Trading knowledge is abundant and badly organised. Books, courses, videos and papers repeat, contradict and quietly copy each other, and almost none of it can be cited at the level of a single claim. NOVAQ set out to build something closer to a scientific literature for trading: each statement addressable, attributed and connected.
Schema first
On 14 May 2026, before any trading knowledge entered the repository, ten system documents defined what a note was, how it should be attributed, which status it could carry and how notes should link. The first extraction gave 198 atoms. By 3 November the vault held about 145,500.
Why atoms
A book contains hundreds of claims, and a summary hides them. One claim, one author, one source and an explicit confidence level make knowledge addressable. Two authors can be compared on one concept, a claim can become a test, and a discredited source can be traced through all its notes. When semantic search arrived 45 days later, the notes were already close to ideal for retrieval, since one claim is one retrievable unit and metadata gives reliable filters.
A pipeline, not a pile
Raw material enters a capture area and never mixes with canonical notes. Sources are catalogued, atoms are extracted, and concepts form the nodes of a knowledge graph of about 2,200 entries. Above them sit 50 author frameworks, 161 setups, one for each trading leg, and 45 execution states that describe when a setup applies. Each stage has its own format and status, so a reader can tell a raw claim from reviewed work at a glance.
From a source to a running AI expert
Getting dates right
During a large video import, two independent agents flagged material published in 2017 that sounded older. In one series the speaker says the year is 2012. In another, four quoted currency prices fit only one window in autumn 2013. Era is now determined from content, by segment, and one video turned out to join two recordings years apart. For teaching that evolves over time, knowing when something was said changes what it means.
Scaling with control
Quant Sources added 42,317 atoms from 3,867 sources in four days. Models triaged 133,273 documents, quotes were verified against sources, and an audit sampled 2,075 atoms. A model flagged 1.0% as serious, and a second check against sources confirmed 0.2%, or 0.5% with borderline cases. In 98.5% of cases the quote was found in the source. The vault went from about 22,000 to about 64,700 atoms in two weeks, with a measured error rate attached, and a second set of 79,914 atoms followed on 3 November.
Atoms in the knowledge vault, May to November 2026
What "canonical" means
About 25,000 Quant Sources atoms and 8,459 notes from one video corpus were promoted after this audit. The word means faithful to the source, not true of the market. Keeping those two ideas apart is the central discipline of the system.
How the knowledge works
The vault is used directly. Claims and setups become pre-registered tests, execution states reach a running AI expert, and agents and the operator search it with filters for author, domain and status. A knowledge engineering problem and a trading research problem turn out to be the same problem at this scale.
Limits
The sample audit covered the first set and used the same model as the extraction. Structure does not make knowledge true, and the separation of claim from evidence exists for that reason.