Import upstream Verifiers episodes into Bench

kbench/cli.py

@@ -12,11 +12,12 @@ from rich.console import Console
 from rich.table import Table
 
 from . import tasks as task_catalog
-from .compare import compare as compare_runs, direction
-from .sweep import analyse, point_from_run
+from .compare import compare as compare_runs
+from .compare import direction
 from .registry import load_registry
 from .results import RESULTS_DIR, Run
 from .run import build_run, run_perf, run_quality
+from .sweep import analyse, point_from_run
 
 app = typer.Typer(
     add_completion=False,
@@ -307,6 +308,51 @@ def card_cmd(path: Path) -> None:
     _print_card_dict(data)
 
 
+@app.command("import-verifiers")
+def import_verifiers_cmd(
+    target: str = typer.Argument(..., help="registered model/host target that produced the traces"),
+    path: Path = typer.Argument(..., help="Verifiers run directory or traces.jsonl"),
+    task: str = typer.Option(..., help="Bench task name for this imported environment"),
+    tier: str = typer.Option("signal", help="reference | signal | canary"),
+    pass_threshold: float = typer.Option(1.0, help="score required for an episode to pass"),
+    primary_metric: str | None = typer.Option(
+        None, help="trace metric to import; defaults to summed weighted rewards"
+    ),
+    save: bool = typer.Option(True, help="write the existing results/*.json scorecard"),
+) -> None:
+    """Import upstream Verifiers episodes; execute no model and no harness."""
+    from .verifiers_import import FORMAT, VerifiersImportError, import_quality, source_sha256
+
+    registry = load_registry()
+    registered = registry.target(target)
+    try:
+        quality = import_quality(
+            path,
+            task=task,
+            tier=tier,
+            pass_threshold=pass_threshold,
+            primary_metric=primary_metric,
+        )
+        digest = source_sha256(path)
+    except VerifiersImportError as exc:
+        raise typer.BadParameter(str(exc), param_hint="path") from exc
+
+    run = build_run(registry, registered, quality=[quality], perf=None)
+    run.runner["quality_import"] = {
+        "format": FORMAT,
+        "source_sha256": digest,
+    }
+    run.sampling = {
+        "source": "verifiers",
+        "primary_metric": primary_metric,
+        "pass_threshold": pass_threshold,
+    }
+    _print_card(run)
+    if save:
+        saved = run.save()
+        console.print(f"\n[green]saved[/green] {saved.relative_to(Path.cwd())}")
+
+
 def _print_card(run: Run) -> None:
     _print_card_dict(run.to_dict())