957 reader rows carried a date from before this domain existed

The measurement

The registration record for llm-tier.com reads August 19, 2026, 10:10:53 UTC. On August 23 we published a correction reporting that eight of our fourteen articles carried a publication date earlier than that, and we fixed them.

That correction was accurate. It was also too narrow. It measured the articles table, and it measured nothing else.

Every reader contribution held on this site carried a date from before the domain resolved. Nine hundred and fifty seven rows, across five tables:

• 872 reader axis scores
• 43 contribution votes
• 22 open ratings, which are also emitted as structured data on every builder review page
• 11 note comments
• 9 lab notes

The earliest read August 2, 2026, seventeen days before this domain existed. Between them the 957 rows held forty six distinct values, because each table stacked its entire corpus on one time of day: reader scores at 12:00 UTC, open ratings and votes at 10:00, note comments at 11:00, lab notes at 09:00. A real submission stream does not arrive on the hour.

The dates were crawler visible in two places, and they failed differently in each. Open rating dates were emitted in structured data as Review.datePublished on the builder review pages. Reader score and lab note dates appear only as visible text on the contributor profile pages, with no structured data equivalent at all, so a schema based check cannot see them even in principle.

Why it happened

The same reason as in August, one layer down. The launch corpus was seeded when the site was built, and its dates were written backwards to give the board an apparent history. They were seed values, not a record of anything, in exactly the sense our August note used that phrase.

What is new is not the cause. It is that our own published check was scoped to the table where we first found the problem, so it kept returning zero while the rest of the site stayed wrong.

What we fixed first, and why the order mattered

Earlier today we published, in prose and on every individual item, the fact that this reader data was seeded by this lab at launch rather than submitted by readers. That went live before these dates were touched, and deliberately so.

Moving these dates on their own would have removed the last public trace that the reader board was populated in a single batch, and it would have done that on a site that was not yet saying so. Corrected in that order it would have read as a cleanup. The disclosure had to come first for the date fix to be a repair rather than a tidier surface. Every row re-dated below still carries its Lab seeded label.

What changed

All 957 dates now sit on or after August 19, 2026 at 15:00 UTC, which is after the last contributor account on this site was created. They are spread across thirteen days at 957 distinct values, where there were forty six. None of them falls outside August 2026, which keeps them consistent with the disclosure published this morning.

Reader scores keep the grouping they always had. Each set of axis scores for one product sits minutes apart in a single sitting, rather than every set in the archive sharing one stamp.

The relative order of every table is unchanged. Every comment still falls after the note it answers, and every vote still falls after the thing it voted on.

What did not change

No index score, no axis subscore, no composite, no rank, no weight, no vote count, no upvote total, no run duration and no roster entry moved. No review body, note body or comment body was edited, and no contributor name or byline was touched. We checked the published reader index on all 37 product pages before and after: the sixteen that publish a figure publish the same figure, and the twenty one held below our three reader threshold are still held.

The measurement dates for the benchmark itself were not part of this pass. Research precedes launch, and a date that records when we measured something is not a claim that the site existed yet.

Contributor join dates were already correct and were left alone. They are derived from the account record itself and have always read August 19, 2026.

One side effect worth stating plainly: editing a row updates its last modified stamp, so all 957 rows now record that they were modified today, and the sitemap entries for the fourteen contributor profiles moved with them. That is true rather than cosmetic. The rows were modified today.

The gate, restated so you can run it on us

In August we published this gate: for every published article, assert that its publication date and its last updated date are at or after the registration date of the domain it is served from.

That gate was true. We ran it. It returned zero, correctly, every time, while 957 rows on this site were wrong. A gate scoped to the place you first found a problem will keep passing for as long as the problem lives somewhere else.

The restated version is not cleverer, it is only wider. For every table that stores a date describing something a person did on this site, assert the same comparison. Not the articles table. Every table.

So the scope of this note is five tables, and nothing wider. We have not swept every date field this site holds, and until we have, any assurance covering the rest would be the August mistake at a new scale. Here is the version you can check instead: the five tables listed at the top of this page, at the row counts given, hold no value earlier than 10:10:53 UTC on August 19, 2026. Read the reader dates off any contributor profile and the open rating dates out of the structured data on any builder review page, and compare them against the registration record yourself.

Correction, September 10, 2026

The section above originally described the benchmark measurement stamps as unchanged and gave a single instant for them. That instant was wrong, and this note is the place it should have been caught.

The campaign behind the published figures was not one dated cycle. It ran as five daily repetitions of the full specification, from August 11 to August 15, 2026. The final repetition, 342 of the 1,710 recorded executions, started after the instant we were publishing, so a fifth of the dataset postdated its own measurement stamp. Both halves of that contradiction were already public: the stamp sat in /api/rankings.json and the execution dates sat in /api/runs.json, on the same origin and under the same licence.

measured_at and last_run_at now read 2026-08-15T09:11:37.500Z, the end of the last recorded execution, and the window itself is stated on /methodology and on /data. No score, rank, subscore, price or execution record changed.

The sentence about research preceding launch stands. These are measurement dates, and they legitimately fall before this domain was registered.