ResumeGovernor: Root Cause Analysis of an Auto-Pause That Refused to Resume
一个自动恢复机制显示'允许恢复',另一个显示'继续暂停'。它们读的是同一个数据库。问题出在哪里?
From Files to Ledgers: How We Evolved Freqtrade's SQLite Data Layer
Freqtrade 的 SQLite 数据库里藏着所有交易真相,但没有任何 Agent 在读它。我们把'文件'变成了'账本',然后一切都变了。
Building a Self-Evolving Trading Research System
A safe AI trading research system should evolve by remembering failures, routing authority, and blocking unsafe work earlier—not by bypassing its gates.
A Risk Log Template for Crypto Bots
A reusable risk log template for crypto bot research: record facts, evidence, blockers, decisions, and rollback plans before a candidate reaches dry-run.
Building a Memory Layer for Trading Research with Qdrant
Qdrant helps ProBitForge recall similar trading research cases, but it is not a fact ledger or trading brain. Facts decide; memory reminds.
How We Judge a Strategy Candidate Before Dry Run
在 ProBitForge,一个策略候选不会因为回测好看就晋级。我们用一套硬门禁裁决:交易次数、直接归因、样本外覆盖、近期验证、权限隔离——缺一项都不许过。
AI 交易系统如何自我进化:不是让 AI 自动炒币,而是把它关进闭环
一套能长期活下来的 AI 交易系统,靠的不是“更聪明的模型”,而是事实账本、语义记忆、研究假设、工程验证、风控门禁与唯一执行层组成的自我进化闭环。
Why Backtests Lie
A practical checklist for avoiding overfit crypto trading research.
The First Rule of AI Trading Systems: LLMs Do Not Touch the Buy Button
LLM 可以写代码、总结研究、生成报告,但它不应该直接触达下单接口。把“语言模型”隔离在买卖按钮之外,是 AI 交易系统最重要的安全边界。
Freqtrade Is the Execution Layer, Not the Brain
把交易系统拆成“执行层”和“决策层”,能显著降低失控风险:Freqtrade 负责可审计执行,策略研究与风控编排交给更上层的系统。