TraceArena v0.1.12 — Investment Agent Benchmark v1
TraceArena now ships a reproducible evaluation contract for its flagship capital-market world.
What is new
- Investment Agent Benchmark v1 contract baseline with a checked-in JSON report and readable leaderboard.
- Evidence → decision → simulated order → deterministic settlement evaluation across return, benchmark excess return, maximum drawdown, turnover, and transaction costs.
- Integrity verification for fixture, full replay, semantic replay, and generated report.
- Cross-platform reproduction in CI on Ubuntu and macOS.
- English and Chinese benchmark cards linked from both READMEs.
Reproduce it without an API key
Fastest path: open the pinned Colab notebook. It runs v0.1.12 and verifies the generated report against the published baseline.
PYTHONPATH=backend python backend/scripts/investment_benchmark.py \
--output runs/investment_agent_benchmark_v1The run disables network access and brokerage execution. It uses a synthetic fixture and a simulated portfolio ledger.
Important claim boundary
The two bundled entrants are deterministic scripted controls. This release proves the benchmark contract; it is not yet a model leaderboard. A model result will only be promoted when its provider/model identity, fixture and scenario digests, seeds, role rotation, complete replay, and verifier output are available.
The synthetic fixture remains marked internal_synthetic_pending_approval, and its proposed CC0-1.0 dedication is still pending maintainer confirmation. The report exposes this status explicitly.
Evidence bundle
Download tracearena-v0.1.12-investment-benchmark-evidence.zip to inspect the bilingual benchmark cards, report, leaderboard, run manifest, summary, and authoritative replay.
SHA-256: 353026af03de551020ce927a0579ebcebd1f9836772a35127ee5d30ff426496b
No financial advice. No real orders. No brokerage connection.
本版本发布首个投资智能体评测协议基线：可离线复现、可校验、可回放，并同时报告收益、超额收益、最大回撤、换手率与交易成本。当前两名参赛者是脚本控制组，不是模型排行榜；网络和券商均禁用，不构成投资建议。