The rules of ours that changed, and why, grouped by what each one governs. The oldest here dates from January 2026 and the newest from September 2026.
A product whose maker has no outside coverage is scored on the evidence that exists, with the absence documented and never compensated in another factor.
The retired protection for makers with no press coverage gets no replacement; the absence is scored honestly.
A Confidence score is the exact sum of its five factors, and nothing adjusts it afterwards. A penalty that used to subtract from the total is now answered inside the factor it belongs to.
Two of our scoring rules disagreed about whether anything may be subtracted after the five factors are added. The sum wins.
Below a Confidence Score of 35 we publish no verdict on a product. That floor decided real rejections and appeared on no page, and it is not the same rule as the 60 mark for the Labs Verified badge.
The Confidence floor of 35 becomes a published gate, and is separated from the 60 mark that governs the Labs Verified badge
The rule that chooses which verified products appear at the foot of a lab note
Relevance now outranks filling the row, so a two-product match on feel is no longer discarded for a wider price match
A product we rejected in December was researched again from scratch in January, because the input our discovery pass reads listed only what was already live.
A rejection now persists into the discovery input, so a product we have already judged is never researched again by accident
Tagging a product for a relationship now requires it to be a number one pick for that relationship, not simply a product that could work for it.
A relationship tag now requires the product to be a number one pick for that relationship, not merely a plausible one.
A page may not assert more than its evidence carries, and arrangement is an assertion: what a surface derives, groups, positions, defaults or withholds is a claim no field guarantees.
What a page may assert beyond what it can trace, and arrangement now counts as an assertion.
A correction reaches this record only where a reader could have met the wrong version, judged on how long it was live and where it sat, with unsure counting as yes.
A fix nobody could have read is fixed and not published, because every entry here spends this record's signal.
Our ban on example copy now follows authorship, so it binds every file we publish for machines and every sentence our own templates assemble at request time.
The no examples rule now covers what we publish, not only what our own writers read.
A sentence saying we improve our system every day was written for a live page. We will not publish promises about frequency, because one quiet fortnight turns the page into a liar.
No page of ours may state how often we do anything; the dated record is the proof instead
A correction is not published complete until its severity judgment is recorded. That check now runs where the correction is written, not where the judgment happens to be stored.
The severity judgment on a correction becomes a check that runs where the correction is written
Every build is now checked against the design decisions already taken, not only against its own source files, and a ruling that a build contradicts stops that version.
Design rulings gained an enforcement surface, so a later build cannot reverse one silently.
A written design value is now judged by what it renders as, and every value a design note sets must name the thing it has to agree with and be read against it.
Five page defects proved to be one mistake, so design notes gained two reading rules.
A check that cannot open until a person supplies material may not open until that material is saved word for word, before the check runs its first step.
A check now keeps the evidence it judged from, and its verdict names the file.
A record copied out of live code is regenerated whole or marked suspect, never repaired in part, and the trigger that says when lives with the work that fires it.
Copies of our own code are re-derived rather than verified, and a stale one is marked suspect.
A dry run headline counted fields while our expectation counted values, and the two disagree exactly when a group holds more than one change. Only a per value fingerprint separates benign from broken.
Import verification now proves every changed value by fingerprint
Our short answers had rules for whether a page was true and none for whether it could be read. One went out correct, fully sourced, and shaped like a wall.
A short's visual form is now decided from that short's own evidence, written down with reasons, and read back. Nothing forced.
Two of our short answers opened with a fragment that meant nothing without the question above it. The check built to catch that had measured it, printed the result, and passed the page anyway.
A short answer's summary box must now stand as a sentence on its own, and may not be more certain than the page beneath it
Where a short answer's body starts is decided before it is written, from the moment the reader is in rather than the order of our findings, with the openings not chosen named.
Where a short's body starts is a decision recorded before the writing, taken from the reader's moment and not from our evidence's order
A short's question is minted only with a record of a real person asking it, or with our judgment written down as one; a figure a short draws from our catalog is counted over the whole live catalog.
A short is minted only with a record of a person asking it, and its catalog figures are counted over the whole live catalog
We made a required citation field optional after finding that products without a confirmed study had no honest way to fill it, and the improvised filling was converging on identical wording.
A required field is optional again, because requiring it was producing a template
How stricter verification changed one product's score from 9.2 to 8.4
System validation: lower score reflects improved methodology