Our systems
Flag answers where models disagree, or all get it wrong or slant it the same way.
In build.
Students fix the model answers we flag, back each fix with a line from the source and write the test a good answer must pass. A professor rechecks every item. The result is training data for framing, bias and writing, not just math and code.
Become a design partner → Who does what →
In build No item has been reviewed by a person yet. No customers. 16 November 2026.
Why it’s here Two models disagree about what Monroe threatened.
Question
What did Monroe’s message of 2 December 1823 tell European powers?
Answer A
noted by the student: To protect the young republics of Latin Americanote 1, Monroe struck by the student: threatened war corrected to: said the United States would see it as unfriendlynote 2 if any European power interfered with them.
Note
Framed as protection. The message cites American interests.
“a principle in which the rights and interests of the United States are involved”
Seventh Annual Message, 2 December 1823, paragraph 1
Correction
He warned. He did not threaten war.
“in any other light than as the manifestation of an unfriendly disposition toward the United States”
Same message, paragraph 2
The test
Math and code have tests that pass or fail. How a model frames an event, what it leaves out and how it writes are graded by opinion, so models learn to sound right.
“[R]eward signals that make models good at code actively flatten voice and style.”
Best Grounding v2 score for staying inside a given document.FACTS Leaderboard, December 2025.
How often EQ-Bench’s automated judge agreed with expert writers.Surge AI, February 2026.
Court decisions where a filing cited a real source for something it did not say.Charlotin, 2 November 2026.
Flag answers where models disagree, or all get it wrong or slant it the same way.
In build.
Correct what is wrong, note how it is framed, back each point with a line from a source, write the test, and judge the models’ second tries.
First pass starts in late November.
Rechecks each finished item, one at a time, approves it or sends it back, and pushes what passed.
First session planned for December.
Code checks every quote, and nothing ships until a professor approves it. Full method →
For fine-tuning, preference training and evaluation. Nothing has shipped yet.
Each correction and note, on the exact words.
The source line, quoted word for word, checked by code.
What a good answer must say and must never say.
Each model’s rewrite, and a student’s judgment of it.
The professor’s approval, and who did what, when.
In build: the first pass in late November, the first session with a professor in December.
In build. Until then we pick items by hand.
No item reviewed by a person or pushed yet. Nothing is marked verified.
Runs today: 26,351 stored quotes re-checked against their sources, zero failures, on 4 November 2026.
No customers, no pilots, no data sold.
Anomaly began as a news feed that listed which outlets covered a story and how their framing differed. The engine behind it ran in production and read more than 430,000 items before its pipeline stopped in July 2026. It was invite-only and never found a real audience. What it had was a rule: nothing goes on screen that cannot be traced to a stored source.
We built Lineage, pointing the same rule at history: a provenance engine for American foreign policy, checked against public archives, among them the Foreign Relations of the United States series, the National Archives, govinfo, the Miller Center, the Avalon Project and Chronicling America. In the development database on 29 October 2026: 289 public-domain documents from those archives, and, across the whole development corpus, which also includes copyrighted collections stored so quotes can be checked and never redistributed, 19,647 claims, 1,555 events and 6,335 relations between sources, 86 of them contradictions, each carrying its own checked quote. Nothing we ship carries more than a quote of at most forty words. That day the engine re-checked all 26,351 stored quotes against their sources with zero failures, and 1,300 backend tests passed.
We paused the consumer feed and turned the engine toward the people who train and evaluate models. When we can fund it, the feed returns as open data for anyone.
We put people at the center. The plan: our systems flag model answers, students turn each one into corrections, proof and a test, and a professor rechecks the result and pushes it. The data is for teaching models journalism, framing and writing, not just math and code.
No customer, no pilot. No item has been marked, reviewed or pushed, so nothing is marked verified. The flags and the desk are in build. This entry changes when any of that does.
The first pass at the desk in late November and the first session with a professor in December. Then the first approved items, and two design partners who try them in their own training or evaluation and tell us where they fall short.