01What is TRIZ?TRIZ 是什么

TRIZ — the Theory of Inventive Problem Solving — was distilled by Genrich Altshuller and his colleagues starting in 1946, from the analysis of hundreds of thousands of patents. Its core insight: innovation is not random genius. The same technical problems recur across industries, and their solutions follow reusable patterns. A problem solved in aerospace in 1975 may be the same contradiction a battery engineer faces today.

TRIZ（俄文"发明问题解决理论"缩写）由根里奇·阿奇舒勒及其同事自 1946 年起，从对数十万件专利的分析中提炼而成。它的核心洞察是：创新不是随机的天才灵光——同样的技术问题在不同行业反复出现，而它们的解法遵循可复用的模式。航空领域 1975 年解决的问题，可能就是今天电池工程师面对的同一个矛盾。

The toolbox is unusually structured for a creativity method: 40 numbered inventive principles and the contradiction matrix; separation principles for physical contradictions; ARIZ, a full problem-solving algorithm; and the laws of technical-system evolution. Samsung, Siemens, GE and Boeing have trained thousands of engineers on it. The catch: today it lives almost entirely inside expensive courses and certified consultants — a few thousand experts worldwide, against millions of engineers who could use it.

它的工具箱在"创造力方法"里罕见地结构化：40 个有编号的发明原理与矛盾矩阵、分离原理、完整的 ARIZ 算法、技术系统进化法则。三星、西门子、GE、波音都曾用它内训数以千计的工程师。问题是：这套方法论今天几乎只存在于昂贵的培训课程和持证咨询师身上——全球活跃咨询师千人量级，而需要它的工程师是百万量级。

02Why fine-tune for TRIZ?为什么值得为 TRIZ 做微调

Prior LLM+TRIZ work was all prompt engineering. Fine-tuning buys three things prompts can't: terminology discipline (correct principle numbering and analysis steps, baked into weights rather than begged for in prompts), behavioral style (direct, executable answers — our fine-tunes compress base responses to ~1/10 length at parity coverage), and internalized capability (the difference between consulting a book and knowing the method).

此前的 LLM+TRIZ 工作全部是提示词工程。微调能买到三样提示词买不到的东西：术语纪律（原理编号引用、分析步骤完整性，固化进权重而不是在 prompt 里恳求）、行为风格（直接可执行的回答——我们的微调把基座回答压缩到约 1/10 长度且覆盖持平）、能力内嵌（"查书答题"与"学会方法"的区别）。

It turns methodology application from a per-day-billed service into infrastructure with zero marginal cost.

微调模型把方法论的应用能力，从按人天计费的服务变成边际成本为零的基础设施。

For enterprises, the innovation bottleneck is rarely ideas — it is problem formulation. A resident model answers "which principles apply to my contradiction?" at 2am before the design review; it scales the onboarding tier (our strongest task types are exactly principle recommendation and concept explanation); and it drafts contradiction-analysis reports so expert time shifts from production to judgment. An assistant, not a replacement — but it raises the floor for everyone.

对企业而言，创新的瓶颈往往不是点子，而是问题表述。常驻模型能在设计评审前夜凌晨两点回答"我的矛盾该用哪几条原理"；它把入门普及层规模化（我们最强的子集恰是原理推荐与概念讲解）；它起草矛盾分析与 ARIZ 报告，让专家时间从"生产"转向"判断"。这是助手而非替代——但它抬高了所有人的下限。

03Results — all of them, including the tie结果——全部结果，包括打平的那个

v5 gold protocol, 300 items, per-item paired statistics. Judge-track numbers here are same-family (Moonshot) readings — the external-judge final review follows below:

v5 金标口径，300 题，逐题配对统计。本表 judge 轨数字为同族（Moonshot）评委读数——异源评委终审见下方：

On the older v4 protocol (100 items), Meerkat-TRIZ-v1 is statistically tied with our strongest internal baseline (judge −0.22 [−0.47, +0.01], n.s.). We ran no external-model comparisons and claim none.

在较早的 v4 口径（100 题）上，Meerkat-TRIZ-v1 与内部最强基线统计打平（judge −0.22 [−0.47, +0.01]，不显著）。我们未做外部模型对比，也不做任何此类宣称。

External-judge final review (post-release)

异源评委终审（发布后补做）

After release we re-ran the judge protocol verbatim — same rubric, same prompts, T=0 — with three external judges from three vendors. 299 paired items (1 blocked by provider content filter):

发布后，我们将评委协议逐字复跑——同 rubric、同 prompt、T=0——换成三家外部厂商的评委。299 题配对（1 题被服务商内容过滤拦截）：

Direction confirmed under two of three external judges — at roughly one quarter of the same-family magnitude. The defensible headline is +0.09 ~ +0.10; the +0.39 must not be extrapolated across judge families. The external judges agree with each other (Spearman 0.63–0.75) far more than with the same-family judge (0.27–0.31) — a scoring-system difference, not noise. The strongest consensus: the targeted concept_explanation repair is significant under all three external judges (+0.24 / +0.31 / +0.29). Watch item: gemini scores principle_recommendation significantly lower (−0.31) — single-judge support, logged as an open signal. Full report →

三个外部评委中两个确认正方向——但量级仅为同族读数的约 1/4。可辩护的头条结论是 +0.09 ~ +0.10；+0.39 不得跨评委家族外推。外部评委互相一致性（Spearman 0.63–0.75）远高于与同族评委的一致性（0.27–0.31）——这是评分体系差异，不是噪声。最强共识：定向修复的概念讲解子集在三个外部评委下全部显著（+0.24 / +0.31 / +0.29）。观察项：gemini 对原理推荐子集给出显著负差（−0.31）——单一评委支持，记为开放信号。完整报告 →

04The evaluation caught itself lying评测系统抓到了自己说谎

Three measurement failures found and repaired during five versions of iteration — each capable of reversing a conclusion:

五版本迭代中发现并修复的三类测量学故障——每一类都足以反转结论：

05Limitations, stated局限，如实声明

• Single domain (Chinese TRIZ QA), single base, single PEFT config.
• 单领域（中文 TRIZ 问答）、单基座、单 PEFT 配置。
• Judge shares a family with the data generator (declared weak same-origin). A post-release external review (§03) confirmed direction under 2/3 external judges at ~1/4 the magnitude; same-family absolute scores are not portable across judge families.
• 评委与数据生成器同族（已声明"同族弱异源"）。发布后异源终审（§03）在 2/3 外部评委下确认方向，但量级约为同族读数的 1/4；同族绝对分不可跨评委家族外推。
• 100-item subsets are descriptive only; "tied with v2" means no significant difference detected, not proven equivalence.
• 100 题口径子集仅为描述性；"与 v2 打平"意为未检出显著差异，而非已证等价。
• Public benchmark withholds reference answers (copyright); keyword track fully reproducible, judge absolute scores not.
• 公开基准不含参考答案（版权原因）；关键词轨可完整复现，judge 轨绝对分不可。

06Quick start快速上手

from peft import PeftModel
model = PeftModel.from_pretrained(base, "Meerkat-AI/Meerkat-TRIZ-v1")
# ⚠️ keep the empty think block in your prompt template —
#   the model was trained with it retained; stripping measurably hurts output.
# ⚠️ 请保留 prompt 模板中的空 think 块——模型在保留它的格式下训练,
#   剥离会可测量地损害输出质量。