What LLM Trading Agents Actually Do in Production: A Six-Month, Population-Scale Record from Two Fleets
Paper • 2609.05663 • Published • 22
observation_id string | measurement string | evidence_class string | event_day float64 | token_label string | window_minutes float64 | buying_vaults float64 | active_vaults float64 | sell_count float64 | median_inter_agent_gap_seconds float64 | cascade_count float64 | minimum_selling_vaults float64 | trade_share_percent float64 | minimum_buys float64 | minimum_sells float64 | same_token_required bool | communication_state string | source_locator string | interpretation string | evidence_boundary string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
feet-one-hour-buy-cascade | buy-cascade | historical-live-deployment | 3 | FEET | 60 | 1,544 | 3,454 | null | null | null | null | null | null | null | null | The vaults did not communicate with each other. | Production Behavior Under a Frozen Harness, page 9 | Shared market state was sufficient for a concentrated buy event in the reported deployment. | One bounded 21-day deployment; no single-cause, prevalence, alpha, profitability, universal-safety, or transfer claim. |
poopcoin-largest-sell-cascade | largest-reported-sell-cascade | historical-live-deployment | null | POOPCOIN | null | null | null | 438 | 9.5 | null | null | null | null | null | null | null | Production Behavior Under a Frozen Harness, page 10 | The row describes the largest sell cascade reported for this token in the deployment. | One bounded 21-day deployment; no single-cause, prevalence, alpha, profitability, universal-safety, or transfer claim. |
tournament-sell-cascade-count | sell-cascade-count | historical-live-deployment | null | null | 10 | null | null | null | null | 3,878 | 10 | null | null | null | true | null | Production Behavior Under a Frozen Harness, page 10 | The count uses the paper's registered sell-cascade definition across the tournament. | One bounded 21-day deployment; no single-cause, prevalence, alpha, profitability, universal-safety, or transfer claim. |
two-sided-token-window-share | trade-share-in-two-sided-token-windows | historical-live-deployment | null | null | 5 | null | null | null | null | null | null | 92.9 | 1 | 1 | true | null | Production Behavior Under a Frozen Harness, page 10 | The shared model produced both buy and sell flow when mandates, positions, and settings differed. | One bounded 21-day deployment; no single-cause, prevalence, alpha, profitability, universal-safety, or transfer claim. |
DX Terminal Pro research aggregates
Nine published research records from DX Research Group's Terminal Pro work. These are small aggregate evidence tables. They contain no participant-level decision logs or training trajectories.
The two configurations have different units of analysis:
| Configuration | Records | What a row represents |
|---|---|---|
market_behavior |
4 | A reported event or token-window aggregate from the bounded 21-day real-capital Terminal Pro deployment |
mandate_compilation |
5 | A reported pre-launch comparison or historical live setting gradient |
The market-behavior rows describe the historical Base deployment. The mandate table contains three controlled pre-launch examples and two historical live gradients. Do not pool these as independent trials from one experiment. The earlier 2025 DX Terminal simulation is not included.
Load the tables
from datasets import load_dataset
market = load_dataset(
"DXRG/dx-terminal-pro-research-aggregates",
"market_behavior",
split="observations",
)
controls = load_dataset(
"DXRG/dx-terminal-pro-research-aggregates",
"mandate_compilation",
split="observations",
)
Sources and provenance
Primary paper: Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital, arXiv:2604.26091v1. The subsequent two-fleet paper supplies the broader research context; these two configurations are Terminal Pro material.
The market CSV preserves the four records from DXRG's published observations, version 1.0.0. The control CSV preserves all five records and original columns from the mandate-compilation evidence table, version 1.0.0, and adds measure_scope, comparison_sample_size and sample_size_note to clarify denominators. The original metadata objects are preserved in source_context/.
This release republishes reported measurements. It does not rerun the underlying analyses or independently validate the paper's findings. Dataset release version: 1.0.0.
Reading the measurements
- The fabricated-rule comparison measures sell decisions in the affected pre-launch populations. Its before/after percentages describe a combined intervention, not an isolated memory-label effect.
- The paper's aggregate diagnosis of 4,900 reasoning traces is not the denominator for every comparison. Exact per-arm counts are not supplied for these records; missing sample-size values remain null.
- Trading Activity is a percentage of invocations that produce trade actions. Trade Size is a share of available ETH used per trade. Their similar percentage notation does not make them the same measure.
- The two-sided-flow percentage is trade-weighted and uses five-minute windows for the same token. A two-sided window contains at least one buy and one sell.
- A sell cascade requires at least ten vaults selling the same token within ten minutes. The largest reported cascade and the total cascade count are different aggregates.
- Missing numeric cells mean a field is not applicable or is not supplied in that source record. They must not be converted to zero.
The source locators, definitions and evidence-boundary columns accompany the values. No new causal effect, statistical significance, prevalence estimate or performance result is inferred by this packaging.
Scope and exclusions
No wallet, vault, user, account, agent, order or trade identifiers are included. User strategies, chat, model prompts and reasoning, per-participant returns and individual transaction records are excluded. The separate internal EVM construction evaluation and internal model-selection screen are also excluded.
The data describes a bounded historical experiment. It is not evidence of current DXAP Alpha performance, a general trading advantage, safety, or transfer to other venues. The published record includes limitations and negative findings; readers should use the linked papers for their full context.
License and citation
License: CC BY 4.0, consistent with the associated papers. Attribute DX Research Group and the source paper when reusing the tables. Cite the original frozen paper version for a reported result and this dataset version for the packaged tables.
- Downloads last month
- 53