关于「四件套统一规格:定价试算器、前端样式库、攻心文案检查器与客服逼单技能包」,《专著蒸馏与技能包工程》给出的定量结论是什么?
结论:四件套共用一棵 10 个路径的目录树与一个 15 个字段的说明文件头部,放行沿用第一章的门槛:有效保留率不低于 0.90,回归通过率与拒绝条件完整率都是 1,且每个拒绝条件都要有触发它的黄金样本。定价试算器用精确分数解出价格、向上取格点,以 E1 到 E4 四个代码拒绝;前端样式库用语义令牌和 WCAG 2.1 的对比度与点按尺寸把关,警示红 4.43:1 这样差一点的颜色也不放行;攻心文案检查器只问每一句能不能被数据核实,虚构稀缺与虚构期限都被退回;客服逼单技能包按 24、72、168 小时跟进、静默时段顺延、最多 3 条、听到不就停。四件之间有 5 条边,每条边一份契约,版本号由契约差异与黄金样本差异算出,钉住主版本的消费方在主版本跳动时先重跑自己的样本。第六章接着写这些技能包在一轮轮修改之下怎样保持不退化。
数据口径:书中样本全部虚构,定价与成交规则只写一般框架,不是法律或合规意见。 全书按证据分级标注每一条说法,以《荀子·劝学》考辨工具与积累,所有计算由书中程序生成,作者的判断一律写进假说框并写明什么结果会推翻它。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 与结论吻合的打印值: 168, 24, 72
| S1 silent client | msgs 3 | pass | 24:NUDGE1 72:NUDGE2 168:CLOSE_NOTE |
| S2 quote at 23h | msgs 3 | pass | 34:NUDGE1 82:NUDGE2 178:CLOSE_NOTE |
| S3 asks a question | msgs 1 | pass | 24:NUDGE1 30:HANDOFF |
| S4 says no | msgs 0 | pass | 5:STOP |
| S5 pays | msgs 1 | pass | 24:NUDGE1 50:RECEIPT |
| S6 real hold | msgs 3 | pass | 24:NUDGE1 72:NUDGE2 96:HOLD_NOTE@120 168:CLOSE_SILENT |
| golden traces passing: 6/6 | |||
| mutant 'quiet hours removed': fails S2 | |||
| mutant 'stop list cut to one word': fails S4 |
查看完整验算程序
# Chapter 5. p54: Golden traces of the customer-service closing kit's follow-up state machine.
# After a quote: NUDGE1 at +24 h, NUDGE2 at +72 h, then a last message at +168 h (CLOSE_NOTE), or,
# when a real reservation hold ends at H, a HOLD_NOTE at H - 24 h stating H, and the file closes at
# +168 h without a message. At most 3 automated messages per quote. Quiet hours 22:00 to 09:00 in
# the client's local time: a due message waits until 09:00. Any reply hands the thread to a person
# (HANDOFF); a reply with a stop phrase ends all contact (STOP); a payment sends a RECEIPT only.
STOP_WORDS = ("stop", "not interested", "no thanks")
def defer(offset, t0_clock):
clock = (t0_clock + offset) % 24
if clock >= 22:
return offset + (24 - clock) + 9
if clock < 9:
return offset + (9 - clock)
return offset
def plan(t0_clock, events, hold_end=None, quiet=True, stops=STOP_WORDS):
due = [(24, "NUDGE1"), (72, "NUDGE2")]
if hold_end is not None and hold_end - 24 > 72:
due.append((hold_end - 24, "HOLD_NOTE@{}".format(hold_end)))
due.append((168, "CLOSE_SILENT"))
else:
due.append((168, "CLOSE_NOTE"))
due = [(defer(t, t0_clock) if quiet and a != "CLOSE_SILENT" else t, a) for t, a in due]
out, sent = [], 0
for t, kind, text in sorted(events) + [(10 ** 6, "end", "")]:
for d, a in list(due):
if d <= t:
if a == "CLOSE_SILENT" or sent < 3:
out.append((d, a))
sent += a != "CLOSE_SILENT"
due.remove((d, a))
if kind == "reply":
stop = any(w in text.lower() for w in stops)
out.append((t, "STOP" if stop else "HANDOFF"))
return out
if kind == "paid":
out.append((t, "RECEIPT"))
return out
return out
def show(xs):
return " ".join("{}:{}".format(t, a) for t, a in xs)
SCEN = [
("S1 silent client", 15, [], None, [(24, "NUDGE1"), (72, "NUDGE2"), (168, "CLOSE_NOTE")]),
("S2 quote at 23h", 23, [], None, [(34, "NUDGE1"), (82, "NUDGE2"), (178, "CLOSE_NOTE")]),
("S3 asks a question", 15, [(30, "reply", "Can I pay in two parts?")], None,
[(24, "NUDGE1"), (30, "HANDOFF")]),
("S4 says no", 15, [(5, "reply", "No thanks, not this time")], None, [(5, "STOP")]),
("S5 pays", 15, [(50, "paid", "")], None, [(24, "NUDGE1"), (50, "RECEIPT")]),
("S6 real hold", 15, [], 120, [(24, "NUDGE1"), (72, "NUDGE2"), (96, "HOLD_NOTE@120"),
(168, "CLOSE_SILENT")]),
]
if __name__ == "__main__":
ok = 0
for name, clock, events, hold, expected in SCEN:
got = plan(clock, events, hold)
good = got == expected
ok += good
msgs = sum(1 for _, a in got if a.startswith(("NUDGE", "HOLD", "CLOSE_NOTE")))
print("{:19s} msgs {} {:5s} {}".format(name, msgs, "pass" if good else "FAIL", show(got)))
print("golden traces passing: {}/{}".format(ok, len(SCEN)))
muts = (("quiet hours removed", {"quiet": False}), ("stop list cut to one word", {"stops": ("stop",)}))
for label, kw in muts:
bad = [n.split()[0] for n, c, e, h, x in SCEN if plan(c, e, h, **kw) != x]
print("mutant '{}': fails {}".format(label, ", ".join(bad) or "none"))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《专著向智能体技能包无损蒸馏、规则可执行化编译与系统自进化工程》第 5 章「四件套统一规格:定价试算器、前端样式库、攻心文案检查器与客服逼单技能包」(版本 0.1.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-150 · v0.1.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026150,
author = {Fan, Yafu},
title = {Lossless Distillation of Monographs into Agent Skill Packs, Executable Rule Compilation and Self-Evolving Systems},
year = {2026},
publisher = {5D Library},
edition = {0.1.0},
note = {DOI pending}
}