submission.env.example

# ─────────────────────────────────────────────────────────────────────────────
# Submission runner config. Copy to `submission.env`, then run:
#
#   Model API in Docker:  make runner-pull && make preflight && make run
# (submission.env is gitignored — it may hold your API key; never commit it.)
# ─────────────────────────────────────────────────────────────────────────────

# ── REQUIRED ─────────────────────────────────────────────────────────────────
# Model identity recorded in the submission. It must be LiteLLM-routable:
#   anthropic/claude-...   openai/gpt-...   openrouter/...   gemini/...
#   openai/<model> together with API_BASE for an OpenAI-compatible endpoint
MODEL_NAME=openai/gpt-5.4
# API_KEY=...        # generic key for the model API endpoint
# Prefer the provider's own env var? Use it INSTEAD of API_KEY, e.g.:
#   OPENAI_API_KEY=...   ANTHROPIC_API_KEY=...   OPENROUTER_API_KEY=...   GEMINI_API_KEY=...

# ── MODEL API (OpenRouter / gateway / local vLLM / Azure …) ──────────────────
# Nothing is hardcoded to anthropic/openai. Reach any model with these:
# API_BASE=https://my-gateway.example/v1
# API_KEY=...          # generic key when no PROVIDER_API_KEY name fits
# MSWEA_COST_TRACKING=ignore_errors
#   # ^ required when your model isn't in litellm's price map (typical for custom
#   #   endpoints) — mini-swe otherwise aborts the session on the cost-calculation
#   #   error. Token usage is still recorded from API responses.

# ── FROZEN LIMITS ────────────────────────────────────────────────────────────
# There are no budget environment variables. Benchmark runs always use
# Source-only Top50 triage, then 50 isolated rule episodes with
# M=10, E=12, P=24, P_solve=10, and exactly S=2 submit attempts per rule.

# ── OUTPUT / LOGS ────────────────────────────────────────────────────────────
# Docker defaults follow.
# OUTPUT=/out/submission.json  # authoritative submission output
# Benchmark runs reject custom prompts, strategies, backends, and logical budgets.

# Version pins, source manifest, and the trusted `pred` path are baked into the image.
# Benchmark runs reject pin overrides and custom `pred` binaries.

# ── METADATA ─────────────────────────────────────────────────────────────────
# SUBMITTED_BY=your-hf-handle   # also fillable at submit time on the Space