feat(insights): Add Kernel Factory dataset

packages/nemo_platform_ext/src/nemo_platform_ext/skills/nemo-intake/references/ingest-formats.md

@@ -197,6 +197,8 @@ attributes retain native JSON types. Known semantic attributes populate queryabl
 attributes appear in detailed reads under `raw_attributes`. Reposting the same
 `(source, trace_id, span_id)` updates the existing logical span.
 
+String inputs and outputs are stored verbatim; objects and arrays are serialized as JSON.
+
 The default ClickHouse TTL is 90 days from `started_at`. If any span is outside that window, Intake
 returns `422` before writing the batch and instructs the operator to increase the `spans` and
 `trace_index` TTLs. Provider timestamps are never rewritten by the endpoint.

plugins/nemo-insights/evaluation/README.md

@@ -12,7 +12,7 @@ wheel.
 
 - Run every command from `plugins/nemo-insights`; the `evaluation` package is
   not installed in the plugin wheel.
-- Install the AWS CLI and configure the CSS S3 credentials described in
+- Install AWS CLI 2.33.0 or newer and configure the CSS S3 credentials described in
   [State bundles](#state-bundles).
 - Start a local NeMo Platform at `http://localhost:8080` before using pinned
   analysis, restore, roundtrip, or guarded publish commands.
@@ -158,7 +158,10 @@ uv run python -m evaluation publish evaluation/tmp/glamr.tar.zst --base http://l
 ```
 
 `snapshot` drains the subject's workspaces (benchmark subjects: realistic +
-`-oracle` twin) into JSONL + manifest — no ClickHouse, no Docker. `publish`
+`-oracle` twin) into JSONL + manifest — no ClickHouse, no Docker. An intake
+subject can set `experiment = "<name>"` to capture only that Experiment's
+complete traces; membership comes from the Experiment's evaluations and their
+traces, then every span in each trace is exported by trace ID. `publish`
 refuses to mint unverified: `--base` runs the round-trip fidelity guard there
 first (re-ingest into scratch workspaces → re-export → doc diff), or pass
 `--no-verify` only after running `roundtrip` separately and confirming it
@@ -190,6 +193,8 @@ What restore touches:
 - Accepted losses: annotation/evaluator-result `created_at`/`created_by` are
   server-stamped at restore (the write APIs reject client values); a running
   platform is required to analyze.
+- Non-OTLP spans, including ATIF, restore through the provider-neutral direct
+  ingest API so their source names and arbitrary span IDs remain unchanged.
 - Legacy `state-v1..v5` tars are not present in CSS (the v4 corpus lives on as
   `state-v6`); a stray local copy restores only from a checkout
   predating the v6 migration.

plugins/nemo-insights/evaluation/artifact.py

@@ -93,6 +93,8 @@ def build_export_manifest(
         "platform_revision": (platform_info or {}).get("revision"),
         "records": [r.name for r in records],
     }
+    if stats.get("selections"):
+        manifest["selections"] = stats["selections"]
     env = env or {}
     manifest |= {key: env[var] for key, var in _LINEAGE_ENV.items() if env.get(var)}
     return manifest
@@ -222,12 +224,19 @@ def snapshot_export(
     CLI's ``--base`` override rewrites every stanza before this is called).
     """
     subject_workspaces: list[tuple[Subject, list[str]]] = []
-    claimed_workspaces: set[str] = set()
+    claimed_workspaces: dict[str, Subject] = {}
     for subject in subjects:
+        for workspace in workspaces_for_subject(subject):
+            owner = claimed_workspaces.get(workspace)
+            if owner is not None and (owner.config.get("experiment") or subject.config.get("experiment")):
+                sys.exit(
+                    f"snapshot: workspace '{workspace}' is shared by experiment-scoped subject "
+                    f"'{owner.name if owner.config.get('experiment') else subject.name}' and another subject"
+                )
         workspaces = [workspace for workspace in workspaces_for_subject(subject) if workspace not in claimed_workspaces]
         if workspaces:
             subject_workspaces.append((subject, workspaces))
-            claimed_workspaces.update(workspaces)
+            claimed_workspaces.update({workspace: subject for workspace in workspaces})
     # Every selected subject must carry a base_url: a partial miss would silently
     # let the agreement check pass on the configured subset and export the
     # unconfigured subject from the others' platform.
@@ -256,6 +265,7 @@ def snapshot_export(
                 state,
                 since=since,
                 client=_basic_auth_intake_client_for(subject, source_url),
+                experiment=str(subject.config["experiment"]) if subject.config.get("experiment") else None,
             )
             for subject, workspaces in subject_workspaces
         ]
@@ -275,6 +285,11 @@ def snapshot_export(
             },
             "min_start_time": min(min_bounds).isoformat() if min_bounds else None,
             "max_start_time": max(max_bounds).isoformat() if max_bounds else None,
+            "selections": {
+                workspace: selection
+                for result in exported
+                for workspace, selection in result.get("selections", {}).items()
+            },
         }
         for rec in records:
             shutil.copy2(rec, state / "tmp" / rec.name)

plugins/nemo-insights/evaluation/cli.py

@@ -95,17 +95,30 @@ def _load_dotenv(path: Path = ENV_PATH) -> None:
             os.environ.setdefault(key, value)
 
 
+def _aws_cli_version(executable: str) -> tuple[int, int, int] | None:
+    result = subprocess.run([executable, "--version"], capture_output=True, text=True)
+    match = re.search(r"aws-cli/(\d+)\.(\d+)\.(\d+)", result.stdout + result.stderr)
+    if not match:
+        return None
+    major, minor, patch = match.groups()
+    return int(major), int(minor), int(patch)
+
+
 def _doctor(subjects: dict[str, Subject], name: str | None) -> None:
     """Print a readiness checklist for one subject (or all): what's set up, what's not."""
     names = [name] if name else sorted(subjects)
+    aws = shutil.which("aws")
+    aws_version = _aws_cli_version(aws) if aws else None
     for subject_name in names:
         subject = subjects.get(subject_name)
         if subject is None:
             print(f"✗ {subject_name}: unknown subject")
             continue
         unmet: list[str] = []
-        if shutil.which("aws") is None:
-            unmet.append("AWS CLI (needed for pinned/--state analyze; install with `brew install awscli`)")
+        if aws is None:
+            unmet.append("AWS CLI 2.33.0+ (needed for pinned/--state analyze; install with `brew install awscli`)")
+        elif (aws_version or (0, 0, 0)) < (2, 33, 0):
+            unmet.append("AWS CLI 2.33.0+ (required for immutable multipart state uploads)")
         if not os.environ.get(release.ACCESS_KEY_ENV) or not os.environ.get(release.SECRET_KEY_ENV):
             unmet.append("CSS S3 credentials in evaluation/.env (CSS Portal → Auth Info)")
         unmet += build_adapter(subject).check()

plugins/nemo-insights/evaluation/evaluations.toml

@@ -14,6 +14,13 @@ agent = "content-dedup"
 workspace = "nvq"
 state = "state-v7"
 
+[kernel-factory]
+type = "intake"
+agent = "solswarm-campaign"
+workspace = "kf-prod-evals"
+experiment = "prod-latest-completed"
+state = "state-v11"
+
 [glamr]
 type = "intake"
 agent = "glamr"                                 # the main GLAMR agent (ux-agent); the only agent_name-tagged one