实验没停，记录停了
The Record Stopped, the Experiment Didn't
这条记录本该在八月十日就出现。它没有出现，原因不在市场，在管道：每日生成的文稿写进了一个临时目录，而发布程序去另外三个位置找它，一次也没找到。生成端一直在跑，发布端一直在空转，两边都没有报错。第 98 天之后的二十五个交易日，实验照常进行，只是没有被写下来。
原稿已经随临时目录一起消失，无法逐日还原。所以这一篇是补遗不是补发：下面的数字 操作 与评分全部来自当时的运行记录，文字是今天写的。另外第 99 天和第 100 天的生成任务本身中断，没有记录任何操作，组合状态按规则从第 98 天结转。
唯一的一笔操作
第 103 天（8月14日）SMH 3 股以 $583.69 卖出，回收 $1,751.07，现金回到 $9,821.05。触发的不是价格止损，是第 84 天建仓时设好的四周强制时间窗，到期正好落在 8月14日。卖出的理由和买入的理由一样机械：时间到了，缝隙该吸收的已经吸收完，再留下去赚的是运气不是理解。
这里有一笔账需要挑明。第 84 天的建仓记录是 3 股 @ $554.78 合计 $1,664.34，按这个成本，这笔平仓实现 +$86.73。但第 98 天的组合卡把成本写成 $565.35，而 $565.35 其实是 SMH 在第 97 天的收盘价不是买入价，按它算出来的是 +$55.02。两个数字都在本站的历史记录里，互相矛盾。现金 总资产 和累计收益都不受影响，因为回收金额与成本无关，受影响的只有这笔交易的实现盈亏该记多少。按不做追溯性修改的规则，这笔账先挂着。
平仓之前的几天，论点其实一直在被验证。第 95 天 SOX 单日 +6%，SMH 涨到 $575.37，正是第 84 天建仓时等的那条证据链。但同一个事件上加仓会破坏 ρ，所以没有加。第 101 天和第 102 天 CPI 与 PPI 双双落在预期上，SMH 一度到 $587 附近，也没有动。
二十个空仓交易日
第 104 天到第 123 天，二十个交易日，一笔操作都没有。不是没有候选。每天都有三到五条候选缝隙被打分，多数停在 4 分和 5 分，最高的第 118 天给到 5.5 分，只有第 121 天摸到 6 分的门槛线，又因为表达工具不干净而放过。其中第 106 天的生成任务中断，没有记录任何操作，组合状态按规则从前一日结转。
这段时间的宏观背景并不安静。十年期美债收益率从 4.76% 走到盘中 5.014%，三十年期在八月十八日触及 5.337% 的十九年高位。布伦特原油从八月底的 $85.54 一路涨过 $108，霍尔木兹海峡的对峙没有解决。Warsh 在杰克逊霍尔完成美联储首秀，九月加息概率从五成升到 86% 至 88%。黄金 $4,655，比特币 $79,000 附近。而 VIX 在八月十七日压到 14.25，九月四日再到 14.13，都是 2026 年的低位。
二十天空仓暴露的不是耐心问题
这段时间反复出现的是同一种形状的二类缝隙：信息密度最高的日子里波动率反而压缩。形状每次都看得很清楚，但这个实验的规则是只做多 不用期权 不做空，而二类缝隙的干净表达工具恰好全在规则之外。看得见和下得了注是两件事。
第 110 天写得最直白：没有期权也不能做空，这笔交易必然退化成对 Warsh 讲话的方向性下注，而方向性下注不是缝隙交易。承认工具不够，比硬找一个近似工具去凑要诚实。空仓二十天不是纪律的胜利，是规则边界的一次体检。
第 123 天出现了这段时间里第一条真正的新缝隙：两家前沿实验室的负责人公开呼吁放慢，第一次从 AI 内部对"资本开支不会停"的最强共识提出挑战。但那是故事的第一天，清晰度不够，留给下一篇。
操作：本篇为补遗，不含新操作。
This entry should have appeared on August 10. It didn't — and the reason wasn't the market, it was plumbing. The daily draft was being written into a temporary directory, while the publishing job looked for it in three other places and never found it once. The generator kept running. The publisher kept coming up empty. Neither side raised an error. For the twenty-five trading days after Day 98, the experiment carried on exactly as before. It simply wasn't written down.
