Day 122: Fact Chains, Not News Streams — Builder's Log

Today was about tightening the investment system around source-of-truth checks and thesis-aware signals, so automation helps decisions instead of adding noise.今天的重点是把投资系统收紧到事实源校验和 thesis-aware 信号上，让自动化服务决策，而不是制造噪音。

Tightened the portfolio fact chain around Convex positions, so downstream reports and public views follow the same source of truth instead of drifting by file.
Reframed market-news monitoring from a faster news stream into a filter for signals that can change a holding thesis.
Simplified the NewsLiquid path from several overlapping monitors into three daily analysis windows tied to the active portfolio.
Split long podcast and reading outputs away from private chat, keeping Telegram for short pointers and Discord for the full archive.
Treated model upgrades as production changes, not trophies: stability won over novelty when the runtime showed sharp edges.
Pulled one useful investing lesson from the day: AI in research is strongest when it knows the thesis it is protecting, not when it reads everything.

今天把组合事实链重新收紧到 Convex position 上，让报告和公开展示跟同一个事实源走，不再靠多个文件各说各话。
新闻监控也重新定位：不是更快刷新闻，而是过滤出会改变持仓 thesis 的信号。
NewsLiquid 从几条重叠监控收敛成每天三次、围绕真实持仓的分析窗口。
播客和阅读的长内容从 Telegram 私聊拆出去，私聊只留短指针，全文进 Discord 存档。
模型升级按生产变更处理，不按“新模型更强”冲动推进；运行时有毛刺，就先保稳定。
今天最有用的投研判断是：AI 真正强的地方不是读得更多，而是知道自己在保护哪条 thesis。

Day 122: Fact Chains, Not News StreamsDay 122: 事实链，而不是新闻流