# Shape-Aware Evaluation Design

**trigger_condition：** full apply 已确认单一 source owner，准备写 core 前的 baseline、protected behavior 与 pre-patch eval plan。**must_read：** full apply 必须读完；Tier A behavior-preserving revise 不触发。**fallback_if_unread：** 不得开始 full apply source patch，只能报告 structural-only evidence，且不得声称 baseline、semantic adequacy 或行为改善。**eval_case：** `prepatch-entry-governor` 触发。

先选择目标项目已有 native eval owner；没有时创建 target-local maintainer cases，不把 `spec-write-skill` promotion validator 变成通用平台。fresh reviewer 只接收 raw source、真实 request 和 artifact，不能收到 intended fix。

| Shape | Baseline and protected behavior | Minimum eval family |
| --- | --- | --- |
| entry governor | with-skill vs bare-menu、route collision、near-neighbor | route/discipline，必要时重复 run 报告波动 |
| artifact producer | artifact contract、错误输出和 consumer | machine assertion + semantic artifact rubric |
| deterministic setup/validation | facts、reason code、failure boundary | unit/integration assertion，不用 prose 替代 deterministic facts |
| prose/agentic workflow | authority、输出 contract、good/bad/why | positive、near-neighbor、failure fresh-source cases |
| long-horizon loop | checkpoint、stop/resume/recovery | authority + checkpoint + recovery path |
| hybrid | 每个承重面各自 evidence | 组合必要 case，不强迫统一总分 |

新增 protected behavior 必须有 `protected_behavior → source carrier → contract assertion → semantic eval case` 映射。不能把 fixture pass、模型自检或 source bytes 当作 runtime quality。反馈/transcript/issue 先是 advisory：绑定 source/host/model，脱敏、最小复现、确认 expected behavior，并取得 eval-source mutation authorization后才可转 regression；否则只记录 observation。

## Model-Family Adaptation

用户要求升级或适配模型系列时，分开报告 `model-configured`、`skill-source-adapted`、`runtime-projected` 与 `optional-capability-validated`；这些是 claim 标签，不是新 schema。修改 prompt 前先用相同 source、cases、tools、endpoint contract 与有效 reasoning 运行 model-only baseline，再把候选 source 作为单变量 treatment。至少记录 source hash、host/runtime、请求模型与实际返回模型、effective reasoning、treatment、quality gates、token/latency/calls/cost、limitations 与 keep/revert/iterate/defer 决策。

如果真实模型调用需要未获授权的外部成本，或缺少 fresh-source 隔离 runner，则对应 treatment 保持 `not_run`。可以继续做有独立依据的 source hygiene，但只能称为 prompt-hygiene candidate；不得从 source 文案、runtime projection 或 structural fixture 推断模型已配置、语义已适配或效率已提升。`model-family-source-adaptation` 覆盖这一边界。

closeout 分开报告 deterministic checks、eval adequacy、五轴 readiness、default-source bytes、field outcome（默认 `not-run`）、not-run reasons、generated runtime status 和 residual risks。`manual_observation` 不能支持 default promotion。