Fleet template is now rebuilt only when deploymentVersion changes (~68 times per
28,800-tick run instead of every tick). Reuses module-level Maps, arrays, and
utilization objects instead of allocating new ones each tick. Replaces 4x
Object.values().reduce() with single-pass aggregation and sorts fleet in-place.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Adds a full simulation harness (game-simulation package) with greedy/random strategies,
36-metric diagnostics, multi-run orchestration via child processes, and a statistical
interpreter. Includes 2.3x engine performance optimizations (research bonus caching,
per-DC dirty tracking, reduced allocations in tick pipeline, single-pass loops).
Fixes a critical balance bug where training pipelines stalled on insufficient VRAM would
permanently block training slots — the engine never re-checked stalled pipelines, and the
greedy strategy didn't pre-check VRAM requirements. This caused 20-25% of seeds to get
stuck in Scale-up era. All three fixes (engine un-stalling, strategy VRAM pre-check,
stalled pipeline cancellation) bring pass rate from 75% to 100% across 20 random seeds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The old overload policy had dead controls (maxQueueDepth, rateLimitPerCustomer never read)
and trivial flat penalties. This replaces it with a full serving pipeline where deployed
models form a fleet, requests route through priority/degradation logic, and policy choices
create meaningful strategic tradeoffs.
New serving pipeline: fleet building from deployed models (size/quant/MoE multipliers),
demand categorization by 5 priority tiers, enterprise capacity reservation, priority-ordered
serving with overflow behaviors (queue/reject/degrade), auto-degradation to faster models
under load, and Batch API to fill idle capacity at discounted rates.
4 new research nodes gate features progressively: Intelligent Request Routing, Priority
Queue System, Request Batching, and Auto-Scaling. New dedicated Serving page with pipeline
metrics, model fleet utilization, and research-gated policy controls.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
josh