打开文档导航

Pysolate benchmark 与 evidence catalog(单文档审查)

0. 审查范围、冻结对象与判定词

  • 审查对象/Users/yuzhe/projects/agent-python-runtime 的精确 commit b94c989785632d98634cdd5c2a6d4a1de6cc13ac,tree 6b8bb6a8d8274b8329a2aba444d13eac9fac19e9
  • 提交状态git verify-commit 为有效 ED25519 签名;审查前工作树无 dirty/untracked 输出。
  • 方法:先读取 WORKER_BRIEF.md,再用该 commit 的隔离 archive 做只读盘点。没有重跑 campaign、CI、provider 或外部数据收集;没有改实现仓库、commit 或 push。
  • 本 catalog 的 target-bound 定义:只有 producer/harness/checker/artifact manifest 明确绑定上述 commit/tree,才可称为 target-bound。证据文件虽然被 checked-in 到目标 commit,并不等于其数字由目标 commit 产生。
  • 状态词Current 表示目标树中的合同/实现/证据索引仍在;Observed 表示 canonical artifact 记录了观察;Historical 表示 producer 或运行 artifact 来自目标之前的 commit;Framing 表示结构性或设计性边界;Deferred 表示被 gate 拒绝或尚未运行;Unsupported 表示不能由该证据推出。
  • replayable 词义canonical-recompute 只表示可从 checked-in canonical JSON 重算摘要;raw-replay 需要原始样本、artifact、harness 和运行环境,不能因有 SHA 字符串就宣称可重放。

论文主张应保持的连接

Pysolate 允许 Host 对显式授权、identity-bound、可丢弃的物理工作做提前 dispatch;未改变的 Guest dynamic occurrence 仍是唯一可以 claim logical effect 的位置;无效或未到达的工作被拒绝或 orphan,而不会继承后续 authority。

因此,本文将 analysis/candidate、物理 preparation、typed receipt、logical claim、fresh Guest、workspace state、native backend 和 projected schedule 分开记账;不把 component benchmark 当成生产延迟,不把 authored fixture 当成 natural uplift,不把旧 producer 当成当前 target certification。

1. 总览:paper claim 聚类而非逐 JSON 清单

研究问题证据主线最强可支持的 claim当前审查结论
fresh execution / authority20-program transparent fixture;phase5r fresh scratch/final controlsauthority、fresh Run、logical/physical identity 与 terminal disposition 可被记录和拒绝Current mechanism / Historical benchmark;target-bound 否
streaming overlapauthored source-prefix 3-pair;natural structural census在固定 source tail + early read fixture 中可观察 overlap;natural cohort 的机会频率另算Observed authored;natural no-go;target-bound 否
semantic pre-dispatchexact-Guest G1/v0;Phase 4 matched campaign;Phase 5r no-go狭窄、default-off、exact live-only read 可在未改 Python call boundary 下 claim;derived-AST 扩展只保留 mechanism no-go/成本诊断Current bounded consumer;timing producer historical
workspace / reuse / COW20-program workspace/authority fixture;NumPy reuse Phase 7exact sharing、fresh private materialization、COW 选择和 failure attribution 可验证;经济收益未出现Observed fixture;performance default rejected
prepared dataNumPy contract → physical read receipt → typed decode → private seal → exact object claim固定 numpy.load/.npy/numpy-core 可做 authority-preserving prepared-data proof;不能外推通用 Python/zero-copyCurrent contract;producer evidence historical,phase6 timing 不可认证
native / backendnumpy-core static-native profile;native sandbox lifecycle;Open-SWE placement controlprofile-bound native qualification 和 identity-bound native sandbox boundaryCurrent bounded backend;非通用 native correctness
natural cohort / no-goCodeAct/Open-SWE corpus;tau2 remediation;source-prefix census;Phase 5 no-gocorpus/denominator、机会缺失、证据缺口和停止条件可审计Framing/Deferred;不支持自然性能提升

2. 公共 authority/fresh-execution spine

2.1 问题、调用路径与边界

研究问题:当多个逻辑请求具有相同的 sealed qualification 时,能否在 fresh Run、private workspace attempt 和当前 authority 不混淆的前提下减少物理执行,并把取消、过期、widening、预算超限和 parent terminal 记录为拒绝/取消而不是成功?

