workflows

GameDesignOS Workflows

These workflow guides describe how existing skills write durable assets into a GameDesignOS project workspace.

中文摘要：工作流文档不是新的 skill，而是规定已有 skill 如何读取上游资产、用 VOI 判断是否需要更多信息、把结果写回 workspace，并在关键节点进入 Human Gate。

Cross-Cutting Route

• Decision to Information: define the decision, current default action, boundary, information actions, EVSI probe, and stop rule before research expands.

Project Governance Checkpoints

$paranoia-ai-system-evolver is a cross-cutting governance layer for every workflow, not a replacement for the domain skill that owns the work.

Each workflow should expose these checkpoints in workflow-run.governance:

1. intake intent: convert vague instructions into an Intent Work Order when the task is open-ended;
2. VOI/UL/RJR boundary: name the Decision Object, current_default_action, highest-value uncertainty, optional ul_state_ref, authority level, residual judgment, and Human Gate trigger;
3. midstream drift: check branch explosion, scope creep, low-VOI research, missing evidence, and over-structured output;
4. delivery gate: verify first-glance acceptance, non-sacrifice constraints, rollback, and unsupported claims;
5. retrospective learning: persist only candidate rules until eval evidence and Human Gate justify promotion.

Default enforcement is shadow. Move to warn or enforce only after representative evals show that the checkpoint improves output without hiding useful domain work.

Core Production Routes

• Idea to Validation
• Media to Diagnosis
• Weekly ED Experiment
• Evidence to Proposal

Shared Rules

Every workflow must:

1. declare its input and evidence boundary;
2. identify the decision, owner, deadline, real options, and current_default_action;
3. inspect accepted decisions and upstream assets before asking the user to repeat context;
4. classify the decision boundary as undefined, far, near, or locked;
5. acquire new information only when a plausible signal can change action, priority, allocation, or a stop condition;
6. generate no more than three candidate information actions per VOI round;
7. prefer the smallest sample or experiment with positive net value after acquisition, delay, attention, privacy, and contamination costs;
8. route to the smallest suitable skill;
9. save outputs to the documented workspace area and update asset relationships;
10. distinguish observations, interpretations, assumptions, model learning, consumption, and decisions;
11. preserve local negative evidence instead of washing it into generic prose;
12. end with a next action, Human Gate, or explicit research stop condition;
13. write workflow governance refs for intent, VOI, optional UL state, RJR, Human Gate, rollback, and candidate learning;
14. avoid publishing private project material back to the public repository.

UL（Uncertainty Ladder，不确定性阶梯）只在系统演化、复杂能力验证或失败归因需要控制未知暴露时启用；普通领域 workflow 不强制生成 UL 文件。完整方法见 paranoia-ai-system-evolver/references/uncertainty-ladder-protocol.zh-CN.md，VOI 方法见 value-of-information-playbook.zh-CN.md。