You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
task: 【Hao AI Lab · Towards Verifiable LLM Infra】 将 vstack 的规格冻结与模型保真检查纳入 RSI evidence 工作流 #402
Independent issue review, scoped to main 61303fa8150472dfbda62ec8d77908bc05c73f3b. Acting as the delegated reviewer under the maintainer's standing review authorization; this is a documentation/search assessment, not acceptance of implementation scope, ownership, or a campaign PASS.
The useful integration boundary is a source-derived reference under the existing rsi-evidence owner, with explicit evidence-type separation. Three acceptance gaps should be made testable:
Do not turn the PD example into a runnable or validated bundle by citation. At vstack 793c9edc2ee2c0cb21ff38fa8658130e5f452439, the regression example is explicitly an adaptation example without its project models, traces, mutations, runner or experimental evidence. Its five-step procedure distinguishes drift/source-derived checks from replay of already-frozen traces. The reference should enumerate those missing inputs rather than imply copying the workflow provides implementation validation. A frozen replay can remain green after a live code edit; the offline cases should test that distinction, not just repeat “proof is not implementation.”
Preserve the actual theorem premises and non-vacuity boundary. The pinned PdCore proof README conditions its theorem on Sizes and Design, including the narrow completion assumption and disabled unquarantine. It also documents that the four-location map makes some larger instance configurations vacuous. An interpretation example should reject a claim of arbitrary deployment/transport coverage even if the design proof succeeds. Likewise, whole-trace completion, invariant satisfaction, and required abort/release reachability need separate observations; a missing event is insufficient evidence, not proof that the event is safe. I inspected the source documentation, not a fresh proof run.
Searchability needs a packaged-consumer test, not only a Markdown link. On the reviewed main, rsi-evidence/metadata.json has empirical_validation: not_run and a LoopX-only provenance readset; its generated knowledge record inventories the existing resources. Actual source-tree CLI queries knowledge search PdLedger --json and knowledge search vstack --json both returned []; rsi-evidence remains discoverable. For the proposed addition, require a wheel installed outside the checkout to search the intended protocol terms, open the resulting reference, and resolve the exact upstream SHA, source path, license and scope. Include the reference in the generated record/resource inventory, and assert the two offline interpretation outcomes. Keep source provenance pins distinct from campaign baseline/candidate pins. Preserve the upstream MIT notice for copied vstack material.
Validation: read the existing skill, metadata, generated record, packaging and fresh-install test path; ran the CLI queries above. The default interpreter initially lacked pytest. After checking an existing workspace venv, reran PYTHONPATH=src <existing-venv>/bin/python -m pytest -q tests/test_skill_catalog.py tests/test_knowledge_find.py: all 91 tests passed (one pytest-plugin deprecation warning). These existing catalog/find checks do not establish acceptance of the proposed reference. I did not build/install a wheel, run TLC/TLAPS, capture implementation traces, or run performance/Agent-utility experiments. No code changed.
Conclusion: these are concrete acceptance requirements for a bounded knowledge-reference slice. Current main does not yet supply that reference; delivery, implementation fidelity and Agent utility remain unverified, separately.
Activity
[-]task: 将 vstack 的规格冻结与模型保真检查纳入 RSI evidence 工作流[/-][+]task: 【Hao AI Lab · Towards Verifiable LLM Infra】 将 vstack 的规格冻结与模型保真检查纳入 RSI evidence 工作流[/+]amy-why-3459 commentedon Nov 12, 2026
这个方向很有价值。对缓存所有权、异步传输、资源回收这类改动,现有的吞吐、延迟和输出样本确实说不清“每种允许的时序下谁拥有这份资源”。第一版只交付
rsi-evidence下的一份可检索参考材料,我认为范围合适。有几点建议:793c9edc…)和 veriserve(8345c9c9…)各自的许可证,以及参考材料是蒸馏重写还是摘录原文。参考材料需要带上游来源记录(仓库、SHA、路径),写法可以参照现有 knowledge 源记录(例如 Add source-bound TTS playout benchmark guidance #339 的 source record),这样“追溯上游 SHA”这条验收能直接复用现有检查。结项标准(文档加检索验收,Agent 效果另做对照实验)我同意。
zengchuang-hw commentedon Nov 17, 2026
Independent issue review, scoped to main
61303fa8150472dfbda62ec8d77908bc05c73f3b. Acting as the delegated reviewer under the maintainer's standing review authorization; this is a documentation/search assessment, not acceptance of implementation scope, ownership, or a campaign PASS.The useful integration boundary is a source-derived reference under the existing
rsi-evidenceowner, with explicit evidence-type separation. Three acceptance gaps should be made testable:Do not turn the PD example into a runnable or validated bundle by citation. At vstack
793c9edc2ee2c0cb21ff38fa8658130e5f452439, the regression example is explicitly an adaptation example without its project models, traces, mutations, runner or experimental evidence. Its five-step procedure distinguishes drift/source-derived checks from replay of already-frozen traces. The reference should enumerate those missing inputs rather than imply copying the workflow provides implementation validation. A frozen replay can remain green after a live code edit; the offline cases should test that distinction, not just repeat “proof is not implementation.”Preserve the actual theorem premises and non-vacuity boundary. The pinned PdCore proof README conditions its theorem on
SizesandDesign, including the narrow completion assumption and disabled unquarantine. It also documents that the four-location map makes some larger instance configurations vacuous. An interpretation example should reject a claim of arbitrary deployment/transport coverage even if the design proof succeeds. Likewise, whole-trace completion, invariant satisfaction, and required abort/release reachability need separate observations; a missing event is insufficient evidence, not proof that the event is safe. I inspected the source documentation, not a fresh proof run.Searchability needs a packaged-consumer test, not only a Markdown link. On the reviewed main,
rsi-evidence/metadata.jsonhasempirical_validation: not_runand a LoopX-only provenance readset; its generated knowledge record inventories the existing resources. Actual source-tree CLI queriesknowledge search PdLedger --jsonandknowledge search vstack --jsonboth returned[];rsi-evidenceremains discoverable. For the proposed addition, require a wheel installed outside the checkout to search the intended protocol terms, open the resulting reference, and resolve the exact upstream SHA, source path, license and scope. Include the reference in the generated record/resource inventory, and assert the two offline interpretation outcomes. Keep source provenance pins distinct from campaign baseline/candidate pins. Preserve the upstream MIT notice for copied vstack material.Validation: read the existing skill, metadata, generated record, packaging and fresh-install test path; ran the CLI queries above. The default interpreter initially lacked pytest. After checking an existing workspace venv, reran
PYTHONPATH=src <existing-venv>/bin/python -m pytest -q tests/test_skill_catalog.py tests/test_knowledge_find.py: all 91 tests passed (one pytest-plugin deprecation warning). These existing catalog/find checks do not establish acceptance of the proposed reference. I did not build/install a wheel, run TLC/TLAPS, capture implementation traces, or run performance/Agent-utility experiments. No code changed.Conclusion: these are concrete acceptance requirements for a bounded knowledge-reference slice. Current main does not yet supply that reference; delivery, implementation fidelity and Agent utility remain unverified, separately.