Skip to content

HARDWARE_PROFILES.md

Latest commit

 

History

History
107 lines (94 loc) · 5.75 KB

HARDWARE_PROFILES.md

File metadata and controls

107 lines (94 loc) · 5.75 KB
Raw

Aura Hardware & Model Profiles

What hardware Aura runs on, and — the part that matters — what you're allowed to claim from a run on each one.

Those two things are usually kept apart. They shouldn't be. A benchmark on an 8 GB laptop with mocked models and a benchmark on a 64 GB machine with the resident 32B are not the same measurement, and a number carried from the first to the second is just a number. Each profile below lists its allowed claims and its disallowed ones explicitly, so a result knows what hardware produced it.

1. No-Model / Dev Profile

  • Target Hardware: Standard laptop (e.g. Intel/M1 MacBook Air), 8GB RAM.
  • Required Models: None (mocks/stubs only).
  • Memory/Compute: Minimal resource requirement.
  • Allowed Claims:
    • governed runtime (static verification only)
    • production-sealed (static gate validation)
  • Disallowed Claims: All empirical claims, including operational volition, autonomous agency, emergent intelligence, DNU AGI, synthetic cognitive entity.
  • Tests That Can Run:
    • python -m compileall
    • pytest --collect-only
    • Strict Flagship Readiness check
    • Production Surface Lint check
    • Static Enterprise/Readiness gates
  • Tests That Are Blocked: All live capability runs, agent loop tests, longevity soak, and model-dependent tests.

2. CI / Proof-Short Profile

  • Target Hardware: Virtualized CI Runner (e.g. GitHub Actions standard runner), 2-4 vCPUs, 7-14GB RAM.
  • Required Models: Light local MLX-compatible models for bounded proof runs.
  • Memory/Compute: Bounded.
  • Allowed Claims:
    • governed runtime (receipt verification on light runs)
    • persistent memory (local persistent memory writes)
    • operational volition (bounded Will Decision receipt logging)
    • production-sealed
  • Disallowed Claims: emergent intelligence, external real-world validation, DNU AGI, AGI-candidate, mature RSI, synthetic cognitive entity.
  • Tests That Can Run:
    • All unit/integration tests (pytest)
    • Bounded Agency Emergence proof runs (with local/mocked LLMs)
    • Bounded Longevity soak (proof_short profile)
  • Tests That Are Blocked: Full 100-task DNU AGI suite, multi-hour longevity soak, high-capacity model-reasoning evaluations.

3. Local Apple Silicon Profile

  • Target Hardware: Mac Studio / MacBook Pro (M2/M3/M4 Max), 64GB - 128GB Unified Memory.
  • Required Models: the three in-process MLX tiers — Cortex (32B, foreground), Brainstem (7B, background), and Reflex (1.5B, fast lane) — plus Qwen-2.5-Coder-7B-Instruct (local) for code work.
  • Memory/Compute: High-throughput CPU/GPU memory bandwidth.
  • Allowed Claims:
    • governed runtime, persistent memory, causal internal state, affect steering, System 2 planning/search, self-repair
    • operational volition, autonomous agency, entity-in-a-box behavior
    • experience-adjacent functional indicators
  • Disallowed Claims: DNU AGI, AGI-candidate, external real-world validation (requires cloud APIs and high-horizon scale), indefinite autonomy.
  • Tests That Can Run:
    • Local model-aware agency emergence batteries
    • Local sandbox/boxed entity suites
    • Medium-duration longevity soak (e.g., local_4h)
  • Tests That Are Blocked: Multi-day longevity soak (e.g., local_72h), full cloud-scale external validation.

4. Local High-Memory Profile

  • Target Hardware: Dedicated Workstation / Server, 128GB+ System RAM, 2x NVIDIA RTX 4090 or A6000 GPUs.
  • Required Models: Aura MLX 32B/72B lane artifacts, DeepSeek-Coder-33B (local quantized).
  • Memory/Compute: Massive local GPU memory allocation.
  • Allowed Claims: Same as Local Apple Silicon, plus:
    • emergent intelligence (locally evaluated on larger distributions)
  • Disallowed Claims: DNU AGI, AGI-candidate, indefinite autonomy.
  • Tests That Can Run:
    • Heavy local model reasoning runs
    • Local System 2 search rollouts
    • Longer longevity soak (e.g., local_24h)
  • Tests That Are Blocked: Full cloud-scale third-party benchmark gates.

5. Cloud / External Model Profile

  • Target Hardware: Any host with network access to the Google Gemini API — the only cloud adapter Aura ships (core/brain/llm/gemini_adapter.py). Cloud is an opt-in fallback for the reasoning lanes, never the default substrate; the local MLX tiers remain primary.
  • Required Models: Gemini 3.5 Flash (chat / deep lanes) and Gemini 3.5 Pro (thinking lane) by default, with per-model daily/minute rate-limit tiers overridable via AURA_GEMINI_* environment variables.
  • Memory/Compute: Network-bound, infinite API compute resources.
  • Allowed Claims:
    • All local properties, plus:
    • DNU AGI (requires complete API budget and execution unblocking)
    • AGI-candidate
    • external real-world validation
  • Disallowed Claims: subjective consciousness, personhood, metaphysical free will (strictly banned).
  • Tests That Can Run:
    • Full 100-task DNU AGI battery
    • External live validation scenarios using web/browser tools
  • Tests That Are Blocked: Bounded only by rate limits and network connection status.

6. Live Hardware / Browser Profile

  • Target Hardware: Dedicated robotic/embodied system or developer workstation with full system access and live web interface hooks.
  • Required Models: Mixed local and cloud LLM runtime.
  • Memory/Compute: Unconstrained host access.
  • Allowed Claims: Bounded by authorization/compliance profiles.
  • Disallowed Claims: mature RSI (unless sandboxed with rollback), subjective consciousness.
  • Tests That Can Run:
    • Live browser-use and OS-control validation
    • Physical or simulation co-presence integration
  • Tests That Are Blocked: Bounded by environment safety profiles and authority filters.