Overview

Terminal task curation and verification

Published batches

Tasks

Search tasks Status Source Batch

Current publication is the manifest-listed pool for new downstream jobs. Retained versions preserve earlier inputs. Validation results shown here are runtime evidence; semantic audits and solved agent trajectories are separate checks.

What this board is reading

Measured artifacts · shared Curator visual system

Current run	Live state, plus the completed report only when its run ID matches.
Published batches	Verified manifest and immutable versions; metadata comes from each task.
Validation	Oracle/NOP report verdicts. Available local trial results and JUnit supply actual rewards and test counts.
Run history	Run index and matching archived reports. Missing measurements appear as —.
Privacy	No task text, source files, solutions, prompts, traces, config, endpoints, host paths, or raw exception text are included.

Oracle must receive reward 1 and NOP reward 0 in separate fresh environments, with meaningful assertions and no trial exception. A reported pass with unavailable raw evidence is labeled “report only.” Semantic requirement/test review and successful Tracer/Evaluator solving require their own evidence.

Counts describe different scopes: the latest attempt may fail while a previous publication remains available. This board does not invent collection totals, difficulty scores, or semantic approvals. Source, task IDs, run IDs, tags, and numeric usage remain visible; review them before sharing a snapshot.