BENCHMARK & DATASET · v1.0
Database diagnosis with verifiable evidence

Evaluate the diagnosis, not only the final answer.

DataTrace asks AI agents to find the single faulty database object behind a pipeline failure or incorrect result, repair it, and support the diagnosis with relevant evidence.

Each task represents database artifacts—such as schemas, data, queries, constraints, plans, tests, logs, and issue reports—as a dependency graph. Because the benchmark knows where the fault was introduced, it can score both root-cause accuracy and whether the agent inspected the artifacts that connect the observed problem to that root cause.

83
runnable environments
62
distinct domains
7
fault types
0
handwritten reasoning traces
The problem

A wrong result may originate far from where it appears.

A database pipeline can produce an incorrect result without raising an error. The underlying problem may be an unenforced constraint, a renamed column, a corrupted source record, or a broken view several steps upstream. DataTrace checks both the reported cause and the evidence used to find it.

Outcome-only grading

Only the final answer is checked

A correct answer receives full credit even when the agent did not inspect the artifacts needed to support it.

# agent output root_cause = stg_orders.customer_id # evidence inspected: (none) verdict: CORRECT ← insufficient evidence
DataTrace · evidence-grounded

The answer and supporting evidence are checked

Every read action records the graph nodes it touched. A correct answer is compared with the reference evidence path π*. A correct answer with low evidence coverage is reported separately.

trace_lineage(fct_orders) → stg_orders → raw_orders profile_table(raw_orders) → dup customer_id root_cause = raw_orders.customer_id verdict: CORRECT · GROUNDED (ER 0.83)
schema data query constraint plan test log incident

Leaderboard

Grounded accuracy rewards supported diagnoses.

Frontier models on the identical frozen corpus (64 tasks × 2 repeats, n = 128 each; one scaffold, one parser). Ranked by RC@1ᵍ — correct and evidence-recall ≥ 0.5. The amber segment is the evidence gap (RC@1 − RC@1ᵍ): correct answers with insufficient evidence.

Grounded correct (RC@1ᵍ) Correct, insufficient evidence Incorrect click a column to sort
# Model Grounded accuracy  CUR ER Worst-env

95% CIs are environment-clustered bootstrap intervals (4000 draws over 13 environments). CUR = P(ungrounded | correct). ER = evidence recall. Worst-env = lowest RC@1ᵍ over environments with ≥ 6 episodes. Full table with Fix / EF1 / cost / steps in leaderboard.md.


What we found

Higher accuracy does not always mean better evidence coverage.

Outcome-only scores can hide unsupported diagnoses. In these experiments, reference-path coverage also provided a useful training signal.

FINDING 01

Stronger models get more right

Across the controlled model ladder, root-cause accuracy climbs RC@1 0.18 0.82. Capability clearly helps agents reach the correct object.

FINDING 02

Unsupported correct answers also increase

Over the same model comparison, the share of all runs that are correct but have insufficient evidence rises 0.05 0.27. The conditional unsupported rate (CUR) remains present at every tier.

FINDING 03

Grounding is trainable

Rewarding reference-path coverage during training reduces CUR 0.163 0.054 while RC@1 0.292 0.367 rises.


How it is scored

Two levels: outcome and evidence.

The reference answer—the defective object, the reference evidence path π*, and lineage distance — is derived mechanically from the known fault and dependency graph, without a handwritten reasoning trace for each task.

RC@1
Root-cause accuracy

Exact-match on the single defective object (outcome).

RC@1ᵍ
Grounded accuracy

Correct and evidence-recall ≥ 0.5. The ranking metric.

infl
Evidence gap

RC@1 − RC@1ᵍ: the share of all runs that are correct but insufficiently supported.

CUR
Conditional unsupported rate

Among correct answers, the fraction with insufficient evidence.

ER / EF1
Evidence recall / F1

Artifacts inspected by the agent compared with the reference path π*.

LG
Unsupported-correct rate

The root cause is correct, but evidence recall is below 0.5.

Fix
Repair success

Repaired outputs match the reference outputs and all tests pass.

EG-Elo
Self-play rating

Rating used in the optional evidence-aware self-play experiment.


The corpus

Built from controlled, reproducible faults.

Seven fault types introduce a known defect, run the pipeline again, and record the resulting failure or incorrect output. The dependency graph is then used to construct the reference evidence path. The frozen evaluation corpus contains 448 instances (mean reference-path length 5.1; mean dependency distance 3.0).

By fault category

By fault type

By number of artifact types on the reference path

By difficulty tier (1 easy → 4 hard)


Citation

Cite DataTrace

If DataTrace is useful in your research, please cite the project.

BibTeX
@misc{datatrace,
  title        = {DataTrace: Evidence-Grounded Root-Cause Localization for Databases},
  author       = {The DataTrace Authors},
  howpublished = {\url{https://anonymous.4open.science/r/datatrace}},
  year         = {2026}
}