.env.eval.example

# Copy to .env.eval and fill in secrets. Not committed to git.

# --- transport ---
# vast  = rent/reuse vast.ai GPUs (default; all vast_eval.py logic active)
# ssh   = fixed bare-metal box (EVAL_SSH_HOST); vast.ai not contacted
EVAL_TRANSPORT=ssh
EVAL_SSH_HOST=91.224.44.227
EVAL_SSH_PORT=50200

# --- vast.ai (used when EVAL_TRANSPORT=vast) ---
# VAST_API_KEY=
# VAST_INSTANCE=42682383   # required: existing instance id (--reuse); no auto-rent

# --- bot ---
# GH_TOKEN=
# FRONTIER=285
# CEILING=366
# BIDIR=1          # Qwen3.5 + Qwen3.6 bidirectional eval (default; set 0 for legacy single-model)
# TRIPLE=1          # legacy alias for BIDIR=1
# PRIMARY_QUANT=Q4_K_M   # Qwen3.5 quant: Q4_K_M | Q8_0 | BF16
# POLARIS=1         # Polaris TDX verifiable receipts (default; set 0 to disable)
# POLARIS_API_KEY=  # TDX attest (https://polaris.computer); preferred when available
# SPARKINFER_POLARIS_PRIVATE_KEY=  # base64 Ed25519 fallback when TDX is down
# POLARIS_API_BASE=https://polaris.computer

# --- Qwythos / Qwen3.5-9B (triple-eval primary) ---
# QWYTHOS_REPO=empero-ai/Qwythos-9B-Claude-Mythos-5-1M-GGUF
# QWYTHOS_MODELS_DIR=/workspace/models35
# QWEN35_9B_Q4K_SHA256=   # pin after first download (sha256sum on eval box)
# QWEN35_9B_Q8_SHA256=
# QWEN35_9B_BF16_SHA256=
# QWEN35_9B_LLAMA_128=224.91    # llama.cpp decode tok/s @128 (RTX 5090, 2026-07-10)
# QWEN35_9B_LLAMA_512=225.10
# QWEN35_9B_LLAMA_4K=224.68

# --- optional ---
# HF_TOKEN=
# OPENAI_API_KEY=