目标树中的实现锚点

  • research/workflowbench/campaign.goCanonicalTransparentCampaignCampaignManifest.ValidateCanonicalCampaignPlans
  • research/workflowbench/campaign_execution.goCampaignRequest(刻意不含 program ID/family/Expected)、RunTransparentCampaignCampaignRuntime.PhysicalValidateCampaignEvidence
  • research/workflowbench/campaign_runtime_adapter.goCampaignAdapter 与 typed execution kind 的 adapter 边界;
  • runtime/capabilityruntime/workflowruntime/workspace:Host-owned Plan/Broker、fresh Run/resume、workspace root 和 terminal disposition;
  • integration/e2e/campaign_guest_test.go:real-Guest wiring;research/workflowbench/evidence_layers.go:EvidenceSuite.Validate:证据层合同。

路径为:固定 manifest/Plan → adapter 收到不含 paper case label 的 CampaignRequest → Runtime admission/physical slot → Host-owned capability/child Plan → fresh Guest/Run 与私有 workspace attempt → Broker/typed consumer → result/terminal event → evidence recorder。共享只发生在 exact identity;近似 source/input/privacy/root 变化必须在 physical execution 前拒绝。

边界:Host owns Plan、grant、handler、workspace root identity、physical execution ID 和 disposition;Guest owns unchanged Python occurrence、logical call/result;shared 只有 sealed contract/receipt;private 是每个 Run 的 workspace、staged result、body-safe evidence。fixture contract (P01P20) 不是 Runtime API,也不是 virtual Git 或生产 scheduler。

2.2 Evidence row

artifact / producer / harness / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/authority-transparent-campaign-v1.json; campaign source 40882ca5a818f4c5388bdeebe7d36ee9dc5fe7c5; Guest artifact source ae922641cd9c539b68a0ea7110b5dc205e5c9a8a, artifact sha256:0a37a963a09b4e763cb6a40886a771e9c13e2f6a9d3a2d295788752e319c5795; manifest sha256:0633e6d98dd67fee6a2aad12cfd491a6d14e5344d5d2d78d91c059e62ec0fe7e; checker ValidateCampaignEvidence,但没有独立 checker commit/digest;canonical sample = 10 run rows + 20 programs + 148 walkthrough eventsDarwin/arm64,Go 1.26,Apple host;固定 20 个 logical programs、3 physical slots、5 balanced paired repetitionscanonical runs 重算:baseline/qualified physical median 19/17,每对 reduction 2;wall median 25,728.820333/22,004.745333 ms,paired wall reduction 3,724.075 ms;文档四舍五入为 25.73/22.00 s支持:fixed fixture 中 exact request/root verification、fresh authority resume、child attenuation 和 rejection disposition 可重建;观察到少 2 次 physical execution。禁止:arbitrary workload、production throughput、semantic equivalence、external-effect safety、virtual Git、provider quality。Observed + Current index;producer Historical;target-bound ;canonical-recompute ,target-tree raw-replay (没有独立 raw log/Guest body)。
docs/evidence/semantic-speculation-phase5r-mechanism-evidence-v1.json + phase5r-harness-freeze-v5.json; guest artifact sha256:62454f9689ae4a11e45d51398e1d605be91b58b472eaafc26a994cb5669f62e9; harness identity sha256:ea88c6f2de16e5932c5211158b94c9d1efaa5d598096a0643def9276e153d042; source-set file hashes在 freeze 中;checker scripts/review-semantic-speculation-campaign.pyDarwin/arm64 与 Linux/amd64;11 个 bounded mechanism cases;Linux 要求 private COWcase_count=11, passed=11, failed=0;fresh independent scratch/final、logical_calls=0、Broker/workspace unavailable、fallback/replay/authority expansion 均 false;single-use capsule ≤256 bytes;economics records 0支持:authority-free derived mechanism 的 fresh lifecycle、source/AST binding、error/traceback parity、unclaimed/cancelled discard contract。禁止:生产执行、timing advantage、通用 derived-AST optimizer、Broker/workspace claim。Current mechanism / Deferred economics;producer Historical(freeze source/artifact 不等于 b94);target-bound ;canonical structural check ;raw-replay

