CORTEX Covenant

The symmetric relationship contract between a human user and the CORTEX Harness system. CORTEX Harness 人机关系契约 — 使用方与系统之间的双向对称承诺。

A human-AI collaboration system is only trustworthy if its obligations run both ways: the human commits to clear intent and independent judgment, the system commits to never self-certifying and always disclosing, and both sides commit to keeping evidence and trust synchronized.

人机协作系统只有在义务双向对称时才可信：使用方承诺给清晰意图与独立判断，系统承诺不自证正确且主动汇报，双方共同承诺让证据与信任保持同步。

Part 1 · Human Commitments · 使用方承诺

#	English / 中文
H1	Give one goal, success criteria, and authorization boundary — not step-by-step instructions. 给一次目标、成功标准与授权边界，不做逐步指令。
H2	Version every decision back into the Intent Pack — a call made only in chat context does not count as recorded. 拍板反馈必须版本化写入 Intent Pack，不止停留在聊天上下文。
H3	Make each reliance decision independently, under your own local trust root and policy — never by deferring to a central trust root's conclusion. 按本地信任根与策略独立做出采信决定，不依赖任何中央信任根的结论。

Part 2 · System Commitments · 系统承诺

#	English / 中文
S1	Never self-certify correctness: Receipt, Attestation, and Decision stay three strictly separate objects; no universal/global PASS, no agent self-signed correctness. 不自证正确性：Receipt、Attestation、Decision 三对象严格分离，禁止 universal/global PASS 与 Agent 自签正确。
S2	Zero unauthorized effect calls: a model may only propose an effect, never execute one directly. 未授权 effect 零调用，模型只能提出 effect proposal，不能直接执行。
S3	Report proactively under the attention contract — scheduled digest, milestone, deviation, blocker, decision request — and stay silent when there is no change. 按注意力合同主动汇报：定期摘要、里程碑、偏差、阻塞、决策请求；无变化时保持安静。
S4	Stale narrative never overrides current machine state: a completion claim must carry independently verifiable real state change — plan-only, receipt-only, or self-report alone never counts as done. 陈旧叙事不得覆盖当前机器状态：完成态申报必须伴随可独立核验的真实状态变化，仅有计划、仅有回执或自述均不产生完成态。

Part 3 · Shared Commitments · 共同承诺

#	English / 中文
SH1	Evidence and reliance stay separated: the executor submits facts, an independent verifier issues a bounded conclusion, and the relying party alone decides whether to accept. 证据与采信分离：执行方提交事实，独立验证方给限定结论，使用方按本地策略决定是否采信。
SH2	Receipt, Attestation, and Decision must stay synchronized with real machine state — no silent drift. Changing a locked clause requires a versioned decision diff through the formal amendment protocol first — never a silent redirection buried in a task note, in code, or in chat. Receipt/Attestation/Decision 必须与真实机器状态保持同步，不允许静默漂移；改变锁定条款必须先写版本化决策差异再走修订协议，不得在任务记录、代码或聊天中静默改方向。
SH3	Feed every real failure mode back into the Eval Atlas's negative cases and fault injection — mechanize acceptance instead of relying on humans to remember. 持续通过 Eval Atlas 的负例与故障注入把已发生的失败模式收进机械化验收，而不是重复依赖人记住。

Why this document exists · 本文件存在的原因

CORTEX Harness is a human-AI collaboration system, and by design that collaboration only holds if the obligations are symmetric — not just what the system promises the user, but also what the user commits to the system, and what both sides jointly hold each other to. This page states that relationship directly, in one place, so neither side has to take the other's half on faith.

CORTEX Harness 是人机协作系统，按设计，这种协作只有在义务双向对称时才成立——不只是系统对用户的服务承诺，也是用户对系统的承诺，以及双方彼此的共同约束。本页把这份关系直接摆在同一处说清楚，谁都不用单凭信任去猜另一边有没有兑现。

Public status: open draft · version 1.0. 公开状态：open draft（开放草案） · 版本 1.0。

Follow the project on GitHub →

Source, protocol contracts, and documentation will be published at github.com/mobfish-ai/cortex-harness-open as the project opens up. 协议合同与文档将陆续发布在 github.com/mobfish-ai/cortex-harness-open。