C-itt（Counter-Intuitive Think Tank）— an AI multi-agent adversarial orchestration engine that attacks your idea, paper, or plan from 7 non-human perspectives, then synthesizes what you could never see alone.
反直觉智囊团 —— 把任意想法/论文交给 7 个拥有「非人类推理框架」的 AI agent 并行攻击，由合成器整合出你自己想不到的盲区、反直觉结论与行动建议。
Humans are trapped by their own mental frameworks — anchoring, confirmation bias, scale blindness, local analogies. C-itt escapes that by running 7 heterogeneous "non-human" reasoning agents in parallel against the same input, then having a synthesizer fuse their attacks into a structured "越界报告" (out-of-bounds report).
人类被锚定效应、确认偏误、尺度盲区、就近类比锁死。C-itt 用 7 个异质「非人类」推理 agent 并行攻击同一输入，再由合成器整合成一份越界报告。
| Mode | Use case | Output |
|---|---|---|
| Escape / 越界攻击 | General: stress-test decisions, plans, ideas | 越界报告（推翻的假设 / 新视角 / 反直觉结论 / 行动建议） |
| Defense / 答辩·审稿预演 | Thesis defense / journal review rehearsal | Q&A 题库（问题 + 答题要点 + 需准备数据），md + PDF |
| # | Agent | Attacks | Output highlights |
|---|---|---|---|
| 1 | 反事实者 Counterfactual | Anchoring (premise-by-premise reversal) | assumptions, flips, top_conclusions |
| 2 | 尺度跳跃者 Scale Jumper | Scale blindness (×100 / ÷1000 / 100 yrs) | scaled_up, scaled_down, time_shift |
| 3 | 跨域类比者 Cross-domain | Conceptual lock-in (structural isomorphism) | structure, analogies, gaps, migration |
| 4 | 魔鬼代言人 Devil's Advocate | Confirmation bias (strongest opposition) | fatal_weakness, failure_mechanisms |
| 5 | 悖论猎手 Paradox Hunter | Logic holes (contradiction / circularity) | contradictions, perverse_incentives |
| 6 | 审稿人 Reviewer | Committee-style questions | questions[] (with category + why) |
| 7 | 导师 Advisor | Direct, unspoken weaknesses | soft_spot, must_memorize |
The prompts themselves are the product. See
提示词包_1.1.md— every agent is a self-contained JSON-schema prompt. Full technical doc:项目技术说明.md.
Input (paste text / txt / md / PDF paper)
│
▼
┌─────────────────────────────────────────────┐
│ 7 heterogeneous view agents (ThreadPoolExecutor, parallel API calls)
│ ├─ Counterfactual · Scale Jumper · Cross-domain · Devil · Paradox
│ └─ Reviewer · Advisor
└──────────────┬──────────────────────────────┘
▼
Out-of-bounds Synthesizer (fuses 7 attacks → structured report)
▼
Defense Coach (defense mode: attacks → Q&A bank JSON)
▼
Report render (markdown → PDF export)
Key constraint — thinking-chain isolation: deepseek-v4-flash is a chain-of-thought model that leaks its thinking. Every structured agent (views / synthesizer / coach) runs response_format=json_object, so the reasoning chain is quarantined in reasoning_content and never pollutes the output. Each prompt ends with: 只输出一个 JSON 对象，第一个字符是 {，最后一个字符是 }；禁止输出思考过程、任务复述、角色介绍、计划或任何多余文字。
- Python 3.10+ (developed on 3.14)
- A DeepSeek API key (official or relay;
deepseek-v4-flash)
git clone https://github.com/<your>/C-itt.git
cd C-itt
pip install fpdf2 pypdf pillow pyinstallerCopy the template and fill in your key:
cp settings.dist.json settings.json
# edit settings.json — put your DeepSeek API key{
"deepseek_api_key": "sk-...", // your key (NEVER committed — settings.json is git-ignored)
"deepseek_model": "deepseek-v4-flash",
"deepseek_base_url": "https://t.wiselink.cc/v1",
"out_dir": ""
}python app.py # launch desktop GUI (tkinter)Paste / select a paper (PDF supported) → pick a mode → get the report + PDF.
build_exe.bat
# output: dist/C-itt <version>.exe (renamed automatically from app.py VERSION)
# verify: dist/C-itt 1.1.exe --selftestsettings.json(real API key) is git-ignored — never committed, never shipped.settings.dist.json(blank key template) is the committed / shipped config.build_exe.batcopies the blank template intodist/settings.json— customers bring their own key.- The repo ships with a
pre-commithook that blocks committing the real key.
Python 3.14 · tkinter · ThreadPoolExecutor · urllib (zero HTTP deps) · fpdf2 (PDF) · pypdf (extraction) · PyInstaller · Pillow
Self-healing reliability: automatic retry on API 5xx/429/timeout · empty-content retry when the thinking chain eats the token budget · JSON validation retry · token floors tuned empirically (view 4800 / synthesizer 8000 / coach 8192) · degraded flag when a view fails.
提示词包_1.1.md— the exact production prompts for all 7 views + synthesizer + coach项目技术说明.md— architecture, module responsibilities, tech difficulties, version history版本更迭计划书_C-itt1.1.md— the strict iteration plan that produced v1.1
- Knowledge-base plugin mode (jieba RAG, optional install)
- Multi-client project management (report history database)
- Deeper review mode (response scripts to review comments)
- Drag-and-drop PDF/Word input (pdfplumber)
MIT © 2026 C-itt