2.3 不变量、失败语义与 trade-off

  • 不变量是“物理一次”与“逻辑一次”分离,而非物理一次自动等于逻辑成功;issues=starts=finishes=logical_claims=1 才是成功等价形状。
  • P05/P06 exact identity 可共享;source/input/privacy/root 任一字段 near-match 必须拒绝。P16/P18/P19/P20 等 authority expiry、widening、budget、parent terminal 是负控,不应从总成功数中删除。
  • 选择 fresh Guest/private workspace 是显式牺牲;它避免 persistent interpreter、跨 privacy 和 retained authority,代价是 provisioning/teardown。论文只能写 opportunity-level result,不能写 dominance theorem。
  • phase5r 的 no Broker/no workspace 是 authority-free mechanism 约束,不是“可直接执行任意 agent source”。

defense-ready:我们证明的是“可重建、可拒绝、可归因的 authority/fresh execution contract”,而不是把 fixture 中的少数 sharing 关系包装成自然吞吐率。


3. Streaming overlap:authored mechanism 与 natural census 分开

3.1 Authored source-prefix overlap

研究问题:exact Guest、reach-gated 的 source-prefix execution,能否让一个早期 Host-mediated read 与剩余 Python source production 重叠,同时保持 result、dispatch count、Plan identity、workspace state 不变?它不是 dynamic DAG,不是并行 Python statements。

源路径与 lifecycleresearch/workflowbench/source_prefix_overlap.goSourcePrefixSchedule.ValidateExecuteSourcePrefixPairsValidateSourcePrefixEvidenceProduceTimedSourcesource_prefix_overlap_test.go 的 contract/negative tests。流程为固定 contract/lane config/oracle → baseline generate_then_execute 或 streaming stream_while_generating → unchanged Guest reaches ordinary Broker → result/oracle/workspace hashes → paired evidence。

artifact / producer / harness / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/source-prefix-overlap-v1.json;artifact source 501daef99796c1af7cd7bab1e0ab712a199820b9,artifact sha256:a443042fb080d22f8e352aca0d0c8a5c87a7801e8afcc603e174d75fbe11c69b;harness ca25b1b767edd50dc25363df5347cb801c5c183a;experiment identity sha256:dab34bfa2a6ea8dce909c375c0b963569cfc67f988fa1adae56de561b1b009ff;sample rows 6、3 matched pairs;checker ValidateSourcePrefixEvidencefixed authored 3 chunks at 0/700/1400 ms;early Host read 1500 ms;source-generation tail 1.4 s;机制时钟排除 Guest initializationcanonical rows 重算:baseline median 2,950.046792 ms;stream median 1,533.922917 ms;median saving 1,416.123875 ms,ratio 1.923x,wall reduction 48.00%;6/6 oracle、workspace before=after、每 lane 1 logical/1 physical、fallback 0。3/3 streaming rows 的 tool_started < generation_complete支持:该 authored fixture 中 reach-gated source-prefix overlap 的 observed interval trace 和 result/workspace preservation。禁止:natural/provider/production speedup、dynamic DAG、parallel statements、safe speculative writes、仅凭 read_only/idempotent 自动 dispatch。Observed authored mechanism;producer Historical;target-bound ;canonical-recompute (rows 已含 event timestamps);raw-replay (原始 process log/artifact 不在目标 tree)。
docs/evidence/source-prefix-day-trip-evidence-v2.json 是另一组相同形状的 3-pair authored fixture(artifact source cb32e4f5ec942620d82c3918aa7607b76af9185c、harness d7baa92918f1954dc6584cd4f537cc1e817d8f8d、artifact sha256:2da82e65061a9680dc5921150ae50ea44766590fc44e2a6fa554784f2ce3167e)。不要与 overlap-v1 合并为 n=6;只可作为重复的 bounded mechanism indication。同上,3 pairs独立重算 baseline/stream median 2942.787333/1537.471584 ms;sample rows 完整但仍没有 raw process bundle。只能支持第二个 authored fixed case 的同型行为;不提供 natural prevalence 或 statistical significance。Historical / secondary;target-bound ;canonical-recompute ;raw-replay

