Trialground
Trialground is the evaluation capability inside Teammately correctness infrastructure. It gives AI engineers a managed proving ground for running exact candidate implementations against exact Benchmark Versions while preserving the response, Rubric outcomes, settings, and metadata needed to explain the result.
Definition
A candidate is represented by a saved Harness version rather than an informal description of a prompt or code state. A benchmark version provides the cases and governed correctness boundary. Benchmark Evaluations binds those inputs into Runs, computes rubric-level evidence, and exposes results through Dashboard, List, Compare, Arena, and run-detail views.
Trialground names the capability; Harnesses and Benchmark Evaluations are the current UI surfaces. The distinction matters when giving instructions: users edit and save a Harness under Assets, then select its saved version for a benchmark evaluation.
Decision checkpoint
| Question | Inspect | Required boundary |
|---|---|---|
| What code or prompt was evaluated? | Harness and saved version | Draft editor state is not the candidate identity |
| What behavior space was tested? | Benchmark version and dataset snapshot | Dataset membership is fixed for interpretation |
| Which standards judged each case? | Applicable policies and rubrics | Suggested or unrelated standards are excluded |
| Why did a candidate pass or fail? | Case response and applicable Rubric outcomes | Aggregate movement is traced to Case-level evidence |
| Can two candidates be compared? | Compare or Arena | Both candidates expose compatible benchmark and settings boundaries |
Managed evaluation
Harness Assets provide a multi-file editing surface, Draft and Saved states, coding-agent assistance, secret references, and benchmark activation. Evaluation setup selects the relevant Harness version and preserves run metadata. Managed execution can scale case work while keeping every result attached to the benchmark and candidate identities.
The current Benchmark Evaluations API does not expose execution trajectories, spans, private reasoning, or tool traces. The visible Traces / Spans segment is therefore a capability fence, not inspectable Run evidence. Improve's narrated trajectories are separate session chronology and must not be described as evaluation execution traces. Rubric outcomes remain the governed evidence within the evaluation contract.
Reading evidence
Start with Dashboard to understand overall progress and distribution, then use List for Run and Case-level results. Compare is a symmetric matrix whose columns are saved Harness Versions and whose rows can group Cases, all results, Policies, Rubrics, Dimensions, Topics, Groups, or Patterns. Arena computes pair evidence from admitted results; it is not a subjective preference interview. Run detail is the place to resolve execution state, output mapping, response evidence, and failures.
When a result is surprising, check versions and configuration before changing the candidate. A different benchmark snapshot, Harness version, output mapping, evaluation setting, or run-metadata value can explain movement that otherwise looks like a behavior change.
Worked example
Retrieval candidate comparison
Two saved Harness versions use different source-selection strategies. Both run against the same benchmark version containing current-source, stale-source, and missing-source cases. Compare shows that one candidate improves current-source grounding but regresses when no controlling source exists. The team opens the failed cases and rubric results before deciding whether the candidate belongs in the next Improvement Session.
Related workflows
Related reference pages
Source confidence
Doctrine-backed: this page defines Trialground as the public capability. Harness and Benchmark Evaluation pages provide code-backed behavior and state details.