# Hive-owned execution authority Owner requirement confirmed on 15 November 2026. This is the required product contract. The bounded complete-model path is now implemented and deployed; required-model partitioning remains a migration objective. ## One product execution path Clients submit work to Hive. Hive and its assigned hub own admission, placement, work identities, leases, cancellation, result acceptance and contribution accounting. Workers provide authorised memory and compute. No Mac, personal laptop, local dashboard or permanent target-model host is required between the client, Hive and enrolled workers. The current Mac is a development machine used to simulate hub components and run controlled experiments. Those experiments establish bounded evidence; they do not define a second product architecture. A deployed Mac can participate under the same worker contract as another supported computer, without acquiring permanent control-plane, main-model or verifier authority. ```mermaid flowchart TB C[Client application] --> H[Hive: request and execution authority] H --> A[Assigned hub: admission and placement] A --> W[Authorised worker pool] W --> E[Results and execution evidence] E --> H H --> R[Client response and contribution view] ``` The pilot currently has one logical management hub. The four planned regional zones are not deployed federation. They must not be counted as active execution sites. ## Model and state ownership Hive resolves a workload manifest to an explicit placement plan. The manifest identifies model revision and hashes, tokenizer, quantisation, supported operators, backend and numerical contract, context and output budgets. Admission budgets weights, KV state, activations, scratch memory and loading peaks for each required participant. The plan selects a complete-model worker or compatible required partitions. Optional draft assistance is a separate route. A verifier, when needed, is an assigned execution role on an eligible worker; it is not permanently identified with a Mac. Capacity expansion through required partitions remains a primary objective. A working draft route must not substitute for that objective or be reported as pooled model memory. Hive authorises the execution data path. That path may use an assigned hub relay or authenticated peer channels with measured reachability and policy. Hive ownership does not require every tensor or token to traverse a global server. Link evidence must describe the actual worker-to-worker or relay path, not only a heartbeat to Hive. ## Implemented boundary The Hive pilot implements association, inventory, policy, telemetry, bounded CPU benchmark dispatch and a separate persistent complete-model LLM lifecycle. Fixed authenticated worker endpoints deliver a registered model, lease generation, accept incremental output and preserve terminal outcomes. Android/iOS 0.2.28 adapters compile; a separate Linux native worker completed real Hive requests without Mac inference or relay. [Implementation, limits and measured evidence](https://github.com/sev7enITA/NodeLift/blob/main/docs/HIVE_LLM_ENGINE.md). This is not required-model partitioning. Direct enrollment alone does not authorize generation: the matching model must be prepared and the worker must be ready, explicitly enabled and permitted by device constraints and operating policy. The existing local LLM harness loads a target and verifies optional phone proposals. Small-model required-partition controls have separate physical evidence. These are migration inputs and historical test tools, not a supported alternate product route that requires a user's Mac. ## Migration work and acceptance gates | Work | Required change | Acceptance evidence | | --- | --- | --- | | Request ownership | Add a scoped Hive LLM request lifecycle alongside the existing CPU job contract | Submit, stream or retrieve, cancel and inspect a request without a local dashboard or local host session | | Placement | Use model/runtime identity, application memory budgets, required state and measured links | Reject incompatible or insufficient workers explicitly; no admission from summed RAM or SGEMM rate alone | | Worker runtime | Expose portable, bounded execution adapters through direct hub-authenticated sessions | A non-Mac worker executes an accepted model request; the test Mac need not run the target or relay the job | | Required partitions | Bind required participants and tensor/state assignments to one plan and fenced attempt | Demonstrate resident allocations, correct output and the declared stop/recovery contract on required-peer loss | | Optional assistance | Assign drafts and verification through the same Hive request authority | Account for accepted, rejected, expired and cancelled work; preserve the declared verification contract | | Interface | Make Hive the entry for pairing, model selection, submission, activity and contribution | No instruction to pair with a Mac to unlock the product's LLM service; missing implementation remains explicit | | Availability and failure | Preserve association and enforce policy, battery, thermal and OS lifecycle constraints | Disconnect or stop the development Mac; direct worker management and admitted execution follow their independent contracts | | Performance and energy | Attribute timing, transfers, useful work and measured/estimated energy to one request and plan | Compare matched complete runs; retain missing sensors, failures, discarded work and uncertainty | Changing a companion message, enabling a switch or exposing a new endpoint does not satisfy these gates. Physical Linux end-to-end Hive execution has been demonstrated for one co-located small-model pilot. This does not qualify mobile execution, WAN placement, model partitioning or a general performance benefit. [Architecture](https://github.com/sev7enITA/NodeLift/blob/main/docs/ARCHITECTURE.md) · [Cloud management](https://github.com/sev7enITA/NodeLift/blob/main/docs/CLOUD_MANAGEMENT.md) · [Engineering status](https://github.com/sev7enITA/NodeLift/blob/main/docs/ENGINE_STATUS.md) · [Hive console](https://github.com/sev7enITA/NodeLift/blob/main/docs/HIVE_CONSOLE.md)