论文claim matrix
| Claim | 状态 | 最直接实现/证据 | 适用范围 | 不应推出 |
|---|---|---|---|---|
| Pysolate把Host-authorized physical work与Guest logical effect分开记账。 | Current | runtime/capability、runtime/semantic、runtime/preparedregion；prepared-data正负例 | 显式Plan、受支持effect class和固定prototype | 任意Python effect均可speculate |
| AST/candidate analysis不能自行生成authority。 | Current | semantic/prepared-data admission tests | exact target-Guest analysis及Host Plan | analyzer是安全边界或authority owner |
| 未到达exact occurrence的提前结果不会产生logical claim。 | Current + Observed bounded | branch-not-taken、earlier-exception、cancel/orphan tests | semantic pre-dispatch与fixed prepared-data lane | 外部write已rollback |
| append-only streaming source可以在不重新解析final AST的情况下，把已分析prefix occurrence绑定到final source。 | Current | StreamingPrefixAdmission.SealFinalSource、ready-record promotion tests | 已到达、完整、按顺序验证的prefix；final source必须extension | arbitrary source rewrite、dynamic code mutation |
| 多个相同capability/arguments的straight-line occurrence按verified source order逐个claim。 | Current | duplicate-arguments order test | 不同span、每处dynamic occurrence 1、顺序执行 | Broker拥有通用program-counter或任意control-flow bijection |
| Pysolate可以在某些effect-adjacent workload中避免把全部工作退化为serial。 | Current mechanism + bounded evidence | semantic/source-prefix/prepared-data | explicit authority、read-like/immutable准备、足够lead | universal wall-clock dominance over EAGER |
| fixed NumPy prepared-data形成Host receipt、typed object与Guest occurrence的one-shot join。 | Current + bounded observed evidence | research/prepareddataset、P3 v2 | numpy_npy_c_v1、固定dtype/shape/profile/copy bridge | generic object ABI、zero-copy、任意numpy.load |
| private-COW允许多个fresh consumer共享immutable baseline并隔离mutation。 | Current Linux mechanism | Wazero COW tests及历史fixture | Linux memfd/MAP_PRIVATE与固定baseline | 跨平台、完整heap snapshot、prepared-data novelty |
| whole-Run reuse、single-flight、typed materialization和prepared runtime是不同复用对象。 | Current taxonomy | runtime/agentfunction、resultblob、Wazero prepared/COW | 各自明确identity/lifecycle | 一种generic cache抽象覆盖全部机制 |
只有typed unsupported/not_started允许WASM→native promotion。 | Current | placement/orchestrator tests | Host-owned placement与fresh child attempt | 普通异常、ImportError或post-effect retry |
| native attempt结束后，后续portable-value request可以重新选择WASM。 | Current design/implementation | state class与placement | 新request、无native process/workspace依赖 | native continuation迁移到WASM |
| Host receipts与terminal disposition提供effect truth和ambiguity信息。 | Current | Broker、native RPC、verification | 系统观察到的dispatch/lifecycle | external-world truth、完整事务rollback |
| Prepared Family允许多个fresh single-use member复用一个bounded NumPy input/image，同时保持Invocation、workspace和terminal state逐member私有；需要capability的member还使用自己的Plan/Broker。 | Current + bounded acceptance | runtime/engine/wazero/prepared_family*.go、docs/prepared-family-v1.md | 显式opt-in、numpy-core、bounded numpycodec dtype/shape；portable private-copy与Linux exact-artifact private-COW | generic object family、scheduler/retry、workspace merge、zero-copy或fan-out speedup |
| Prepared Family的physical disposition可在同一合同下选择private-copy或qualified Linux private-COW；显式COW失败不会静默降级。 | Current API；Experimental COW substrate | prepared_family.go、COW/real-Guest tests | `auto | private_copy |
| synthetic exact-Guest control可以通过source-bound pre-dispatch gate。 | Current + Observed bounded | source-bound-pass-authored-workload-evidence-v2.json | 1个显式构造的eligible read，1 admitted / 0 rejected | 自然workload中常见 |
| 两个checked-in synthetic overlap fixtures观察到early read与source tail重叠。 | Historical Observed | source-prefix-overlap-v1.json、source-prefix-day-trip-evidence-v2.json | 固定1.5秒read与约1.4秒source tail；1.914x-1.923x | natural或production speedup |
| 六个coding-shaped cases与36-event sample没有admitted prefix opportunity。 | Observed bounded negative | source-bound v2 census、source-prefix opportunity census | 两个固定样本的prevalence | 机制没有positive workload或无法加速 |
| compile-time source-pass registry可以新增一个不在central name switch中的whole-program pass，并以unchanged original RunRequest执行exact-Guest patch。 | Current + Observed bounded | runtime/passplugin、runtime/sourcepatch、source-pass-plugin-v1.json | pure_scalar_cse的closed scalar program内相邻int64/bool子集；call/control/introspection整条拒绝 | 动态插件系统、任意paper pass已兼容，或性能收益；该fixture treatment实测慢14.61% |
| stratum的exact constant-folding kernel可以收窄为独立whole-program pass。 | Current + Observed bounded | pure_scalar_fold、paper-pass-absorption-v1.md、pure-scalar-fold-paper-pass-v1.json | closed top-level bool/signed-int64 program及total +/-/*；import/call/control/introspection整条拒绝 | stratum完整DataOp optimizer、其他论文kernel已实现或一般speedup；保留fixture treatment慢4.49% |
| prepared-data P6比较的是component timing下的projected schedule。 | Historical model evidence | phase6 campaign/aggregate | 固定fixture和lead-gap model | observed interval trace、生产critical path |
| NumPy result reuse在固定Phase 7矩阵没有达到break-even。 | Historical Observed negative | phase7 economics rows | 指定payload/consumer/gap/platform grid | reuse永远不经济 |
主claim推荐措辞
Pysolate demonstrates that fresh isolated execution can retain optimization opportunities when authority, physical work, and effect truth remain Host-owned: eligible work may start early, while the original Guest occurrence remains the only point that can commit the logical effect.
EAGER比较推荐措辞
Compared with an EAGER persistent-interpreter baseline, Pysolate does not provide a universal latency dominance theorem. Its contribution is that explicit authority and discardable Host-owned preparation expose safe overlap and reuse opportunities that need not serialize merely because later execution may contain effects.