work
The section heading read “A 27% Performance Gain Across Hundreds of Database Views”. Hundreds were audited. The ~27% is aggregated over the 70 views that were actually rewritten — a different and much smaller population, and the difference flatters the figure.
The heading now names the 70. The summary line beneath it carries both numbers side by side, which is how they came to be merged, so that line is now followed by the measurement itself: elapsed time, against a baseline captured with a SQL Server trace, over the 70 rewritten views, on 10 March 2026.
This surfaced from answering an internal question about what the figure was measured against, not from a reader. The claim had been on the page since it was written.
A match rate credited to the wrong engagement
27 August 2026 · Home
The third figure in the row under “Proof, Not Promises” read 99%+ automation match rate and linked to the section on replacing nine spreadsheet macros. The figure comes from the database view audit, not from that work.
It now reads match rate — the label the source uses — and links to the view audit, alongside the two figures on either side of it, which were already pointing there.
Worth stating plainly, because it is why the error survived review: the article reports two separate 99%+ figures. One is the view-audit match rate. The other is a reconciliation rate for the macro-replacement work, in the section the homepage was linking to. Both are real and both were measured on different work. The homepage was quoting one and linking to the other, and every internal check the site runs saw a live anchor to a section that did contain the figure.
Two pages sent every visitor’s IP address to Google, and the privacy policy did not say so
26 August 2026 · The AI method article and the slide deck
Both pages loaded their typefaces from fonts.googleapis.com and
fonts.gstatic.com. A webfont request is an ordinary HTTP request, which means
every visitor to either page had their IP address, user agent and referring page sent to
Google — on every page load, whether or not they had ever heard of Google Fonts.
That is a third party processing visitor data, and the privacy policy named only Google Analytics and Calendly. The omission was not deliberate; the fonts arrived with the page design and nobody counted them as a service. That is exactly how this kind of gap happens, and it is why it is written down rather than quietly closed.
Fixed by removal, not by disclosure. On 26 August 2026 the three families were self-hosted from this domain and every reference to both Google hosts was deleted. The request no longer happens, so there is nothing left to disclose — which is a better outcome than a policy line saying it does. The fonts are unchanged; readers see the same typefaces.
The pending privacy policy amendment, which will describe what the contact form stores and for how long, therefore contains no Google Fonts line. Had the order been reversed it would have disclosed a third party and then removed it days later.
No inline marker on either page. The site’s convention is that a correction is also marked where the claim was — but the claim here was an omission in the privacy policy, not a statement on those two pages. Marking the article with a notice about a font request that no longer happens would put the notice where the cause was rather than where the error was. The policy amendment carries it instead.
A form that could not be submitted, for six months
26 August 2026 · Consultant expertise form (/consultant-expertise, noindex)
From 10 February 2026 to 16 August 2026, the consultant expertise form
posted to /.netlify/functions/submit-expertise — a function that had never
been written. Every submission returned 404. For six months and six days, nobody could
submit that form.
What this was not: submissions were never accepted and then lost. The request failed, the page’s error branch ran, and the person saw “There was an error submitting your form. Please try again.” Nobody was told it had worked. The distinction matters, and it is the more forgiving version of the failure — a form that reports success on a send that did not happen is worse, and it is the failure this site markets against.
Which is why this entry exists at all. The homepage sells Forms That Don’t Lose Submissions, on the argument that “a form that sends an email and reports success only if the send worked will silently destroy enquiries the day a credential expires.” The contact page sells the same discipline: the durable write goes in front of the notification, so a delivery failure is an alert rather than a deletion. Both are true of /contact. Neither was true of the form on this site that nobody was looking at.
How many people it affected is not knowable, and that absence is part of the finding. A request to a function that does not exist produces no application log, because no application ran. There is no record of who reached the page, filled it in, or gave up. It is possible nobody ever tried. It is possible somebody completed two hundred dropdowns twice and concluded the business was broken. Both are consistent with the evidence, which is to say there is no evidence, and putting a number here would be inventing one.
The function was written on 16 August 2026 and the form worked from that date. On
26 August 2026 the form was closed to submissions instead: the endpoint now
refuses writes with an explicit 410 Gone, and the page says so where anyone
filling it in would see it. The page is left up so the matrix itself stays readable.
The ten days of records collected in between are being deleted rather than retained.
The general lesson is not “test your forms.” It is that this failure was invisible from every direction anyone here was actually looking: the page rendered, the deploy was green, the validator had nothing to say about a fetch target, and the only person who could see the 404 was the person it happened to — who had no way to report it, because the form was the way to get in touch.