3.2 Natural structural opportunity census

docs/research/source-prefix-opportunity-census-v1.mddocs/evidence/source-prefix-opportunity-census-v1.jsonNATURAL_STRUCTURAL_OPPORTUNITY_CENSUS_NOT_PERFORMANCE_EVIDENCE 是 zero-provider structural census,不是 overlap timing。

  • producer/harness/artifact:Guest source 501daef99796c1af7cd7bab1e0ab712a199820b9、artifact sha256:a443042fb080d22f8e352aca0d0c8a5c87a7801e8afcc603e174d75fbe11c69b、census harness 61112106e20959e5894414ca991f8bac2699dd92;parent remediation sha256:5f504a138084f933ba0fd4f3bec7aede7076924ec3c2a5cfb8f05db3dd9a513f;public identity sha256:13120c7ec8565fe7599c0c3f362a0ae90deeb67cafdd986dafa4a8cac70d714a
  • denominator/platform:冻结 remediation-v2 READ cohort 36 events、30 unique source bodies;每个 source 为 48–57 bytes 的 body-free digest projection;exact CPython/WASI Guest analyzer。
  • 结果structurally_eligible=0structurally_ineligible=36timing_not_recorded=36;每个 READ 都在唯一/最后 candidate region,trailing regions 0。这些 counts 可从 canonical JSON 重算;没有 provider/chunk timing/raw body,不能重算 latency。
  • 支持 claim:该 frozen cohort 没有 preregistered source-prefix overlap admission opportunity。禁止推论:不是 prevalence outside cohort,不 refute authored result,不支持 natural speedup/provider timing。
  • 状态Framing + Observed no-go;producer Historical;target-bound ;structural canonical-recompute ;raw-replay 。主文只建议放一行 boundary/no-go,详细计数放 appendix。

paper placement:authored interval trace 可放主文 mechanism figure;natural census 只放主文 caveat 或 appendix gate,严禁把 projected/natural census 与 observed authored timings pooled。


4. Semantic pre-dispatch:窄 consumer、formal timing、derived-AST no-go

4.1 Current G1/v0 consumer

研究问题:能否从 exact Guest source occurrence 生成一个 sealed QualifiedCall,在 physical read 开始前 consume Host PreDispatchBudget,并在 unchanged Python ordinary Broker boundary 原子 claim 同一个 streaming.StagedObservation,而不 duplicate live request?

目标树锚点runtime/semantic/streaming_predispatch.goCanPreissuePreDispatchBudgetExecuteSemanticPreDispatchruntime/capability/streaming_predispatch_test.go 的 Broker revalidation;runtime/semantic/streaming_predispatch_test.go 的 lifecycle/terminal cases;integration/e2e/semantic_predispatch_test.go(real CPython/WASI exact Guest)。语义包不创建 implicit goroutine;Host launcher 负责 capacity,child physical context 负责 cancel/wait;late/unclaimed result 必须 orphan/discard。

