High-performance kernel and self-tuning engine for StarCraft II AI, with generic multi-language adapters.
SC2AIOptimize exists to make expensive SC2 AI computation cheaper only when measurement proves the optimisation is worth using. The baseline implementation is always a valid candidate; an optimisation that does not produce positive measured net benefit is rejected or disabled.
SC2AIOptimize is the public, generic optimisation engine. It supports public or private consumers through bot-neutral adapters and the shared integration contract.
Private consumer repositories, private integration/orchestration layers and private bot topology are intentionally not enumerated or linked from this public README. That information belongs in the private consumer repositories. This keeps SC2AIOptimize reusable and policy-neutral while avoiding unnecessary disclosure of private integration design.
The public project must not acquire bot-specific strategy, private orchestration policy, private map/spatial authority or private training/runtime ownership merely because a private consumer uses it.
The 32-slice V1 software programme is 32/32 implemented. The public API contract is 1.1 with status v1-candidate: component/build/toolchain proof is in place, while representative target-hardware, real bot/framework integration, behavioural, ladder/runtime and BotOps release evidence remain separate certification gates.
API 1.1 is an additive extension of 1.0 that adds the shared bot-neutral integration compatibility contract. Existing 1.0 consumers remain valid.
Production seed policies and multi-language edge adapters are maintained as post-programme V1 polish. They are defaults and optional integrations, not bot-performance limits.
See docs/STATUS.md and docs/V1_CERTIFICATION.md for the exact proof boundary.
The runtime is deliberately small:
- SC2Registry — immutable, build-pinned SC2 facts. Seeded from MicroMachineEvo v0.9.2Z / Base75689 data, with provenance retained.
- Kernel — framework-independent compute.
Hotcontains tiny proven hot-path primitives; larger compute belongs in Kernel only when profiling justifies the native boundary. - Tuner — cheap selection and tuning. Baseline/champion/challenger, confidence and context live here; expensive discovery is not paid on every hot call.
- Adapters — direct generic C++, optional Python buffer binding, optional stable C ABI, and thin Rust/Go/.NET/Java/Node edges. Consumers adapt to SC2AIOptimize; SC2AIOptimize does not depend on a bot framework.
- Shared integration contract — one startup compatibility authority for API generation, neutral view ABI, pinned registry identity and generic capabilities, exposed through the language edges rather than duplicated per bot.
The stable C ABI allows older native consumers such as a C++14 bot to use the shared connector without importing the C++20 public surface. The Python edge exposes the same compatibility authority for Python consumers. A non-ready handshake keeps the consumer on its baseline/OFF path; optional loading and package-presence fallback remain consumer/deployment responsibilities.
The core has no dependency on SC2API, python-sc2, MetaSwarm, MicroMachineEvo, BotOps or an ML runtime.
The V1 engineering priors are intentionally conservative and consumer-overridable:
- operational evidence: 20 samples
- certification evidence seed: 50 samples
- default promotion margin: 2% and 1 microsecond per declared representative workload
- confidence separation: z = 1.96
- tuner/control overhead ceiling: 1% and 50 microseconds
- workload seeds: 8, 16, 32, 64, 128, 256, 512, 1024
- batch seeds: 16, 32, 64, 128, 256, 512, 1024
- spatial-cell seeds: 2, 3, 4, 6, 8, 12
These values are starting priors, not universal truths or hard ceilings. A consumer may tighten, relax or replace them when representative evidence justifies it.
The compute-budget governor is OFF by default. A consumer may explicitly delegate a full compute envelope; the helper then seeds resume/soft/hard hysteresis at 55% / 70% / 95%. SC2AIOptimize never invents or silently imposes a bot compute budget.
Correctness first. Measure total cost. Keep an optimisation only when its net benefit is positive.
Total cost includes adapter, dispatch and bookkeeping overhead. OFF / baseline is a first-class champion.
See docs/ARCHITECTURE.md, docs/ADAPTERS.md, docs/PROGRAMME.md, docs/TUNER.md, Adapters/AIArena/README.md and docs/MMEVO_Z_MIGRATION.md.