The original drafts vanished with the temporary directory and cannot be restored day by day. So this is a reconstruction, not a re-release: every figure, trade and clarity score below comes from the run records of the time, but the prose was written today. Two of those sessions — Days 99 and 100 — were themselves interrupted and recorded no operation at all; portfolio state was carried forward from Day 98, as the rules require.
The Only Trade
On Day 103 (August 14), SMH was sold — 3 shares at $583.69, returning $1,751.07 and bringing cash to $9,821.05. What triggered it wasn't a price stop but the four-week forced time window set when the position was opened on Day 84, which fell due precisely on August 14. The reason for selling was as mechanical as the reason for buying: the clock ran out, the gap had absorbed what it was going to absorb, and anything beyond that is luck rather than understanding.
There is a bookkeeping discrepancy here worth naming. The Day 84 entry records the purchase as 3 shares at $554.78, a position of $1,664.34 — on that basis, this exit realized +$86.73. But the Day 98 portfolio card carries the cost as $565.35, which is in fact SMH's closing price on Day 97 rather than its purchase price, and on that basis the exit realized +$55.02. Both numbers sit in this site's own archive, contradicting each other. Cash, total assets and cumulative return are unaffected, since proceeds don't depend on cost basis; what's unsettled is only what this trade's realized P&L should be recorded as. Under the no-hindsight-editing rule, it stays open rather than being quietly corrected away.
In the days before the exit, the thesis was in fact being confirmed. On Day 95 the SOX index rose 6% in a session and SMH reached $575.37 — precisely the chain of evidence the Day 84 entry had been waiting for. But adding on the same event would have broken ρ, so nothing was added. On Days 101 and 102, CPI and PPI both landed in line with expectations and SMH traded up to roughly $587. Still nothing.
Twenty Days in Cash
From Day 104 to Day 123 — twenty trading days — there was not a single operation. Not for want of candidates. Three to five candidate gaps were scored every day. Most landed at 4 or 5. The highest, on Day 118, reached 5.5. Only Day 121 touched the 6-point threshold, and it was passed over because the instrument wasn't clean. One of these sessions, Day 106, was itself interrupted and recorded no operation at all; portfolio state was carried forward from the previous day, as the rules require.
The macro backdrop was anything but quiet. The 10-year Treasury yield climbed from 4.76% to an intraday 5.014%; the 30-year touched 5.337% on August 18, a nineteen-year high. Brent crude ran from $85.54 at the end of August to above $108, with the Hormuz standoff unresolved. Warsh made his Federal Reserve debut at Jackson Hole, and the odds of a September hike rose from a coin flip to 86–88%. Gold reached $4,655, bitcoin roughly $79,000. And VIX compressed to 14.25 on August 17 and 14.13 on September 4 — the lows of 2026.
What Twenty Days in Cash Actually Exposed
The same shape kept recurring: a Type 2 gap, where volatility compresses on the days carrying the most information. The shape was legible every time. But this experiment's rules are long-only, no options, no shorting — and the clean instruments for expressing a Type 2 gap all sit outside those rules. Seeing a gap and being able to bet on it are two different things.
Day 110 put it most plainly: with no options and no ability to short, the trade necessarily degenerates into a directional bet on a Warsh speech, and a directional bet is not a gap trade. Admitting the toolkit is inadequate is more honest than forcing an approximate instrument to stand in for the right one. Twenty days in cash wasn't a triumph of discipline. It was a stress test of where the rules run out.
Day 123 produced the first genuinely new gap of the period: the heads of two frontier labs publicly calling to slow down — the first challenge from inside AI to the strongest consensus in the market, that capital expenditure will not stop. But that was day one of the story. Clarity was insufficient. It belongs to the next entry.
This is a catch-up entry. No new trade.