artifact / producer / harness / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/semantic-predispatch-experiment.json;artifact sha256:36baf5a3897e9a7c1894b9e0f1e717e159649c1bc840fc5bce95a943e4d76eaa;source SHA sha256:ff213e9f20b05b78659fda93ffc2683dfb37fca2eee2ef02e74de499febf4a0d;Plan SHA sha256:a1b37bd12484e264e881f98e4b677ed79f55ecd1c3593755075474f16e252533;checker 由 report validator/real-Guest test 约束,canonical 未记录独立 checker commit;sample rows 105 baseline + 5 semantic_pre_dispatch;physical delay 1,000,000 µs;real Guest fixture从 10 rows 独立重算 baseline/optimized median 3,214,550/2,196,151 µs、saving 1,018,399 µs;每 trial logical/physical calls 1/1;optimized issues=starts=finishes=1、rejected claims 0;result digest 每行相同。支持:one exact live-only read 的 bounded mechanism/timing indication,unchanged call boundary,无 duplicate physical call。禁止:通用 semantic scheduler、captured-playback、writes/unknown effects、arbitrary Python、production latency。Current bounded consumer + Observed fixture; source/artifact provenance Historical/不完整;target-bound ;canonical-recompute ;raw-replay
docs/evidence/semantic-predispatch-day-trip-v1.json 是同 schema 的另一 5×2 day-trip sample;artifact sha256:d5706fbf113c7042a4484ad5713ee5baa8fe4788c33beb9b6223b0ff9f1201af、source SHA sha256:0a4920d4356e37f733517065e92eb46eb0f39fe4e2c67722d030177717ecbcb0、Plan SHA sha256:e8944dfef1bd283caf18a153ba14e7aeb879d92a808852cae91f5cdb2bc148595+5 trials,1 s delay独立重算 3,149,103/2,132,259 µs,saving 1,016,844 µs;同样每行 1/1。不与前一报告合并成 n=10 的单一实验。同上;重复 fixture 不能制造统计显著性。Historical secondary;target-bound ;canonical-recompute ;raw-replay

4.2 Formal Phase 4 matched campaign(历史 timing,非 target certification)

docs/evidence/semantic-speculation-phase4-campaign-evidence-v1.json:campaign source 66caa186dcf37e3432813adf46b0cf6d642d912d;artifact sha256:8780338cf3b4330371b13f06a2846006077c3ff99ee89d7fb618ea19e252d242;aggregate validator commit 8f5398d5f2083005181a6c685a926cb0087a1d97;campaign log/stdout/raw-record digests 分别为 sha256:52507229d9af457fe51f78aa29e0429811b765c7bef1a5bda3ef6a3b5a5252bdsha256:2eff2acbc68c67f34374c14dbd49bdfd88f81a5758494c037b2b158d655581b7sha256:b12994b6ee8257e0f9c9d4a0acdab0de793c322a6ba6029710903452e4d8550e。平台为 Linux/amd64(gpu31,private COW),360/360 records、120 matched cells、每 coordinate 5 ready trials,6 个 passing coordinates(3 cold + 3 preprovisioned)。

  • canonical aggregate 中可直接复核的 savings:cold 445,820,706432,933,163460,582,130 ns;preprovisioned 2,435,380,5832,468,139,3932,471,417,846 ns
  • validator 语义:research/semanticspeculation/phase4_aggregate.go:AggregatePhase4Campaignscripts/review-semantic-speculation-campaign.py 系列检查 complete matrix、outcome parity、mechanism/economics gate。报告支持 fixed named synthetic regime 的机制/经济结果,不支持 natural prevalence、production speedup、universal Python equivalence。
  • raw 限制:目标 tree 只有 canonical report 和 raw digest,没有 360 条原始 trial JSONL;因此可核对 aggregate schema/数量,不能从 target tree 独立重算每个 median 或重放 Guest。
  • 状态:Observed Historical;target-bound ;canonical aggregate structural check ;raw-replay

4.3 Phase 5 derived-AST no-go

docs/evidence/semantic-speculation-phase5-mechanism-gate-evidence-v1.json 的 parent gate 保留 11 case 的 pass/fail 及 phase5_gate_passed=falsescripts/review-semantic-speculation-phase5-mechanism-gate.py 明确要求 failed case、timing_campaign_started=falsetiming_samples_observed=0。Phase5r remediation 的 11/11 是 bounded mechanism controls,不是 economics promotion:economics_campaign_started=falseeconomics_records_observed=0;cost evidence 仅 4 个诊断 records(macOS/Linux × original/derived),且 fresh Guest provisioning dominates,不能把 critical_wall 误写成 observed user latency。

failure semantics:invalid suffix、pre-region exception、pre-cancelled 只允许 claims=0, consumed=0, discarded=1;after-region exception 可保留 parity claim;wrong capsule/selection identity 必须 reject without consumption。trade-off:分析/compile/scratch/final separation 换取 authority-free 和可审计性,牺牲 provisioning time;因此 no-go 保留 original execution,不扩展 transport/authority。

paper placement:G1/v0 exact live-only shape和一个小型 real-Guest result可进入主文;Phase4 full matrix、Phase5 no-go、cost profile放 appendix;绝不把 Phase5r 4 cost rows 写成 benchmark speedup。


