Use ServingStudio Analyzer

Use Analyzer whenever a user-visible claim depends on an existing ServingStudio Sim
result. The conversation backend owns lifecycle and ownership links; Analyzer
owns result values, descriptors, curves, plots, and hardware limits.

Select the exact resource

Prefer evidence in this order:

1. the explicit Analyzer selection supplied by the UI;
2. the stable Analyzer resource ID returned by the managed workflow just run;
3. a resource ID the user named;
4. recent ready candidates from
   /api/v1/sweeps?status=ready&limit=5, or
   /api/v1/sweeps/latest when one newest candidate is useful.

Catalogs are discovery evidence, not result evidence. They intentionally expose
opaque IDs. Before choosing a candidate, compare its display_name, ordered
axes, deployments, traces, status, and time with the user's request.
“Latest” means newest ready result, not automatically the semantically correct
result. If multiple candidates remain plausible, show the concise candidates
and ask instead of silently choosing.

A ready result produced by another conversation in the same workspace is valid.
The producing conversation and turn are provenance, not read authorization.
Never cross a workspace boundary unless the product supplies an explicit
attachment/import mechanism.

Read through Analyzer MCP

Use the read_analyzer_resource MCP tool. Set:

• source="host" for the shared Analyzer service, including a result selected
  in the Analyzer UI;
• source="workspace" for a result that exists only in the current Agent
  workspace.

source selects the data location; it does not change workspace ownership.
After selecting a resource ID, read its exact typed endpoint. Common endpoints
include:

/api/v1/sweeps/{sweep_id}/payload
/api/v1/runs/{run_id}/...
/api/v1/predictions/{prediction_id}/...
/api/v1/kernel-profiles/{profile_id}/...
/api/v1/kernel-measurements/{measurement_id}/...

In a managed UI turn this returns one compact evidence block:

{
  "resource": {"id": "e_...", "name": "20260801_3_llama3_8b_tp"},
  "axes": ["tensor_parallel", "request_rate"],
  "metrics": {
    "total_tps": {
      "label": "Total throughput",
      "unit": "tok/s",
      "objective": "maximize",
      "citation": "exp.throughput"
    }
  },
  "rows": [
    {
      "coordinates": {"tensor_parallel": 2, "request_rate": 20},
      "values": {
        "total_tps": {
          "raw": 30124.2,
          "citation": "exp.tp2.rate20.throughput"
        }
      }
    }
  ]
}

The compact block is the user-visible numerical authority. Exact run and
prediction reads return their citation beside the result. Kernel profile and
measurement reads return one compact result plus a citations map keyed by
the metric, panel, or plot represented in that result. Their published token
namespaces are respectively run.*, pred.*, kprof.*, and kmeasure.*.

Do not replace an Analyzer result
with /api/jobs, launcher status text, an implementer summary, or direct report
file parsing when the Analyzer resource is ready.

Write the answer

• Report the exact raw value and its declared unit.
• Copy the adjacent citation token unchanged as Markdown inline code.
• Never assemble a token from axes or metric names, and never invent a token.
• Cite a panel-wide qualitative claim with the metric citation when suitable;
  cite coordinate-specific numerical claims with the matching row citation.
• A derived value may be calculated from returned raw values, but cite every
  input coordinate used and state that the result is derived.
• Label provenance accurately: simulated prediction, measured result,
  catalog fact, or derived from Analyzer values.
• Do not emit Analyzer URLs, navigation target JSON, percent-encoded strings,
  or workspace file links as substitutes for evidence citations.
• Writing a token never navigates automatically. The user chooses whether to
  click the rendered citation.

Example:

At TP2 and 20 req/s, simulated throughput is 30,124.2 tok/s.
`exp.tp2.rate20.throughput`

Follow-up turns and typed resources

Re-read the exact Analyzer resource in every turn that reports its numbers.
This registers a current citation dictionary while historical messages keep
their already-frozen targets.

Sweeps, individual runs, timing predictions, kernel profiles, and kernel
measurements all have first-class citation contracts. Use the token returned by
the exact endpoint being interpreted:

• sweep aggregate: exp.*;
• individual simulation run: run.*;
• timing prediction: pred.*;
• kernel profile: kprof.*;
• kernel measurement: kmeasure.*.

These examples identify namespaces only. Never derive the remainder of a token
from a metric name, slot, curve, or plot. Copy the complete token returned by
Analyzer beside that exact result.

If a resource is missing, pending, failed, ambiguous, or lacks the requested
metric/coordinate, say exactly that. Do not estimate a replacement value.