- Status
- RESEARCH
- Date
- 2026-08
- Field
- inference · model editing · benchmarks
Summary
Hybrid SSD/attention inference with prefix-state reuse, parity-gated benchmarks, a cryptographic neuron catalog, and surgical weight editing with hash-verified rollback.
Facts
- Runtime
- PyTorch, local, 2 consumer GPUs
- Method
- Parity-gated benchmarks
- Editing
- ROME / MEMIT with hash-verified rollback
Tags
inferencemodel editingbenchmarks
