Skip to content

CLAUDE.md

Latest commit

 

History

History
72 lines (46 loc) · 4 KB

CLAUDE.md

File metadata and controls

72 lines (46 loc) · 4 KB
Raw

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Environment

All CUDA-Q and numpy work must use the dedicated venv:

/home/diamondnode/venv312/bin/python

System Python lacks cudaq, numpy, cupy, scipy, and jax. Key packages in the venv: cudaq 0.14.2, numpy 2.4.4, cupy-cuda13x, jax/jaxlib 0.10.0, scipy 1.17.1.

Hardware: NVIDIA GTX 1650, 4 GB VRAM. Thermal threshold 89.6 °C (see bench_gpu_telemetry).

Ollama is at /usr/local/bin/ollama with llama3.2:3b (2 GB) loaded.

Path overrides via env vars: DIAMOND_STATE_DIR, DIAMOND_LOG_DIR, DIAMOND_VENV.

Commands

Run one QUBO iteration (uses venv python internally):

/home/diamondnode/venv312/bin/python scripts/mycelial_qubo.py [--shots 512] [--outer-rounds 3] [--json]

Run the full benchmark suite (saves report to reports/):

/home/diamondnode/venv312/bin/python scripts/benchmark.py [--suite all|qubo|llm|gpu|cudaq] [--json]

Daily health check (GPU telemetry + ttectra + ollama status):

/home/diamondnode/venv312/bin/python scripts/daily_health.py

LLM interpretation of current network state (appends to logs/llm-interpretations.jsonl):

/home/diamondnode/venv312/bin/python scripts/llm_interpret.py

CUDA-Q smoke test (256 shots of a Bell circuit on qpp-cpu):

/home/diamondnode/venv312/bin/python scripts/_cudaq_probe.py

Architecture

Core simulation — mycelial_qubo.py

Models fungal hyphal-network growth as a QUBO over edge variables on a 16-node 4×4 grid. The QUBO diagonal encodes three competing terms: distance penalty (lam_dist=0.4), redundancy reward for already-active edges (lam_redund=-0.2), and resource-proximity reward for edges incident to resource nodes 0 and 15 (lam_resource=-0.8).

Solving is split into subspaces of 10 edges each. Each subspace is solved by a 1-layer QAOA circuit via cudaq.sample (gamma=0.3, beta=0.2). Subspace assignments are stitched back into the global edge vector, and the best global assignment is kept across outer_rounds.

Double-loopback resilience:

  • Inner loopback: if cudaq.sample raises, retries once; falls back to random assignment on second failure.
  • Outer loopback: whenever a candidate energy beats state.best_energy, state is checkpointed to state/mycelial_checkpoint.json. This checkpoint is the recovery target if outer rounds diverge.

State persistence

MycelialState (dataclass) is serialised to state/mycelial_state.json after every iteration and to state/mycelial_checkpoint.json on any new best energy. The state carries active_edges, energy_history, best_energy/best_edges, node positions, and a full run_log.

Benchmark runner — benchmark.py

Five benchmarks (gpu_telemetry, cudaq_probe, qubo_iteration, state_persistence, llm_latency) each return a uniform dict {name, passed, duration_s, value?, unit?, threshold?, notes}. Results are saved as JSON to reports/benchmark-<UTC timestamp>.json. Exit code is 0 only when all selected benchmarks pass.

LLM interpretation — llm_interpret.py

Reads state/mycelial_state.json, builds a one-sentence summary (node count, active edges, best energy, trend, resource nodes, iteration), and sends it to ollama run llama3.2:3b asking for a 3–4 sentence assessment and a single parameter-change recommendation. Output appended as JSONL to logs/llm-interpretations.jsonl.

Arena integration cross-ref (gc-arena-controller + OFFLOAD): High-value Ouroboros matches (from gc-arena-controller 5-phase FSM on shared D1) can participate in the H(s) > 8.5 OFFLOAD → notion-bridge soul-capsule (DB 21e41606...) → diamondvault → gc-mcp propagation story via the live /v1/arena/matches/:id/offload (or future gc-mcp-beta arena tools). See top ~/AGENTS.md "Arena + Conductor Soul-Capsule Integration", gc-arena-controller/README.md, and Task 4. Arena matches provide richer context (phase history, strategy, bench_trace, demand_score) than raw VRAM telemetry.