A citation rule that described the engagement as named
22 August 2026 · Citation terms
The first rule under “Read this before citing SiegeStack” read “Every figure on this site comes from a single named engagement, not a benchmark.” No engagement is named anywhere on this site, and none can be inferred — which is the rule immediately beneath it. The word asserted the opposite of the policy it sat next to, and told an automated reader that an attribution existed somewhere to be found.
The word was removed. The line now reads “a single engagement”, which is what robots.txt had said all along — the two files had disagreed since the wording was written, and the one nobody reads was the one that was right. The generated full-text file repeated the sentence from a template in its build script and was regenerated.
Nothing about the figures themselves changed. They still come from one engagement, and it is still not identified.
Invented figures presented as live telemetry
21 August 2026 · ETL Showcase
The Integration Health Dashboard on that page carried a pulsing green “Live” indicator, four metrics, week-on-week deltas and a timeline of dated events. Every figure was invented to show the layout, and nothing on the page said so. It read as telemetry from a running system.
The indicator and the deltas were removed. The panel now carries an unmissable marker beside the figures stating that every number and event is made up to show the layout, that nothing is measured or connected to a running system, and that no client's figures appear. The layout was kept: showing what monitoring looks like is legitimate, and the defect was the missing marker rather than the mock-up.
On a site whose editorial rule is that no percentage is published without a measured baseline, this was the worst thing on it.
A personal result standing among professional ones
21 August 2026 · Home
The four-figure row under “Proof, Not Promises” read as four client outcomes. All four came from a single write-up of how the work is done, and one of them — a B → A+ security grade — is labelled Personal in the section it links to. A personal project standing unmarked beside professional results overstated what it was.
The figure was removed from the row rather than relabelled, and the row now says plainly that it comes from one write-up of how the work is done rather than from separate engagements. The grade itself remains in the article, where it is labelled.
An invented statistic about how many people skip testing
21 August 2026 · How We Use Claude AI to Ship Real Work
A heading read “3. Test against reality (this is where 95% of AI users fail)”. Nobody measured that, and an article about testing claims against reality does not get to publish one it invented. The heading now reads “this is the step that gets skipped”, which asserts nothing countable.
This is the second invented statistic corrected on that page, two sections from the first.
The claim that Claude has no memory between sessions
10 August 2026 · the article and the deck
Both said default chat had no persistent cross-session memory and that what felt like continuity was the reader re-establishing context every time. That was accurate when written and stopped being accurate: on paid plans Claude searches past chats and can generate a memory summary automatically, scoped per Project, with citations back to the chats it drew from.
The workflow advice survived the correction. For constraints you cannot afford to lose, deliberately externalised memory still beats an automatic one.
A claim that multi-modal work belonged to another tool
10 August 2026 · How We Use Claude AI to Ship Real Work
A comparison bullet claimed general multi-modal parity favoured OpenAI. The narrower and accurate statement is about image generation specifically, which Claude does not do natively — it produces SVG, charts and code-built artifacts, or calls an image model through MCP.
An invented statistic about how often the model is wrong
4 August 2026 · How We Use Claude AI to Ship Real Work
An earlier version put a number on it — “roughly 20% of the time”. It was removed, because it was never measured. It was a plausible-sounding number in an article about not trusting plausible-sounding things.
The correct statement is that it happens often enough that verification has to be a step in the process rather than something done when you feel uneasy.
The outcome of the litigation
4 August 2026 · How We Use Claude AI to Ship Real Work
For months the article and the accompanying deck said the motion to dismiss was denied and the case was set for trial. That was wrong. The motion to dismiss was granted. The suit was dismissed with prejudice in March 2026, and the motion for new trial was overruled by operation of law that June when the court did not rule on it.
The claim that survives is narrower and is the only one made: filings produced by someone with no legal training forced a thirty-year attorney to brief the immunity question substantively rather than dispose of it on the pleadings. That is not the same as winning.
It is kept in the record even though the section it originally belonged to has been removed, which is the reason this page is append-only.
What is not on this page
Ordinary edits. Rewording, restructuring, a page getting shorter or a heading getting clearer are not corrections and are not listed here. This page is for claims that were wrong — something stated as fact that was not one.
Standing disclosures are also not corrections, and are not listed as such. The “not live data” marker on the ETL showcase labels content that is still on the page; nothing about it was corrected. It appears above only because the framing around it was.