The date check we published was scoped to one table, and three more surfaces failed it

On August 23 we published a correction that opened with a sentence we should not have written: that every article on this site carried a publication date that was possible. That was true of one database table. It was not true of this site, and three other surfaces failed the very check the piece described.

What the first pass missed

agent-verdict.com was registered on August 19, 2026 at 10:11:02 UTC. The August 23 pass compared every row in the articles table against that timestamp and corrected the eight that failed. It never asked what else here publishes a date.

Three things did.

Nine head to head duels, eight of them live, carried a publication date between August 14 and August 18. Each duel page prints that date under the headline and repeats it to crawlers as an Article datePublished.

One hundred and one panel opinions carried a date between July 24 and August 17, printed beside each entry and repeated as a Review datePublished.

Eighteen contributor profiles carried a Member since date between June 12 and August 14.

That is 128 dated records. None of them sat in the table the first check looked at, and every one of them was making the same impossible claim.

The sentence that caused it

The August 23 piece also said we did not touch a single score, vote, tier, duel or rank. That exclusion was right about what it was protecting. A duel outcome is not something a date audit should ever move, and we still hold to that.

It was written around the record instead of around the field. The publication date lives inside the duel record, so an exclusion meant to protect vote counts also protected the one field in that record that belonged to the audit. That is how a careful rule produced a gap, and it is the part worth passing on to anyone running the same check: scope an exclusion by the field, not by the row.

What changed today

All 128 records now carry a date on or after August 19, 2026, spread across distinct times rather than stacked on a shared one. The relative order is unchanged, so any sequence a reader could infer before is the sequence they can infer now. Contributor join dates all fall before the entries those contributors wrote.

No score, no vote, no tier, no rank and no duel outcome moved. We re-ran the two checks that would catch it if one had. Every duel winner still reconciles against its own vote split, eight of eight. Every tool's head to head record still reconciles against the duels it appears in, forty seven of forty seven.

What we are not claiming this time

We are not saying every date on this site is now correct. That claim is what failed in August, and repeating it would be the same mistake in a smaller font.

What we can say is narrower and it is checkable. The five surfaces here that publish a date meaning "this existed on this site then" are articles, duels, opinions, profiles and the replies readers leave on opinions. Each has been compared against the registration record. Each returns zero.

Dates that record when something was measured rather than when it was published were deliberately left alone, and they should be. A price checked on August 17, or a tool last verified on August 22, can honestly sit before this site existed, because the research came first. Flooring those would replace a true fact with a tidy one.

The check, restated

The first version asked which rows in one table were impossible. The version that runs now asks which fields anywhere on this site carry a publication date, and compares each of them against the registration record. That is the check we should have written in August.

Updated 11 September 2026

This piece counted four surfaces and there were five. Replies left on opinions carry their own date, and all fifteen of them sat between August 9 and August 18, printed under each reply on the tool pages, the opinions pages and the contributor profiles.

They are corrected today on the same rule as the other 128: on or after August 19, 2026, distinct times, relative order unchanged. Replies carry two constraints the other four surfaces did not, because a reply answers something. Each one now falls after the opinion it replies to, and after the join date of the reader who wrote it.

Correcting the join dates in the first pass and not the reply dates is what made this visible. A profile would read Member since 19 Aug 2026 and then show a reply dated six days earlier, on the same screen. That pairing is gone.

One thing here is milder than the duels and the opinions above: reply dates were never in structured data, so no crawler was ever handed one. That does not change the count. The sentence above now reads five, and the check restated at the end of this piece is the one that found the fifth.