The eight proposed adaptive classes
1 Hardware Detection: measured CPU, RAM, storage, GPU and qualified VRAM. 2 Performance Optimization: choose available algorithms and chunk sizes. 3 GPU Adaptation: compatible kernels, precision and residency. 4 Memory Management: streaming, caching and bounded intermediate buffers. 5 Distributed Computing: explicit supported partitions dispatched to authorized physically distinct FarmNodes. 6 AI Acceleration: quantization, efficient batching and optimized inference. 7 Game Optimization: dynamic quality, texture level, resolution and offline asset processing. 8 NIKI Resource Architect: propose, test, and retain new execution solutions.
Device contribution and consent
Each device must present its real identity and capabilities; verify the owner, grant, licence, task ABI and local consent; enforce bounded CPU/GPU/RAM/storage quotas; preserve auditable start/stop and resource-use receipts; and exclude unrelated private user information from any use or trade. Merely advertising RAM must not create a participation asset or authorize execution. A network-attested physical machine is distinct from a logical node record.
What local tests show
Measured local CPU/RAM/storage selection, bounded workloads, safe streaming fallback, eligibility rejection, protection against claiming 24 GB GPU from 8 GB, and consent checks were tested in the 9/9 adaptive subsystem. FarmNode GUI/device/login integration reports 32/32 scoped tests passed, and an adaptive Windows executable built successfully. The tests deliberately prevent fabricated remote capacity from counting as a completed execution or paid contribution.
Qualification required
Before declaring operational distributed computing: enrol genuinely separate physical machines with trust anchors, authenticate shared ENTITY and Bridge authorizations, enforce resource and privacy policy on the actual operating systems, perform failover and latency tests, meter only executed work, reconcile usage with exchange service rights, and test payout settlement separately. Remote workloads do not add physical GPU RAM to the client and may be unsuitable for latency-sensitive graphics or live robotic control.
Sources, qualification and related records
FarmNode system guide · Unified account design.
Figures refer to scoped local Windows engineering logs as at 25 October 2026. A passed test count does not equate to measured customer deployment. Proprietary source, keys, model credentials, sensitive identity and private account information are not published.