How the record is made

Korvaax is a running record of artificial intelligence, kept month by month from September 2026 and researched backward to March 2016. It is written for adults who live with the technology but do not build it. The fuller account of what it is and why it exists is on the about page.

Mohit Prasher is the editor. He sets the editorial standard, reviews every entry before publication, and signs off every publishing wave. A named human is accountable for every entry on this site, and that human is him.

An entry passes through five stages before it appears on the site.

1. Research. A dated fact sheet is compiled from sources for the span being written: the event, the day, the people and organizations involved, the primary documents.
2. Drafting. The entry is drafted against the editorial standard: a sentence-case headline that states a fact, a summary of what happened, one grounded sentence on why it mattered, and the sources underneath.
3. Fact-checking. Every claim is checked against its cited sources before publication. The record is re-verified in waves; where a claim cannot be verified, the entry is corrected, held back as a draft, or omitted.
4. The build gate. Publication is mechanical as well as editorial. The site build rejects an entry with bad structure, and an automated check enforces the headline and summary length limits, the banned hype words, and the sourcing minimums. A page that fails these checks does not publish.
5. Sign-off. Nothing reaches the site without the editor reading the entry, following its sources, and deciding it is right.

Every entry carries at least one primary source: the paper, the official announcement, the filing, the transcript, the court’s own opinion. Where a reputable secondary account exists, it is cited alongside. Where no primary document exists, two independent secondary sources are the minimum. No source, no entry.

• Dates are given to the day where the day is known, and marked as approximate where it is not.
• Names, versions and numbers are given exactly. A model is called what its maker called it.
• Claims are attributed. The record reports what was claimed and who claimed it, then what was disputed and who disputed it.
• A forecast is never written as a fact. The record does not predict, rank companies, or score models against each other.
• Jargon is defined on first use or linked to the glossary.

Research and drafting are done with AI agents working under the editor’s direction. They compile dated fact sheets from sources, draft entries against the editorial standard, check links, and validate the structure of each file.

Every fact is checked against cited sources before publication, whether the first draft was written by a person or a machine. Fact-checking runs in waves across the whole record, and the corrections those waves produce are public and dated.

The editor signs off on every publishing wave. Where a model wrote the first draft of a sentence, the responsibility for that sentence is still the editor’s.

This is stated plainly because the record asks to be trusted. The machines do the searching, fetching and drafting. The judgment about what is true stays with the named human.

Corrections are public and permanent. When an entry is wrong, the text is fixed and a dated line is added at the foot of that entry saying what changed. Entries are never quietly rewritten and never deleted.

The corrections log