9. Execution and Performance - Execution Modes and Performance Targets¶
9.1 Mode 1 — Single Node¶
- Semantic Core only.
- Fast.
- Minimal overhead.
9.2 Mode 2 — ELM-Assisted Mode¶
- Semantic Core + one or more role-based or domain-specific experts.
- Typically sequential or lightly parallel execution.
- Used when specialist assistance improves quality without full swarm overhead.
9.3 Mode 3 — Swarm Mode¶
- Multiple nodes execute.
- Weighted consensus or arbiter-mediated synthesis.
- Reputation-based selection.
9.4 Mode 4 — Agent Mode¶
- Multi-step execution involving memory, grounding, verification, and optional tool use.
- Secure tool path enforced through the Tool Intermediary.
- Used for higher-complexity workflows.
9.5 Execution Strategy Principles¶
- Local-first, distributed-second: Execute locally when possible and escalate only when justified.
- Small effective cognitive sets: Select the smallest effective set of ELMs and services rather than activating a large swarm by default.
- Roles over raw scale: Prefer assigning the right expert role over scaling one model indiscriminately.
10 Performance Targets¶
Metric
Target
Tokens/sec
≥ INT4 baseline where comparable
Memory usage
≤ practical low-bit deployment envelope
Grounded quality
≥ baseline single-model factual reliability
Verification / formatting quality
measurable improvement
Multi-node scaling
near-linear up to initial swarm targets where network conditions permit
Tool safety overhead
bounded and separately reported from inference latency
Customization efficiency
choose the lowest-cost path among retrieval, memory, private ELM invocation, and swarm consensus that still satisfies quality and policy requirements