Corrections

This is the log. It is appended to and never rewritten. Entries stay up after they are resolved, including the ones that are embarrassing.

If you find an error, email contact@therezaali.com with the page and the claim. Corrections go to the front of my inbox.

2026-08-06: I retitled a published article, and nine new ones failed their first check

Two entries in one, because they are the same story.

The postmortem was retitled. It went out as “The bot that published 900 articles under my name” and is now titled after the finding rather than the incident. The URL is unchanged, so nothing broke, and not a fact, figure, quotation or citation in it moved. What changed is the order: the general defect now leads and the incident is the evidence for it. I am logging a presentation change because on a site that publishes its corrections, quietly restyling a live page is exactly the thing a reader would want to know about.

Nine articles nearly shipped with seven false claims. They were drafted from my own posts, fact-checked, then given to verifiers that fetched every primary source independently. The verifiers found seven blocking errors, and they are worth naming because each one is a way a confident sentence goes wrong:

• I published that a 2.8-trillion-parameter model at 4 bits is a 1,400 GB download. The real checkpoint is 1,560.9 GB. The model is not uniformly 4-bit: attention, shared experts, the dense MLP and the vision tower are exempted in its own config, and I had multiplied as though they were not. The article had invited readers to check the arithmetic with a calculator.
• An Elo rating and a full chip-design specification, both stated as fact, neither traceable to any source I could produce.
• An article whose entire argument is that quoted numbers drift from what was measured opened by dividing two numbers and getting the wrong answer, because the denominator it used was not the one the sentence named.
• The same article accused Microsoft of omitting a methodology detail that Microsoft states in the very next sentence, and missed the caveat that actually mattered.
• A cited study was reported backwards: three of the five signals it lists as making writing read as machine-generated were described as making it read as human. The paragraph built on top of that inherited the inversion.
• A claim that “several outlets” described a licence a certain way. Five candidate outlets were checked. None of them did.

Every one was caught before publication, by a machine instructed to disbelieve me. That is the process working rather than a story about my carefulness: I wrote those sentences and I believed them.

2026-08-06: My legal name was stated wrong

The about page said my legal name is Mohammad RezaAli. It is not. That is how I introduce myself formally. My legal name is Mohammad Mahdi RezaAli, which is what appears on the Seneca Polytechnic certificate linked from the certifications page.

It mattered because the certificates were already published here and two of them carry a form of the name that the site did not acknowledge. Anyone comparing a certificate against this site had to decide for themselves whether it belonged to the same person, on a page whose whole argument is that its claims can be checked.

There is now a short section on the about page setting out all three forms and what each one is for, and the structured data carries the legal and formal versions as alternate names so a machine can resolve them too. A build check refuses to compile the site if a fourth form appears anywhere.

2026-08-06: Wrong employment dates, and three things that were missing

Corrected the day after the rebuild went up, from my own records. Nobody reported these; I checked the page against the record and it did not match.

Jumbula was September 2019 to March 2026, not 2019 to 2024. The 2024 end date was wrong by nearly two years and appeared in eleven places: the about page in both languages, the Jumbula page, the disclosure page, the Yakkyo and Yakkyofy pages, the corrections entry above, an article, and the homepage credentials block. All eleven now read September 2019 to March 2026. Six years and seven months.

Yakkyo started in March 2026, not 2024. The hasOccupation start date in the structured data said 2024, which would have had me at two full-time jobs simultaneously for two years.

Cafe Bazaar was missing entirely. Android Developer, 2013 to 2016, remote. It was left off the rebuild because the site it replaced showed the logo with no role and no dates anywhere, and I was not willing to guess at a job title. It is on the record now, and it is the part of the record that explains the rest: I wrote software before I marketed it.

Two Seneca Polytechnic programmes and the Messina degree were missing their dates. English Language and Literature, January 2022 to January 2023. Computer Programming, January 2023 to April 2024. BSc in Computer Science (Data Analysis) at Messina, September 2025, in progress and stated as in progress.

None of this was invented and none of it was rounded. Rule 2 of the editorial policy says a date on this site is a date I can produce a record for, and an approximate date is worth less than no date, because it looks equally confident.

2026-08-05: Site-wide retraction and rebuild

Everything published on this domain by the automated pipeline has been deleted. Three separate failures, all mine.

1. 739 machine-generated English articles, 2,922 files, all removed

From 31 October 2025 to 27 April 2026, an hourly Cloudflare Worker ran unsupervised on this domain. It scraped five RSS feeds, passed 450-character excerpts to gpt-4o-mini, and committed the resulting posts straight to the main branch. No person read them before they went live. That is 842 Add post: commits. Publishing stopped on 27 April 2026; the archive stayed up until I deleted it on 5 August 2026.

What survived to the end was 739 English articles. Each was machine-translated into Italian, Arabic and Simplified Chinese, which is how 739 articles became 2,922 files.

The quality gate made things worse rather than better. It required every post to contain at least one number and at least two named entities. The model was working from a 450-character excerpt, which rarely contains a usable statistic, so the requirement did not filter out invention. It demanded it. It only ever read the model’s own output, so the model could always satisfy it by writing the number down.

