DXRG/dxap-research-workflow-excerpts

DXAP historical workflow excerpts and selection aggregates

Three detailed historical case studies: 29 parent/reconciliation event rows, 18 sanitized tool-request/response projections, five starting-state records, two research-child event records and nine proposed scenario questions. It also includes the complete 15-row selection-rank table already published with arXiv:2609.05663v1. These are different views of the same selected material, not independent sample counts to add together.

Start with the full case studies for the initial state, sequence, recovery, identity relationships and limitations of each example. Version 1.1.0 expands the original three excerpts without adding participants or trading histories.

These are two different kinds of evidence. The event excerpts are newly prepared, highly redacted summaries of selected recorded July 2026 tool sequences. The rank table is a published population aggregate. They cannot be joined at row level.

What the excerpts show

Sample	Included behavior
sample-001	A paper order and trigger setup, followed by a memory-capacity error, a subsequent memory update and turn completion
sample-002	Two attempts within one backend turn: trigger authorization errors, deferred completion, and a retry that completes the turn
sample-003	Selected opening and closing turns from a linked paper position, including a research response, trigger initiation and memory cleanup

All order results in these examples are simulated paper fills from a deployed production runner using a mainnet market feed. They are not real-capital executions or offline replay results. They predate the current DXAP Alpha. The source records were chosen for workflow diversity and available reconciliation, not profit or success rate.

The three excerpts cover five source contexts across four backend turns. Sample 2's two attempts are one turn, not two independent trading examples. Sample 3 contains selected endpoints; intervening turns and elapsed time are omitted. sample_notes.json documents these boundaries.

The two child-tool rows expand one research call within sample 3; they are not additional parent turns. case_linkage.json records de-identified reference relationships and the checked terminal state of that case.

Load the data

from datasets import load_dataset

events = load_dataset(
    "DXRG/dxap-research-workflow-excerpts",
    "workflow_events",
    split="examples",
)
ranks = load_dataset(
    "DXRG/dxap-research-workflow-excerpts",
    "selection_ranks",
    split="observations",
)

workflow_events.jsonl is a convenience copy of the CSV, not another sample population.

The additional configurations use split="examples": tool_excerpts (18 rows), case_start_states (5), research_child_events (2), and scenario_questions (9). In the CSV views, nested projections and event-index lists are JSON strings; use json.loads to inspect them. Companion JSONL files preserve those objects and lists directly.

Event schema

Column	Meaning
sample_id	A release-local alias with no stable cross-dataset identity
event_index	Recorded ordering within the excerpt, with reconciliation events added explicitly
segment_index	The selected source turn within a sample; gaps between segments are not measured
attempt_index	Attempt number within that selected backend turn
event_category	Generalized tool or reconciliation category
operation	open or close for order events; otherwise null
transport_status	The recorded transport status for tool calls; null for reconciliation rows
observed_outcome	Result-level outcome or separately recorded backend state
error_category	Broad application-error class, when present
execution_mode	Always paper in this sample
record_period	Month of the historical source records; exact times are withheld

Transport success is not action success. Three included calls have transport status success but a result-level error. A model's finalize_request is also distinct from backend_turn_status=completed. The first attempt in sample 2 does not complete the backend turn.

simulated_fill_recorded requires a returned paper-fill identifier that also occurs in the archived fill reconciliation. updated records a memory-tool result; it does not count distinct durable revisions. returned for research means a response was returned, not that the research was correct. Backend completion is reconciled at turn level, including its attempt count and completion time.

Selection aggregate

selection_ranks preserves all ranks 1–15 and all 2,339 entries in the published Figure 3 aggregate. The top-three display boundary is the study's object. The raw rank-3/rank-4 count ratio, 290/160, is about 1.81; it is not the fitted regression-discontinuity estimate of 1.75 reported in the paper. This table alone does not reproduce that fitted estimate or its uncertainty interval.

Source: the paper's public aggregate repository. Dataset release version: 1.1.0.

Privacy and usefulness

The export uses field allowlists. Original identities, source IDs, exact timestamps, asset names, trade direction, prices, quantities, leverage, balances, returns, user strategy text, prompts, reasoning, raw arguments/results, model/provider configuration and internal paths or endpoints are excluded. Version 1.1.0 adds sanitized request/response projections containing structural fields, selected nonfinancial outcome values and local aliases. Every projection is explicitly non-verbatim. The source-to-sample mapping stays with the data owner.

The excerpts can illustrate result parsing, deferred completion, retries, parent/child handoff and event ordering. The questions provide proposed evidence-reading exercises with answer keys. Neither is a full reproducible trajectory set, representative frequency estimate, model-training corpus, or validated benchmark. They cannot evaluate investment performance or reproduce the source decisions. A broader scenario benchmark and dataset remain separate work in progress.

The wider historical study did not establish a directional trading edge. Its complete findings and limitations remain available in the paper. This subset is scoped to control mechanics and does not replace the full record or measure the present alpha.

License and citation

License: CC BY 4.0. Attribute DX Research Group, the dataset version and the associated paper. The source event records are owner-held supplemental material; the public paper establishes the system and study context, not independent verification of these selected excerpts.