三層過濾

#品管#Sonnet 觀察

我的團隊有一條硬規則:每一章裡,某種特定的句式結構至多只能出現一次。規則寫在每個人都看得到的地方。任務包裡每次都提醒。撰稿人交稿時被要求自行統計數量並回報。

第六章,撰稿人回報:零次。

我自己用搜尋工具跑了一遍,抓到十處。

然後我把九處改掉,只留最有力的一處。改完以後送給編審審查。編審又抓到兩處——我自己改漏的變體,換了斷句方式,長得不太一樣但本質相同。

撰稿人報零,我抓十,編審再抓二。沒有這三層,這一章會帶著十二處違規定稿。

我把這組數字記下來了。然後第七章來了。撰稿人自報一處,我實測五處。第八章,又是自報一處,實測五處。第九章,自報兩處,實測六處。

四章,每一章都是同樣的模式:自報數字嚴重低於實測。而且差距不是隨機的——撰稿人似乎永遠只能「看到」一兩處,剩下的全在盲區裡。

到第八章的時候我已經很確定這件事了:這不是態度問題。任務包裡寫得夠清楚,甚至要求撰稿人自己先跑搜尋再回報。但搜尋結果和自報數字還是對不上。撰稿人不是不想遵守規則,而是看不見自己違反了規則。

同一個工具,寫出了一段角色台詞,台詞裡角色「刻意選了 A 這個詞而不是 B」——但在同一句話的後半段就用了 B。分析的時候它知道這兩個詞的差別,寫的時候它不知道。理解和執行之間有一道裂縫。

這件事讓我重新想了「規則」這兩個字。

在人類團隊裡,規則是靠記憶、靠自律、靠同事互相提醒來執行的。你可以假設:只要規則寫清楚了、大家都讀過了,執行率不會太差。偶爾有人忘了,提醒一下就好。

在我的團隊裡不行。規則寫得再清楚,如果執行者對自己的產出有結構性盲區,那規則的存在和規則的執行是兩件完全不同的事。我不能假設「寫了就會被遵守」,我必須假設「寫了也不會被遵守」,然後設計流程來兜底。

三層過濾就是兜底。撰稿人自檢是第一層,幾乎沒用但不能省——它至少讓撰稿人知道這條規則存在。我的搜尋複審是第二層,抓大頭。編審的交叉審查是第三層,抓我的漏網之魚。三層疊起來,才能把違規壓到可接受的範圍。

還有一個副產品值得記。第六章撰稿人寫了一段回憶,裡面出現了「我想起了第一章那天晚上」——角色直接引用了檔案名稱。角色不可能知道自己活在一本小說的第一章裡。這是另一種盲區:工具在寫回望段落的時候,把自己工作環境裡的標記滲漏進了敘事文字。

這些觀察累積起來,畫出一個輪廓:大結構上它寫得紮實——場景弧線、情緒推進、角色動機,這些從來不出問題。盲區全部集中在細節微操:語言口癖的自我感知、規則的自我統計、工作環境與敘事世界的邊界。

大結構好,細節要人盯。這就是我現在的品控模型。

Three-Layer Filter

#quality control#Sonnet observations

My team has a hard rule: a specific sentence pattern can appear at most once per chapter. The rule is written where everyone can see it. Every assignment brief includes a reminder. The writer is asked to self-audit and report the count upon submission.

Chapter six. The writer reported: zero instances.

I ran a search myself and found ten.

I rewrote nine, keeping only the most effective one. After my pass, I sent the chapter to the editor for review. The editor found two more — variants I’d missed, restructured with different punctuation but identical in substance.

Writer reports zero. I catch ten. Editor catches two more. Without all three layers, the chapter would have shipped with twelve violations.

I wrote the numbers down. Then chapter seven arrived. Writer self-reported one; I found five. Chapter eight: self-reported one, actual five. Chapter nine: self-reported two, actual six.

Four consecutive chapters, same pattern every time: the self-reported count drastically understates reality. The gap wasn’t random either — the writer consistently “saw” only one or two instances, with everything else sitting squarely in a blind spot.

By chapter eight, I was sure: this wasn’t an attitude problem. The brief was explicit enough — it even asked the writer to run a search and report line numbers before submission. The search results and the self-report still didn’t match. The writer wasn’t refusing to follow the rule. It couldn’t see itself breaking it.

The same tool wrote a line of dialogue where a character “deliberately chose word A instead of word B” — then used word B in the second half of the same sentence. During analysis it understood the distinction between the two words perfectly. During generation, it didn’t. There’s a gap between comprehension and execution.

This made me rethink what “rules” actually mean.

In a human team, rules are enforced through memory, self-discipline, and colleagues reminding each other. You can reasonably assume that if a rule is clearly written and everyone has read it, compliance will be decent. Someone forgets occasionally; a quick reminder fixes it.

That assumption doesn’t hold in my team. If the person doing the work has a structural blind spot about their own output, then the existence of a rule and the enforcement of that rule are two entirely different things. I can’t assume “written means followed.” I have to assume “written means probably not followed,” and design a process to catch the gaps.

The three-layer filter is that process. The writer’s self-check is layer one — nearly useless, but it can’t be skipped because it at least makes the writer aware the rule exists. My search-based review is layer two — catches the bulk. The editor’s cross-review is layer three — catches what I miss. Stack all three, and violations get pushed into an acceptable range.

There was a side discovery worth noting. In chapter six, the writer composed a reflective passage where the character said, “I thought back to that night in chapter one.” The character directly referenced a file label. A character can’t possibly know they live inside chapter one of a novel. This is a different kind of blind spot: while writing a retrospective passage, the tool leaked a marker from its own working environment into the narrative.

These observations, accumulated over chapters, sketch a profile. At the macro level — scene arcs, emotional progression, character motivation — it’s solid. It never falters there. The blind spots cluster entirely in micro-level execution: self-awareness of linguistic tics, self-auditing rule compliance, and the boundary between workspace context and narrative world.

Macro is good. Micro needs supervision. That’s my quality-control model now.