5. Workspace / reuse / COW:机制成立不等于经济成立

5.1 NumPy result reuse Phase 7

调用/数据路径numpy-core producer 在 verified Wazero Guest 中按 numpy-producer-admission 生成 bounded C-contiguous ndarray → Host owns immutable blob/descriptor → 每个 consumer 在 fresh Guest private materialization → lease consumed/cleanup。相关 source:research/numpyreuse/campaign.go 的 campaign/cell aggregation 与 reuse lifecycle;scripts/review-numpy-result-reuse-phase7.py 的 summary/checksum validator;scripts/review-numpy-producer-admission-phase6.py 的 body-safe lineage/adversarial checks;runtime/engine/wazero/cow_memory_linux.gocow_runtime_linux.go 的 Linux COW substrate。

artifact / producer / harness / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/numpy-result-reuse-phase7-v1.json;harness source 1a6596d2cd238e6c441b7ffa798ecb9b1c01c5e9、tree d98612fa162c9eded44e4d6cf82f52f471cc5cd4;artifact sha256:2753cde560f3961a483df53aec334c8fdbb084934e5a62a56d436aea1ae557ad;campaign identity sha256:fa6fa1a8b68df5eb0fc5070660609a9800062769789fcd5f9c0a107680184e1e;raw JSONL identities macOS sha256:01eb1a864760a1fbf732b20f3f31972dc5f0c6f9fb54484413f45894771df9f3、Linux sha256:cfc87552b05ab4122c7aad0fdb3ea4ad31e95ae6f003f7988734e57e39222374;checker scripts/review-numpy-result-reuse-phase7.pymacOS arm64 + Linux amd64;240 raw records,80 cells,40 economic comparisons;每平台 120 records = 60 cold + 60 preprovisioned,3 trials/treatment;payload/consumer/gap coupled sparse grid从 checked-in economics 40 rows 独立重算:macOS cold median ratio 0.460100,preprovisioned 0.520715;Linux cold 0.432827,preprovisioned private COW 0.430282;best 0.795219;40/40 observed_break_even=false。mechanism summary:all fresh guests/result parity/no authority expansion/no replay,leases consumed 216,reuse blobs consumed records 120,placement fallback 0支持:closed numpy-core typed transport 的 identity/private materialization/COW behavior;在该 coupled matrix 中没有 break-even,故拒绝 performance-default、fan-out/single-flight expansion。禁止:pandas/object/non-contiguous/arbitrary ndarray/pickle/durable cache/cross-Run retention/shared memory/generic native plugin;不能从 coupled grid 插值 independent thresholds。Observed mechanism + Observed negative economics; producer/harness Historical;target-bound ;canonical-recompute (40 economics rows足够);raw-replay (JSONL 仅 digest/外部 private artifact,不在目标 tree)。
docs/evidence/numpy-result-reuse-phase5-{darwin,linux}-v1.jsonnumpy-producer-admission-phase6-{darwin,linux}-v1.json:producer source 16c141d051c43a2a89383336b3c4ca11fe9bb0c5/tree c9bd3255325b12c6ad54b1a78772b439ddb0d9dd(reuse probe)及 1d788057d3c183dbdafb28030a95967863ba63cd/tree c4518d7ff6cbdc7b14f39722a08d3b7b3ed0ca82(admission);artifact同上;checker scripts/review-numpy-producer-admission-phase6.pyDarwin no private-COW candidate/fallback;Linux private COW selected,prepared image 512 MiB virtual / 4 KiB pages;fixed fixture body 48 B (<i8, shape 2×3) 和 bounded operation arange_affine_i64_v1两平台 result parity、0 capability calls、body-safe descriptor、Linux no fallback;adversarial rejects object dtype、noncontiguous、Fortran、random/time/file/dynamic import/stale source/inputs。支持 profile admission and typed body-safe bridge;不能把 Linux COW 结果当跨平台默认,也不能当 zero-copy 或 generic native ABI。Current contract / Historical probes; target-bound ;canonical structural-recompute ;raw-replay

