ZTL Judge

io.github.inventor1975v1.0.0更新於 Sep 29, 2026

Zero-trust logic judge: your AI writes a claim as a ZFL table, the ZTL core judges it.

已驗證Streamable HTTP可網頁執行AI & MLProductivity & Workflow

概覽

AI 產生的概覽

讓助理把自然語言主張轉換成形式化的 ZFL 表格，再由確定性的 ZTL 核心裁決，給出結論、保證等級與通行證。

功能
助理協助把主張形式化為 ZFL，也就是由若干列加一條主張組成的表格，每列包含狀態（T 已核驗、F 已駁斥、Z 未核驗）、依據，以及用文字說明的含義。接著由確定性的 ZTL 核心裁決，回傳二值結論、保證等級、未核驗輸入的通行證，以及完備性表。對於自指系統，它會報告有根據的部分、隔離集合，以及每個元件的通行證，例如 PARADOX、UNDERDETERMINED、INPUT 或 DOWNSTREAM。確定性的回讀會複述核心實際讀到的內容，可選的 AI 解釋只複述結論而不重新裁決。
適用情境
當你希望助理協助精確表述某個主張或悖論，再由確定性的非 LLM 元件裁決，而不是依賴模型自身的推理時，適合使用。它適用於邏輯、形式化與一致性檢查類工作，這類工作要求把翻譯步驟與裁決步驟分開。對於一般問答或無法化約為形式化表格的任務，它較不適用。
執行需求
遠端端點 streamable HTTP；清單宣告不需驗證、環境變數或標頭。README 描述了一個本機 Python 工作室（python3 ztlstudio.py），只用標準函式庫，ZTL 核心已內建。AI 翻譯是可選的：沒有金鑰時以 pro 模式手動填表；有金鑰時可使用 Groq、Anthropic、OpenAI、OpenRouter、DeepSeek、Gemini、xAI 或 NVIDIA 等供應商。
安裝前請注意
可選的 AI 翻譯步驟會把你的主張文字傳送給第三方模型供應商；API 金鑰來自設定欄位、環境變數（如 GROQ_API_KEY、ANTHROPIC_API_KEY）或本機 . _key 檔案。README 稱金鑰保留在本機且被 gitignore，但能存取該機器或檔案的人仍可讀取。AI 解釋被標註為未經核驗，不能當成裁決結果。

安裝

在 SourceWeft 中

  1. 開啟 儀表板中的 ZTL Judge，將其新增到工作區。
  2. 為需要使用其工具的對話啟用該服務。

Web executable，透過 Streamable HTTP。 遠端服務在工作區中設定後即可從網頁執行環境執行。

其他 MCP 客戶端

把它新增到你客戶端的 mcpServers 設定中。

{
  "mcpServers": {
    "ztl-judge": {
      "type": "http",
      "url": "https://api.vitalyreznik.com/mcp"
    }
  }
}

README

ZTLStudio

The AI translates; the measured core judges — truth is never granted on credit, not even to the translator.

A local studio for judging claims and paradoxes. You state one in natural language (any language); an LLM only translates it into ZFL, the formal table language — it never judges. A deterministic, measured ZTL core does the judging: verdicts with warranties, quarantine passports for self-referential systems, and a deterministic back-reading that verbalizes exactly what the core read from your table.

The pipeline embodies the logic it serves: the LLM's output is an unverified input (the mark Z), and the core is the customs house — truth is never granted on credit, not even to the translator.

human ──meta-chat──► the AI fills a ZFL table (rows + a claim), you sign off                         │                         ▼ validator ──► the deterministic core judges                         ▼ back-reading (no AI — the second auditor)                         ▼                   verdict · warranty · passport · stipulations

Run

python3 ztlstudio.py        # → http://localhost:8190

Python stdlib only; the ZTL core is vendored in ztlcore/, so a clone is self-contained (no submodules, no dependencies). The AI is optional: with no key the studio runs in pro mode — fill the ZFL table by hand. To enable AI translation, open ⚙ Model, pick a provider + model + key, or set the env var, or drop a key into a local .<provider>_key file (all gitignored — no keys ship).

What you hand it: one table, no genre to declare

ZFL v2 is a single table of rows plus a claim. Each row states a fact, its status (T verified / F refuted / Z unverified — the zero-trust default), its ground, and — importantly — what it means in words (the polarity auditor: it lets the back-reading catch an encoding that says the opposite of what you intended). You never declare whether this is a "statement" or a "paradox": the genre is computed, and whichever instruments apply fire — a verdict + warranty for a claim, a passport for a self-referential system.

The studio ships 41 worked examples — open one to see the exact shape of the table, then edit it. The back-reading verbalizes what the core actually read, so your translation is audited by a component that cannot hallucinate.

The workflow

  1. Meta-chat — describe the claim in your language; the AI fills the table's rows and asks only when formalization is genuinely blocked. It knows its boundary: arithmetic, quantities and numeric wordplay get an honest "does not formalize into propositional ZTL", never an invented encoding.
  2. The table — a grid of rows, the grounds bar, and the claim line, all hand-editable (pros skip the chat entirely). Run validates and judges; validator issues are machine-readable and can be fed back to the AI to repair.
  3. The report — the core's verdict, its warranty grade (hereditary / sound / until-verification), the passport of unverified inputs, and the completion table — followed by the deterministic back-reading and an optional AI explanation that retells the verdict and is forbidden to re-judge (labeled unverified by definition: the pipeline applies its own logic to itself).

What the core reports

  • Claims — the verdict (T/F — verdicts are always two-valued; Z is a mark on an input, never a verdict), the warranty grade, the passport of unverified inputs, and the completion table showing how the verdict behaves under every reading of the unverified rows.
  • Self-referential systems — the grounded part (identical in every fixed point), the quarantine set, and a passport per component: PARADOX (no classical solution — permanent refusal, with the oscillation period), UNDERDETERMINED (refusal until stipulation), INPUT (until verification), DOWNSTREAM (inherited).

Providers

Keys stay on this machine, read in order: the Settings field, the env var (GROQ_API_KEY, ANTHROPIC_API_KEY, …), then a local .<provider>_key file. Supported: Groq, Anthropic (Claude), OpenAI, OpenRouter, DeepSeek, Gemini, xAI, NVIDIA. A stronger model formalizes cleaner; the core judges the same regardless of who translated.

Related

  • ZTL — the logic itself: the kernel, the papers, and the ZFL language.
  • introspect — the same zero-trust core applied to code: a taint analyzer for seven languages.

AI disclosure

Built by Claude (Anthropic) as architect and implementer, with Vitaly Reznik as human curator and decision-maker, under a strict honesty discipline: mark boundaries honestly, measure — don't guess, and never claim more than was verified.

License

Dual-licensed under MIT and Apache-2.0 (see LICENSE-MIT, LICENSE-APACHE).

來源：README.md，提交 c7e6f19

工具

0
工具後設資料尚未被收錄。

版本歷史

1
  1. v1.0.0最新Sep 29, 2026