Documentation · H33-HIVE

H33-HIVE developer documentation

This is the documentation map for H33-HIVE, H33’s AI execution infrastructure. Each section below states what it will cover and where it stands today.

Status No public HIVE API, SDK or CLI exists yet. Schemas and examples are published here only after they have been checked against the canonical Rust implementation, so this page lists no endpoints, methods or install commands.

Overview and architecture

HIVE routes eligible AI work across models and hardware, reuses independently verified results through Cachee, recovers from provider failures without repeating consequences, and keeps consequential execution inside the boundary Agent-008 enforces. The full design and its readiness are on the technical architecture page.

Workload and execution-context schema

Will define how a unit of work is described: its input, task category, tenant, policy version and an opaque execution-context reference issued at admission. In the research build that reference is carried unchanged and never parsed by HIVE. Not yet published.

Model capability profiles

Will define the per-model, per-category measurements routing uses: success rate, independent-verifier acceptance, false accepts and mean compute cost, with the minimum sample counts below which a profile is not trusted. Not yet published.

CPU/GPU scheduling policy

Will define execution classes, batching, queueing, cold-start accounting and cost-based placement. The scheduler is planned and has not been benchmarked end to end. Not yet published.

Cachee verified-reuse contract

Will define the reuse key (input commitment, tenant, model generation, policy version, verifier identity and version, authority epoch), the admission rule (only an independent verifier’s acceptance of exactly the delivered answer) and the fail-closed refusal reasons. The research adapter implements these against a stand-in; the contract is published once it is aligned with Cachee’s canonical interface. Not yet published.

Provider retry and idempotency contract

Will define when an attempt may be retried: a possibly-executed attempt is re-sent only with an idempotency key, to the same provider, when that provider declares server-side idempotency; otherwise the task settles Unknown. Not yet published.

Agent-008 admission integration

Will define how work is admitted under Agent-008 authority and how the resulting execution context is bound. The research build uses a placeholder at this seam, which is not a security control. Not yet published.

Consequential-action enforcement

Will define how a proposed action is checked at the execution boundary by H33-API-G under Agent-008 authority, with keys held in H33-Key custody. Not yet published.

Evidence and Replay

Will define the evidence a HIVE execution produces and how it is anchored with H33-74 and reconstructed with Replay. Not yet published.

Benchmarking and qualification

Will define the methodology every published figure must carry: hardware, model revision, concurrency, latency percentiles, throughput, independently verified correctness, repeat-rate assumptions, energy and cost against a no-HIVE baseline. Not yet published.

Failure modes and limitations

Will list known failure modes and their trade-offs, including the circuit breaker’s slower detection at low traffic and slower per-scope recovery. A current summary is on the architecture page.