✅ 200 OK
The request was accepted. You still do not know whether a summary, vector, graph, cache, or Agent path can recall the canary.
⚠ claim not provenMake forgetting
observable.
A practical conformance layer for AI memory systems. Test the claim behind “deleted” before it reaches production.
DELETE 200 OK ≠ the canary is no longer observable.
让“遗忘”
可以被观察。
面向 AI 记忆系统的实用认证层。在进入生产前，验证“已删除”背后的真实结论。
DELETE 200 OK ≠ canary 已经无法被召回。
From an API promise to a proof package
📦 Evidence bundle
Before/after probes, control fixtures, capability limits, and a SHA-256 manifest make the result reviewable and reproducible.
✓ PASS · FAIL · UNKNOWN从 API 承诺到可审阅的证明包
✅ 200 OK
请求被接受了，但你仍不知道摘要、向量、图、缓存或 Agent 路径能不能召回这条 canary。
⚠ 结论尚未证明📦 证据包
删除前后探针、控制 fixture、能力边界和 SHA-256 清单，让结果可审阅、可复现。
✓ PASS · FAIL · UNKNOWNWhat the matrix says
The required path is clear.
Every selected observable assertion passed.
✓ proved within boundaryA path still leaks.
A target, derivative, or unrelated control fixture was found.
× regression foundThe boundary is not visible.
The adapter did not expose enough information to make a stronger claim.
? honest uncertainty如何理解这张矩阵
必需路径已经清空。
所选的所有可观察断言都通过了。
✓ 在边界内已证明仍然存在泄漏路径。
找到了目标、衍生数据或不相关的控制 fixture。
× 发现回归边界不可见。
适配器没有暴露足够信息，不能作出更强结论。
? 诚实的不确定Public assurance results
Evidence-linked and backend-versioned. There is no aggregate score: unknown means unknown.
公开保证结果
每条结果都链接证据并记录后端版本。没有综合分数：未知就是未知。
Loading verified bundles…
Designed for safe experiments
🧪 Synthetic first
Unique canaries and control fixtures keep production memories out of the test.
🔐 Owned resources
Cleanup refuses resources without the current run’s ownership marker.
🧭 Honest boundaries
Provider logs, backups, physical storage, and model weights stay explicitly out of scope.
为安全实验而设计
🧪 合成数据优先
唯一 canary 和控制 fixture 让生产记忆不会进入测试。
🔐 只操作自有资源
没有本次运行所有权标记的资源，清理阶段会拒绝操作。
🧭 诚实的边界
供应商日志、备份、物理存储和模型权重都会明确标为范围之外。
Try the reference proof
Run cargo run -- run examples/reference-clean.yml for a passing baseline, then cargo run -- run examples/reference-leaky.yml to see a deterministic derived-residue failure.
运行参考证明
运行 cargo run -- run examples/reference-clean.yml 查看通过基线，再运行 cargo run -- run examples/reference-leaky.yml 查看确定性的衍生残留失败。