git:20260926.80ef163 to git:20260926.2d8ce6a

24 added, 13 removed. Audit A to A.

---

name: cohesix-model-rollout

- description: Compare a Mac MLX candidate with a CUDA deployment and trace an approved model choice through transfer, canary, promotion or rollback. Use for cross-host model rollout decisions, not for training an individual PEFT adapter.

+ description: Choose and verify a model rollout on a selected Mac MLX host, Linux CUDA host or both. Use for model comparison, optional transfer, canary, promotion or rollback; not for training an individual PEFT adapter.

license: Apache-2.0

---

-

+

# Roll out a selected model

- Use a matching installed release and the deployment's selected Mac and CUDA

- profiles. Read the same-version [Mac operation guide](../../docs/HOST_TOOLS.md#swarmui),

+ Use a matching installed release and the deployment's selected host profile:

+ Apple Silicon macOS with admitted MLX, Linux AArch64 NVIDIA with admitted CUDA,

+ or both. A Mac or Linux controller can coordinate a remote selected host.

+ Read the same-version [host operation guide](../../docs/HOST_TOOLS.md#swarmui),

[private release guide](../../docs/PRIVATE_LORA_RELEASE.md) and

[Release B host/client contract](../../docs/BUILD_PLAN.md#release-b).

The Mac's local MLX workbench is useful for exploration; its local observation

- does not itself admit a Cohesix job or certify a remote CUDA deployment.

+ does not itself admit a Cohesix job. A Linux CUDA result must come from its

+ selected native executor. Use only the stages and transfer path advertised by

+ the installed profile.

+ Inspect each candidate host's OS, accelerator, model format, runtime and

+ capacity before comparing or activating. If a combination is unsupported,

+ explain the exact incompatibility and offer a compatible selected host or a

+ separately qualified format/runtime change; do not assume Mac and Linux

+ installations can load the same artifact.

## Make the choice reproducible

Before work starts, name the application task, held-out input set, quality

measure, budget, selected base/model/adapter digests, runtime versions and

- provider/host for each stage. Compare actual outputs and resource use on the

- relevant Mac and NVIDIA hosts. Report unsupported formats, stale capacity or

- incomparable evaluation as blockers, not as a winner. A vMLX model response

- can be a serving observation; its MCP client role and any A2A peer using it

- require separate configured compatibility.

+ provider/host for each stage. Compare actual outputs and resource use on each

+ selected host. For a single-host rollout, compare candidates on that host;

+ for a mixed rollout, keep Mac Metal and Linux CUDA measures separate. Report

+ unsupported formats, stale capacity or incomparable evaluation as blockers,

+ not as a winner. A vMLX model response can be a serving observation; its MCP

+ client role and any A2A peer using it require separate configured compatibility.

Write a decision record with the chosen candidate and why it beat the

- incumbent. Preserve stage/job identities and original admission IDs. When

- the selected operation needs to move model bytes, first establish that the

- installed release and host pair actually advertise verified distribution.

+ incumbent. Preserve stage/job identities and original admission IDs. A

+ single-host rollout has no cross-host transfer step. When the selected

+ operation needs to move model bytes, first establish that the installed

+ release and host pair actually advertise verified distribution.

A Queen reference or FUSE listing is only a reference. Record the authorised

source, destination and payload route; require the full destination digest

and receipt before a **separate** activation decision. Do not infer transfer

from a client timeout, URL or chunk acknowledgement.

## Prove the generation users reached

Preflight the destination's model/runtime compatibility and available

capacity. Admit the selected release, observe the native load, then send a

real application canary to the exact candidate generation. Promote only after

its predeclared quality and service checks pass. On failure, use the original

job lineage to inspect or recover, then observe a client reaching the known

good incumbent after any rollback. A successful control action, a `g1`

diagnostic label and a running server process are not serving proof.

Return a compact flight record: candidates and measures, chosen host at each

stage, source and destination hashes if bytes moved, admission IDs, canary

generation and output, promotion or rollback result, shared-verifier outcome

and unresolved questions. Route adapter-specific training/import and held-out

release mechanics to the [private adapter skill](../cohesix-private-adapter-release/SKILL.md);

route contested evidence to the [evidence skill](../cohesix-evidence/SKILL.md).