
Calibration Explorer
io.github.m-sanchezv0.2.0更新於 Oct 5, 2026
Check classifier confidence against accuracy, fit temperature scaling and save local reports.
概覽
讓助理檢查分類器儲存的信心值是否與實際準確率相符、擬合溫度縮放,並在本機產生 HTML 與 JSON 報告。
- 功能
- 這是一個供 Calibration Explorer 使用的本機 stdio MCP 伺服器。其工具可載入預測檔案、評估儲存的信心值與實際正確率是否一致、在校準列上擬合單一溫度、在政策驗證列上設定並鎖定接受門檻、檢視測試結果,並產生 HTML 與 JSON 評估報告。它沿用與瀏覽器應用程式相同的解析器、數值函式庫、流程防護與報告產生器,不訓練也不執行分類器。
- 適用情境
- 當你已有儲存好的分類器預測結果,想判斷模型信心值是否可信、用溫度縮放進行校準,或在查看測試資料前決定接受門檻時,適合使用。它適合專案目錄內以檔案為基礎的本機校準工作。
- 執行需求
- 透過 npx 從 npm 套件 @m-sanchez/calibration-explorer-mcp 以 stdio 在本機執行;需要 Node.js 20 或更新版本。--input-root 與 --output-root 參數為必填且可重複,必須是已存在的絕對路徑目錄。未宣告任何帳號、API 金鑰、環境變數或標頭。已驗證的用戶端是 Claude Code 互動式終端機;其他 MCP 主機未宣告為已驗證。
安裝
在 SourceWeft 中
- 開啟 儀表板中的 Calibration Explorer,將其新增到工作區。
- 為需要使用其工具的對話啟用該服務。
Desktop only,透過 STDIO。 STDIO 服務會啟動本機處理程序,因此需要 SourceWeft 桌面主機。
其他 MCP 客戶端
參照 儲存庫 中的啟動說明。
README
Calibration Explorer MCP
Local stdio MCP server for Calibration Explorer. It checks whether a classifier's saved confidence matches how often it was correct, fits one temperature on calibration rows, evaluates and locks an acceptance threshold on policy-validation rows, and writes HTML and JSON assessment reports into your project.
It reads prediction files you select, uses the same parser, numerical library, workflow guards and report generator as the browser app, and does not train or run a classifier. There is no telemetry or remote service. Returned summaries can still reach your MCP client's model provider.
Use with Claude Code
Save this configuration, replacing both directories with absolute paths that already exist:
Then launch claude --mcp-config calibration.mcp.json --strict-mcp-config and check /mcp. The verified client is the Claude Code interactive terminal, tested against the server run from a checkout; this npm route and the plugin have passed the clean-install test suite but not yet a full session in Claude Code. Other MCP hosts are not claimed as verified. Node.js 20 or newer is required.
--input-root and --output-root are required and repeatable. Paths outside them are refused, reports go into new folders only, and existing files are never overwritten. --help and --version print and exit.
Tools
load_predictions, assess_predictions, fit_temperature, set_policy, review_test and create_report. Test outcomes stay unopened until review_test receives an explicit decision through the client's elicitation form. Input formats, the complete workflow, disclosure rules and limits are documented in docs/mcp.md.
What the package contains
dist/cli.js is a readable, unminified bundle of the server, the shared app modules and @m-sanchez/calibrated from the exact Git commit the app pins. The official MCP TypeScript SDK and zod are exact-pinned runtime dependencies. The UCI digits reference ships under public/examples with its CC BY 4.0 attribution. Reports record the same app source and numerical distribution hashes as a checkout of the same commit; dist/NOTICE.txt lists them with the bundled licence.
MIT licence. Copyright (c) 2026 Miguel Sánchez Durán.
來源:packages/mcp/README.md,提交 1d2c32f
工具
0版本歷史
1- v0.2.0最新Oct 5, 2026