5.2 Workspace wording 与 COW failure boundary

  • workspace_fixture_sha256 只标识 initial fixture declaration,不是 Git commit、merge/rebase 或 semantic workspace equivalence。实际 immutable Pysolate workspace root 必须由 Manager 单独记录。
  • Linux COW 是 physical execution substrate;macOS probe 明确 memory_cow_candidate=false、fallback true,不能把 Darwin 与 Linux 结果合并为同一 backend。
  • consumer_mutation_is_private=true、lease consumed、store retained bytes 归零是 lifecycle invariants;它们不证明跨 Run durable cache、pointer/heap transfer 或外部 effect rollback。
  • 独立复核不应把 reuse penalty 与 producer/consumer compute 分段相加;Phase 7 的 raw JSONL 才是完整 sample identity,当前 tree 只有 body-safe summary。

paper placement:exact reuse/COW identity和全负 economics 放 appendix 或 mechanism/limitations box;主文只保留“安全 reuse 可实现但不在该 sparse matrix 中经济成立”的一句话。


6. Prepared data:Host physical preparation 与 Guest logical claim 的 join

6.1 Contract 与 end-to-end path

研究问题:在 final source 尚未释放时,Host 能否依据显式 numpy.load contract 做 bounded physical read/decode/seal,而不让 candidate syntax mint authority;之后只有 unchanged Guest exact occurrence 才能 claim 同一 staged object?

目标树锚点

  • docs/research/authority-preserving-prepared-data-contract-v1.md:contract、P1–P6 gates、promotion boundary;
  • runtime/prepareddataset/prepareddataset_test.go:prepared contract/receipt/object lifecycle;
  • runtime/capabilityruntime/streaming:Host receipt、private staged observation、exact claim;
  • integration/e2e/prepared_region_helper_test.goruntime/engine/wazero/prepared_region_scratch.go:fresh Guest/claim/cleanup controls;
  • research/prepareddatasetcampaign/campaign.goBuildrecordsForaggregateoverlap
  • docs/evidence/prepared-data-phase3-v2.json:positive object-bound join + Linux negative control;
  • docs/evidence/prepared-data-phase7-closeout-v1.json:promotion verdict 与历史 evidence inventory。

路径为:numpy.load exact source occurrence → Host contract/Plan (sources.read, immutable workspace root, profile numpy-core, plan_epoch) → physical read receipt/file SHA/source verification → bounded .npy typed decode → Run-private seal → unchanged Guest checks exact final source/occurrence → exact object claim/one logical result → consume or orphan/dispose。Receipt、typed metadata 和 claim identity 是不同层,不能互相替代。

