Batch 4: AI / SLM / Fine-Tuning Platform (`1.0.7-beta`)

[debuggerone]

Source of truth: BACKLOG.md on development/1.0.7-beta.

This issue tracks one release batch. Work should be split into branch-local SPRINT.md slices before implementation.

Batch 4: AI / SLM / Fine-Tuning Platform (1.0.7-beta)

#AI-1 Distributed Model Placement And Inference Execution

Goal:

• Move from proven single-node/model-fit surfaces to real fleet-aware model placement and distributed inference execution.

Checklist:

• Define network-aware model placement across multiple nodes.
• Define distributed inference execution across multiple servers.
• Define future MoE-ready multi-node expert routing.
• Add placement tests that use node hardware profile, health, memory pressure, model artifact locality, and network cost.
• Add fail-closed behavior for no-fit, no-healthy-node, and stale placement state.

Done:

• Placement decisions are deterministic, explainable, and backed by real multi-node proof.

#AI-2 Prompt, Cache, And Checkpoint Persistence

Goal:

• Persist model-facing runtime state where claimed, using King-owned object-store/runtime contracts.

Checklist:

• Define object-store-backed prompt persistence where applicable.
• Define object-store-backed response/cache persistence where applicable.
• Define checkpoint persistence for long-running AI workflows.
• Define ownership, retention, encryption, and invalidation semantics.
• Add restart/resume tests for persisted prompt/cache/checkpoint state.

Done:

• AI workflow state survives restart only where explicitly claimed and tested.

#AI-3 Fine-Tuning And Training Data Workflows

Goal:

• Add real fine-tuning workflow contracts instead of only inference and retrieval surfaces.

Checklist:

• Define training-data extraction pipelines from stored data.
• Define dataset-building pipelines for task-specific SLMs.
• Define dataset versioning and lineage.
• Define fine-tuning workflows for small models.
• Define fine-tuning artifact storage and recovery.
• Define fine-tuned model registration and reuse.
• Define evaluation and validation surfaces for fine-tuned models.
• Finalize the public contract for fine-tuning workflows.

Done:

• A fine-tuned model can be produced, registered, evaluated, and reused under a documented King contract.

#AI-4 Advanced Model Extensions

Goal:

• Keep advanced model work extensible without bloating the built-in SLM surface.

Checklist:

• Define advanced model capabilities as extension-based functionality.
• Define external providers as extension-based functionality.
• Define larger model families as extension-based functionality.
• Define advanced routing, fallback, and policy layers as extension-based functionality.
• Define advanced multimodal capabilities as extension-based functionality.
• Define clear public contract boundaries between the built-in AI platform and later model extensions.

Done:

• Built-in AI platform and extension-based advanced model features have a clean, tested boundary.