A structured library of LLM-generated research on systematic intraday futures trading (ES/NQ focus): data integrity, market microstructure, ML/RL methods, strategy validation, and research governance. 50 research prompts were run against frontier LLMs (49 have at least one raw report); the outputs were reconciled, decomposed into atomic claims, and organized for retrieval by both humans and LLM agents.
- Everything here was written by AI. No claim, citation, or number has been independently verified unless its ledger row explicitly says so. Most rows are marked
NOT_CHECKED/SOURCE_REPORTED— that labeling is deliberate and honest. - Verify before you rely. Citations may be imprecise or wrong; summaries may misstate what a cited author actually wrote. Credit for underlying ideas belongs to the cited authors; errors in the summaries are ours (or the models').
- Not financial advice. Nothing here establishes that any strategy is profitable. The library's own standing verdict on profitability claims is
HOLD/BLOCK.
The human-readable core is the seven synthesis papers in library/60_papers/drafts/ — long-form documents, one per research area:
| Paper | What it covers |
|---|---|
| Data and market inputs | What data exists beyond OHLCV bars, and what it takes to trust it |
| Targets, labels, and rewards | How to define what a model should predict or optimize |
| Models, RL, and memory | Supervised, sequence, and reinforcement-learning methods and when each is justified |
| Strategy discovery, validation, and governance | Finding edges and proving they're real (multiple testing, execution parity, red-teaming) |
| ML algorithm landscape | Broad survey of ML algorithm families and their applicability to intraday futures |
| Data certification methods | A candidate procedure for certifying a dataset before research use |
| ML research practice and systems | Research workflow, engineering, and capability curriculum |
Want the receipts behind a statement? The evidence tables in library/40_atomic_knowledge/ hold the atomic claims, the contradictions between model runs, and every citation with its verification state — GitHub renders the CSVs as sortable tables.
For topic-by-topic browsing, start at library/README.md and follow its topic routes.
Point your agent at library/00_control/SESSION_PREAMBLE.md and let it follow the routing (pointers → topic MOCs → filtered evidence ledgers). See llms.txt.
| Path | Contents |
|---|---|
library/ |
The governed library: routing, topic maps, synthesis papers, and CSV ledgers of claims, contradictions, citations, and numerics — each row carrying an evidence state |
arena_ai_prompts/ |
The 50 research prompts and the raw LLM reports they produced |
synthesis/ |
Early integrated synthesis (pre-library, kept for history) |
source_corpus/ |
Not included — see its README; original third-party documents stay private, citations reference public sources |
Every knowledge object has a stable ID, version, evidence state, and evidence tier. Raw LLM reports are Tier F inputs — agreement between two model runs measures coverage, not truth. The claim/contradiction/citation ledgers under library/40_atomic_knowledge/ are the machine-readable core; the papers are narrative views over them. If you find an error, an issue or PR correcting a specific ledger row (with a source) is the most useful contribution.
Content is licensed CC BY 4.0. Cited third-party works remain the property of their authors.