Knowledge
Trading Brain turns a large body of trading literature into a structured, attributed and searchable system that research and AI agents use directly.
Trading Brain
The basic unit is the atom: one note, one claim, one author, one source. Atoms link to concepts, which form a knowledge graph. Above them sit author frameworks, mechanical setups and execution states describing when a setup applies.
| Layer | Count |
|---|---|
| Atoms | about 145,500 |
| Concepts | about 2,200 |
| Frameworks | 50 |
| Setups (trading legs) | 161 |
| Execution states | 45 |
| Author profiles | 384 |
| Edge notes | 17 |
How the knowledge is used
- In research. Setups and claims from the vault become pre-registered tests. A programme of 217 claims and a round of 15 setups from 7 authors and groups both started there.
- In the running system. Execution states reach the market context expert as a versioned, hashed digest.
- In daily work. Agents and the operator search it with metadata filters for author, domain and status.
From a source to a running AI expert
From claim to validated result
These categories are never merged.
| Category | Meaning | Proven in the market? |
|---|---|---|
| Attributed claim | An author said it | No |
| Canonical note | Faithful to its source and quality-checked | No. Canonical means faithful, not true. |
| Research finding | NOVAQ measured it with a stated method | Within the study's limits |
| Validated edge | Passed in-house validation when tested | Tested, not guaranteed, subject to review |
Author numbers and NOVAQ measurements are stored separately. When they agree, that is a reproduction, not a merged statistic.
Provenance
Each important note records its origin, and video corpora record the video, the timestamp and the original quote. Publication dates are checked against content: NOVAQ found videos dated 2017 in which the speaker says the year is 2012. Era is now determined from content, and missing evidence is recorded as unclear.
Search
The vault is indexed for semantic search by a locally run embedding model, with 148,442 notes and about 239,000 passages. An average query at depth 40 takes about one second. An earlier benchmark passed 100 of 100 golden queries, before the Quant Sources expansion. Atomic notes suit retrieval naturally, since one claim is one retrievable unit and metadata gives reliable filters.
Quant Sources
Quant Sources is a corpus built from public materials of quantitative firms and tool vendors: 294 organisations and 181,615 collected documents. Models triaged the documents, a curated core was selected and atoms were extracted with quotes verified against the source.
| Measure | Value |
|---|---|
| Atoms imported (set A) | 42,317 |
| Atoms imported (set B) | 79,914 |
| Sources (sets A + B) | 16,384 |
| Quotes found in source | 98.5% (96.5% exact) |
| Confirmed serious errors in the audit sample | 0.2%, or 0.5% with borderline cases (of 2,075 atoms) |
The sample audit covered set A and used the same model as the extraction, which is a stated limit. The second set, B with transferable methodology (portfolio, risk, validation, machine learning, execution), has been in the vault since 3 Nov 2026: 79,914 atoms from 12,517 sources, 43,601 of them canonical under the same rule. Including an organisation's public material implies no relationship with it.
Growth
| Date | Atoms |
|---|---|
| 14 May 2026, first extraction | 198 |
| 20 May 2026 | about 9,200 |
| 27 June 2026 | about 10,500 |
| 17 Aug 2026 | about 19,800 |
| 17 Oct 2026 | 22,187 |
| 1 Nov 2026 | about 64,700 |
| 3 Nov 2026 | about 145,500 |
Atoms in the knowledge vault, May to November 2026