[Epic] Deliver the Hermes Harness Optimization Blueprint with NeMo Platform Insights

## Objective

Deliver a reproducible Hermes harness-optimization blueprint that uses Relay telemetry, trace insights, and NeMo Platform optimization capabilities to demonstrate measurable improvements in token consumption, cost, and task accuracy.

The result must include public-safe evaluation data, opinionated task configurations, and enough evidence for another user to reproduce the baseline and optimized results.

## Product boundary

This Epic optimizes the agent harness through prompts, tools, task configuration, and supported routing inputs. It does not modify model weights or include model post-training.

Hermes is the initial harness. Expansion to other agent runtimes requires separately accepted scope and evidence.

#7937 owns the supported Relay and Switchyard telemetry and routing capability. This Epic consumes that capability and must not reimplement its collection, credential, policy, or routing contracts.

## Scope rules

- Select a public-safe task set that represents meaningful Hermes behavior.
- Define token, cost, and accuracy measurements before running the final evaluation.
- Record the exact Hermes, NemoClaw, model, provider, Relay, Switchyard, prompt, tool, and task-configuration identities.
- Measure the baseline and optimized configurations against the same evaluation boundary.
- Use Relay telemetry and trace insights to identify optimization opportunities.
- Apply an accepted NeMo Platform insight and optimization workflow to the experiment loop.
- Keep raw evaluation inputs, scoring rules, and aggregate results available in an approved public surface.
- Publish opinionated task configurations rather than only reporting experimental observations.
- Report regressions and tradeoffs. Do not present cherry-picked tasks or incompatible baselines as improvement.
- Keep credentials, private prompts, internal data, and conversational content out of published evidence.

## Evaluation workflow

| Stage | Required outcome |
|---|---|
| Task selection | A public-safe task set, scoring method, and success thresholds are defined before final evaluation. |
| Baseline | Token consumption, cost, accuracy, and relevant diagnostic measurements are recorded for the unoptimized configuration. |
| Analysis | Relay telemetry and traces identify prompt, tool, configuration, or routing opportunities. |
| Optimization | Candidate configurations are produced through documented experiments and the accepted NeMo Platform workflow. |
| Validation | Baseline and candidate configurations are compared using the same task set, models, provider boundary, and scoring method. |
| Publication | The selected configurations, results, reproduction steps, and known limitations are published. |

## Relationships and readiness

Implementation of the task set, scoring method, baseline harness, and configuration format may proceed while #7937 is in progress.

Final telemetry-backed evaluation and completion cannot occur until #7937 provides the accepted Relay and Switchyard capability or records an approved replacement.

## Completion criteria

- [ ] A public-safe Hermes task set and objective scoring method are published.
- [ ] Success thresholds for token consumption, cost, and accuracy are defined before final evaluation.
- [ ] A reproducible baseline records the exact software, model, provider, prompt, tool, and configuration identities.
- [ ] Relay telemetry and trace insights are captured through the supported capability from #7937.
- [ ] The accepted NeMo Platform insight and optimization workflow is documented.
- [ ] At least one opinionated configuration demonstrates measurable token, cost, and accuracy improvement against the baseline.
- [ ] Results include repeated or otherwise statistically defensible measurements appropriate to the task set.
- [ ] Regressions, limitations, and tradeoffs are recorded alongside improvements.
- [ ] Opinionated Hermes task configurations are published in an approved reusable surface.
- [ ] A fresh environment can reproduce the evaluation from the documented configuration and exact artifact identities.
- [ ] Published evidence contains no credentials, private data, or private conversational content.

## Non-goals

- Model fine-tuning, reinforcement learning, or other model post-training.
- Reimplementing Relay telemetry or Switchyard routing owned by #7937.
- Building a general-purpose observability platform.
- Optimizing every Hermes workflow, model, provider, or deployment topology.
- Claiming improvement from private-only data or unreproducible evaluation.
- Expanding the blueprint to every NemoClaw agent runtime.