Chapter 2.8 follows chapter 2.7, The closeout chain of a batch, and precedes chapter 2.9, The configuration files and their defaults, within part 2, the harness.
A cell can take a paid route, a request path that reaches a hosted model service, or a local route, a request path that reaches the campaign machine. The paid route sends the harness request to the hosted model. The local route sends it to the campaign inference machine. The route names in agent_backends.py are dgx_codex and dgx_claude; agent-coder, the alias for the coder-facing local setup, selects that configuration.
source: operations/site-ia/A6-page-briefs.md, section ”# H7.”
2.8.1 Route selection of the measurement harness
The harness selects the route for a cell before the model request begins. The paid route depends on the hosted provider. The local route depends on the Spark machine and its serving service. The route selection code is in agent_backends.py. The route names identify the request family, while the alias identifies the model-facing configuration. The local route therefore has a machine-side setup that the paid route does not require.
2.8.2 Serving controller coordination of the measurement harness
Fabric, the serving controller for the Spark machine, coordinates access to the local model service. A lease, a time-bounded reservation of serving access, gives a cell permission to use that service. The launcher obtains the lease before it starts the model endpoint. Fabric records the lease activity and the request events. When the cell finishes, the serving process and its lease can be released. source: operations/site-ia/A6-page-briefs.md, section ”# H7.”
2.8.3 Endpoint process of the measurement harness
The batch launcher starts an isolated endpoint process for the cell. The process serves the selected local model through the lease and stays separate from other cell processes. The serving scripts prepare the machine, download the model, start the endpoint, and verify access from the development box. The environment example supplies the values used by those scripts. Fabric telemetry records the serving lifecycle for later inspection.
2.8.4 Model aliases and profiles
The hosted route is pinned to claude-haiku-4-5-20251001, recorded on 2026-08-09.
source: model_matrix.json key models
The local serving profiles are glm-4.7-flash.env, gpt-oss-120b.env, qwen3.6-35b-a3b-nvfp4.env, and qwen3-coder-next.env.
source: folder listing
The profile names are configuration handles. They are separate from the model names used in the campaign records. Appendix N section 5 identifies deepseek-v4-flash and Kimi K3 as candidate models. source: Appendix N section 5
The model_class column of run_index.csv identifies Colibri 1.8.0 where that name occurs in the run records.
source: model_class column of run_index.csv
2.8.5 Request protocols of the measurement harness
The harness can use the OpenAI Responses protocol, a request and response format for the OpenAI-compatible endpoint, or the Anthropic Messages protocol, a request and response format for the Anthropic-compatible endpoint. The selected protocol determines the message shape sent to the endpoint and the response shape returned to the harness. The route and the protocol are separate choices: either route can be paired with the protocol required by its model endpoint. source: operations/site-ia/A6-page-briefs.md, section ”# H7.”
2.8.6 Ledger basis of the measurement harness
One 95-turn pass in run set 035 seed 2 used 8,240,883 ledger tokens. source: run_defaults.json line 7
Figure D-H7-1. The paid route and the two local routes from the cell runner to a model
%% figure D-H7-1 %% protocol: str = "responses" flowchart TB subgraph serving["Serving controller"] n1_1["agent_backends.py"] --> n1_2["dgx_fabric.py"] n1_2 --> n1_3["FabricClient"] n1_3 --> n1_4["resolve_fabric_endpoint()"] end subgraph servingkit["Serving kit"] n2_1["HeartbeatTask"] --> n2_2["spark/01_bootstrap.sh"] n2_2 --> n2_3["04_verify_from_devbox.sh"] n2_3 --> n2_4["spark/profiles/*.env"] n3_1["config/model_matrix.json"] --> n3_2["preflight_batch.py"] n3_2 --> n3_3["probe_dgx_host()"] n3_3 --> n3_4["protocol: str = "responses""] n1_4 --> n2_1 n2_4 --> n3_1 end
source: operations/site-ia/A3-diagram-specification.md
source: operations/site-ia/A6-page-briefs.md, section ”# H7.”
D-H7-1. The paid and local routes converge on model requests, while the local route passes through Fabric and its lease. source: operations/site-ia/A6-page-briefs.md, section ”# H7.”
source: operations/site-ia/A6-page-briefs.md, section ”# H7.”