Every count below is over the 739 English articles, case-insensitive, each file counted once:

• 180 of the 739 English posts cited a source called “Internal Analysis” that does not exist. Pattern Source:.*Internal. No such analysis exists. It never existed. It is a phrase the model produced to satisfy the citation requirement.
• 470 of the 739 carried an invented before-and-after metrics table. A conversion rate, a retention figure, a company that is never named. None of those tables came from anywhere.
• 91 of the 739 credited a figure to Forrester with no link and no report title. 46 did the same to HubSpot. Neither firm published what was attributed to it, and no article said which report it meant, so there was nothing for a reader to go and check.
• 158 of the 739 shipped with a title that stops mid-sentence in an ellipsis. Nothing in the pipeline looked at the output, so a truncated headline went live and stayed live.
• A HubSpot writer’s first-person account was republished under my byline. Someone else’s personal experience, rewritten in the first person and signed with my name. That is the worst item on this page and it is the reason for rule 8 of the editorial policy.

The full account, with the worker source, the validator code and the derivation of every count above, is the postmortem. This page is the dated log; that page is the evidence.

All 2,922 files are deleted. I am not selectively rescuing the ones that might have been fine, because I have no way to tell which those are, and neither did you.

2. The online-registration-software benchmark, removed

The benchmark previously published on this domain has been taken down for two independent reasons.

Fabricated ratings in structured data. It emitted AggregateRating values for third-party products. Those numbers were not derived from any review process, any panel, or any sample. They were presented to search engines as aggregate ratings, which is a machine-readable false statement about other companies’ products.

An undisclosed conflict. The benchmark ranked Jumbula. I worked at Jumbula from September 2019 to March 2026. The benchmark did not say so anywhere on the page. Full account on the disclosure page.

The benchmark has not been corrected and republished. It has been withdrawn.

3. Retracted metrics survived in three languages for 150 days

Performance metrics were retracted from the English About page. The Italian, Arabic, and Simplified Chinese versions of the same page kept serving them for a further 150 days.

The cause was procedural: the translation pipeline pushed English changes outward but a retraction in English did not trigger a re-check of the pages already translated. The result was the worst available outcome: a site that could be shown, side by side, contradicting itself, with the retracted version still serving to the readers least able to find the corrected one.

Those pages have been removed along with everything else.

What replaced it

The rules are on the editorial policy page. The short version: source URLs live in front matter, the build fetches every one of them and fails if any stops returning HTTP 200, every article names its own limits and its confidence level, no automated job can write to the published branch, and no number about my own work gets published at all.

Fewer pages. Every one of them checkable.

Logged 2026-08-05.

2026-08-05: Counts in the entry above re-derived and pinned to one definition, before publication

Caught during the pre-launch review of the rebuild, and logged here rather than fixed silently, because this page claims entries are appended and never rewritten and I have just rewritten one.

• The internal-source count went out on five pages as a lower figure than the one that stands, with no definition attached and three other numbers for the same thing elsewhere in the draft. A reader had no way to reconcile them, which is the failure this site exists to correct, reproduced in miniature. That first count came from a narrower pattern and is superseded; it is not repeated here, because a superseded figure printed next to the correct one is one more number for a reader to reconcile. The count that stands is Source:.*Internal, matched case-insensitively over the 739 English files as they existed at the commit before the deletion, each file counted once: 180. Every page now says “180 of the 739 English posts” and names the pattern, and the postmortem records the command that produces it.
• “The same fabricated conversion-rate table appeared twice” understated it by two orders of magnitude. 470 of the 739 English articles carry an invented before-and-after metrics table. “Twice” made the failure look smaller than it was, which is the wrong direction for an error on this page to run in.
• The Forrester and HubSpot attributions were described loosely rather than counted. 91 of the 739 name Forrester as a source with no link and no report title; 46 name HubSpot the same way. Those are the numbers, and they replace every adjective that stood in for them.

Two related wordings were fixed at the same time: this site’s 2,922 was a file count, not an article count, and pages that called them “2,922 posts” have been changed to “739 English articles, 2,922 files”; and pages saying the pipeline ran “until August 2026” now give the real window, 31 October 2025 to 27 April 2026, with August 2026 as the deletion date. The two dates are four months apart and conflating them made the pipeline look like it was running until the day I pulled it.

One exception to the deletion, also found in that review: a single LinkedIn post from the old pipeline, content/post/linkedin/2025-11-20-post1.md, had survived the deletion commit that removed everything else. It carried the misspelled byline and a banned word. It has now been deleted too, so the claim above, that nothing was selectively rescued, is true as written.

Logged 2026-08-05.

2026-08-05: Three claims about myself, withdrawn

These are separate from the pipeline. I wrote all three by hand.

• A Coursera certificate was described as “graduate-level”. The certificate says “an online non-credit course”, which is close to the opposite claim, and the certificate is the document that decides. Every certificate on the certifications page now uses the issuer’s own wording, and the page states in its own first section what a certificate of that kind does not prove.
• My employer was written as “Yakkyo and Yakkyofy”, in the prose and in the worksFor structured data. Yakkyofy is a product of Yakkyo S.p.A. There is one employer and one job. Two names in that field read as two roles at two companies, which inflates a CV without adding a fact to it. Corrected on the Yakkyo page and the Yakkyofy page.
• This domain carried third-person “reference profile” pages about my employer and its product. They were written in a detached institutional register, on my personal domain, with no authorisation from Yakkyo and no disclaimer. Nobody at Yakkyo asked for them or saw them. Both pages are now in the first person and open with a disclaimer saying they are not company communications.

Logged 2026-08-05.