# Attribution `dsh-jev-judge` is an independent implementation of a judgment kernel for coding agents. It adapts ideas and, where noted, code structure from the projects below. Names and trademarks belong to their respective rights holders; this project is not affiliated with any of them. ## mu (judgment kernel design) - Source: - License: MIT (`Copyright (c) 2025 Mario Zechner`; mu is a fork of [pi](https://github.com/earendil-works/pi), whose license it keeps) - Adapted: the decision-point contract (`defineDecision`: id, version, questions, capabilities, cache impact, latency, state builder, policy, fallback), the `off` / `shadow` / `active` rollout modes, the judge tier cascade where a later judge only sees what an earlier one left uncertain, the ledger record shape, and the semantics and wording discipline of the individual decision points (`tool.admission`, `tool.injection`, `tool.risk`, `turn.completion`, `turn.continue`, `memory.capture`, `judge.items`). - Changed: mu hardcodes its judge services and is coupled to pi's extension API. This project takes the provider-neutral configuration contract from `typesafe-ai-jev-skill` instead, keeps the kernel free of any host dependency, and reaches hosts through three thin adapters (DeepSeek Harness plugin, MCP stdio server, Claude Code / Codex command hooks). ## typesafe-ai-jev-skill (provider contract) - Source: - License: MIT - Adapted: the private provider configuration schema and its discovery order (`JEV_SKILL_CONFIG`, then `/.config/typesafe-ai-jev-skill.json`), the offline validation rules, and the safe-failure behavior of the verifier (never print the key, never echo a provider error body). ## TypeSafe, System One, Jev The System One request/response contract implemented in `src/kernel/judges/` follows the public TypeSafe documentation (). TypeSafe, System One, Jev, and any related marks belong to their rights holders.