ProjectDXAI/continuous-record-llm-trading-agents

What it claims

Live LLM trading agents' behavior is set mainly by the operating layer: prompt contents, sliders and posture settings. What users write in strategy text matters much less. The record covers two fleets:

• DX Terminal Pro: 3,505 user-funded vaults trading real ETH on Base for 21 days.
• DXAP pre-alpha: 500–599 user-created agents on Hyperliquid perpetuals for 69 days.

The authors say the fleet as a whole showed no directional edge, although individual agents, configurations and test periods produced widely different P&Ls.

Evidence

Numbers here that the summary does not give:

• Risk slider: each slider level adds +0.425 to chosen leverage.
• Leaderboard routing: a nine-row gainers leaderboard is rendered in the prompt, but only ranks 1–3 are shown. Entries jump at that render boundary, with a regression-discontinuity ratio of 1.75x [1.49, 2.06].
• Volatility-blind sizing: median leverage is 5.0x in every volatility sextile. Median realized return over the same range falls from −10.6 to −98.2 bps.
• Liquidation concentration: one posture-plus-slider cell holds 11% of positions and 62% of liquidations (Mantel-Haenszel OR 22.37).
• Liquidation-distance prompt: stating the liquidation distance in the prompt text changed nothing. This is a direct negative result for prompt-level risk controls.
• Capture gap: 43.2% of positions reached +300 bps of open profit within 24h. Of those, 49.3% still closed negative, and median profit capture was 2.0%. The exit analysis covers 6,400 closed positions.
• Fixed exits: a fixed 2%/4% bracket beat every discretionary exit measured, by +39.0 bps per position.
• Model replay: in 416 replayed production scenarios, regret intervals overlap for all three frontier models, but flip rates on identical repeats do not. The page does not name the models.

How to use it

• paper.pdf: the paper itself, ending in a 17-rule methodology canon. Three of the program's own results were retracted along the way and appear only as retractions.
• figures/: all eight figures as 300-dpi PNG.
• data/: aggregate extracts behind Figures 1, 2, 3, 5 and 8, the same files as the arXiv ancillary files. data/README.md maps each file to its figure. There are no per-user or per-position rows.
• cards/: five 1600x900 share cards.
• CITATION.cff / CITATION.bib: citation records.
• Two Hugging Face datasets, each with a dataset card and datasets.load_dataset configurations (the page does not give the dataset IDs):
• Terminal Pro research aggregates: four historical market-behavior records and five control/setting records, with source definitions and sample-size limits.
• DXAP historical workflow excerpts: three de-identified July paper-execution cases, sanitized request/response projections, starting states, research-child records, proposed scenario questions, and the complete selection-rank aggregate.
• Discussion guide: separates the findings from the open questions, and is pitched as a starting point for replication, benchmark and agent-control work.
• dxrg.ai/research: further versioned artifacts, including a benchmark card, guardrail matrix, execution-reconciliation matrix, trace-feedback registry and prompt-compilation ablation registry.
• Related paper: arXiv:2604.26091 covers the Terminal Pro reliability engineering.

Licence: paper text, figures and cards are CC BY 4.0. Dataset terms are set separately on each dataset card, so check them before reuse.

Limits

• Aggregates only. Per-position analysis cannot be reproduced from this repo.
• Excerpts are illustrative. They withhold participant identities, prompts, strategy text and trade-specific values. The authors call them "a small explanatory sample, not full reproducible trajectories or a validated trading benchmark."
• Benchmark not released. The broader scenario benchmark and the harness/model-training test collection are still work in progress.
• Pre-alpha only. The DXAP record ends August 15, 2026. The authors say it is not a performance summary of the current DXAP Alpha, which has since changed agent controls, execution, exit handling, market context, evaluation and training. They promise forward results separately, so none of these numbers describe the shipping product.
• Authors run the product. DXRG operates DXAP and calls the current system "strong" without published evidence. Weigh the post-cutoff claims accordingly.