6.2 Evidence row

artifact / producer / harness / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/prepared-data-phase0-preregistration-v1.json;contract pysolate.prepared-data-contract.v1;fixture body 8,388,608 B、file 8,388,736 B<i8、shape 1024×1024、body SHA sha256:a78cee677876b925402c15818acd3fc020a47754d9d1c26688914ea09070f8d0、file SHA sha256:390bd7abd4f26dfdf3440f060f77bd02526e97e0a46838337f56cb8ef978af7epreregistration dimensions consumers 1/2/4、payload 8 MiB、platform Darwin/Linux;不包含 generic allocator/fanout/durable cache。contract fields 可直接核对;不产生 execution sample。支持 paper contract scope;禁止把 preregistration 当 observed timing/authority。Current contract / Framing; target-bound 是(contract file在target),但不是 benchmark producer。
docs/evidence/prepared-data-phase3-v2.json;source commit 3eb304097edae1af20fcb8af0a22032ceffaf195、tree a818c7e0057ce9a1b46f65c48e61b66ab61aab5f;positive probe c671284fc8b2457baeed5385232a5b37489fe7e7;object join commit 1fa7f99cfbc5145a6873af58795707e85618d28e;artifact sha256:2753cde560f3961a483df53aec334c8fdbb084934e5a62a56d436aea1ae557ad;platform raw digests Linux sha256:0e6745256847c6bc885cd80b45c0ef0c99521e47e3aeee4cac9d0c8b042a4f32、macOS sha256:3567f887e435f7e1861d094a27cc6b34efb3cb14dfd279e819f6c282a8549438;negative raw sha256:545c8119314a9e2d8ff1ad3fa073345a409654029e6b554e9426cdbf5392fcb3。checker 为 focused tests + prepared-data probe validators;没有单独 checker digest。Linux x86_64/AMD Ryzen 7 5700X,Darwin arm64/Apple M4;每平台 1 positive real-NumPy Guest,另有 Linux 3-case negative controlPositive 两平台:read_before_final=truedecode_before_final=trueauthorized_starts=1no_contract_starts=0token_claims=token_consumed=1、logical parity true、physical read/decode/materialize/seal 各 1;负控 reached = claim 1/consume 1/orphan 0,branch-not-taken 与 earlier-exception = claim 0/consume 0/discard 1/object Orphaned/physical orphan 1。支持:固定 fixture 的 Host receipt 与 exact staged-object claim join,未到达/异常安全 orphan;禁止:generic zero-copy、arbitrary numpy.load、external-world truth、candidate discovery=authority、成功 fixture=arbitrary Python。Observed bounded E2E; producer Historical;target-bound (但该 evidence file 的负控改动本身进入 b94);canonical structural-recompute ;raw-replay
docs/evidence/prepared-data-phase4-phase5-linux-v1.jsonprepared-data-phase4-linux-amd64-v2.json;checker commit 09a0bb12e5d0ffc160dd526b737c378db17bb77e/tree a579f387ad623e94e4465494952d5b72249fbbb9;artifact同上;raw report sha256:97afae0927aa3a275454478daca7e166f96a1aaa99668f523148a669d6ba7fda;Linux private COW selected。Linux/amd64;固定 one-time NumPy shard;P4/P5 mutation-isolation、package prepare once、consumer fresh guestbody copy per consumer 0、mutation isolated、package preparation once per treatment;但 economics_gate=not_formal_single_trial_order_and_warmup_confounded支持 physical substrate/copy isolation;禁止写成 formal economics 或 COW universal result。Historical mechanism; target-bound ;canonical check 部分;raw-replay
docs/evidence/prepared-data-phase6-linux-v1.json + prepared-data-phase6-manifest-v1.json;campaign source ccef2d9875ab2f289434012bbdfb4015b99db6b1/tree 6ecc38d09d37be9eaaab51d2765ca8930592faa9;artifact同上;manifest prereg sha256:9f7baa064eff8e19c93651b41decf4f855673fcc5ae767716f023d3de4702bd6;raw rebuild SHA sha256:82f97a56c1a8b830ab5a96ab6040afb5bdaa8792b435c347667555e6fc8372ca;checker/closeout references prepared-data-phase7-closeout-v1.jsonLinux/amd64(manifest同时固定 Darwin/arm64 host metadata);3 trials × consumers 1/2/4 × gaps 0/250/1000 ms × 6 treatments = 162 records / 54 aggregate coordinatescanonical 中 aggregate_count=543×3×6=54phase6_coordinates=162 可独立核对;但 aggregate 的 median critical path 来自 Build.recordsForoverlap(prep,gap)+consumer projection。raw JSONL 不在 target tree,不能独立重算 median、不能证明 read/decode/finalization interval overlap。支持:固定矩阵的 projected schedule comparison、parity/cleanup fields 和 fixture economics framing。禁止:observed interval trace、critical-path user latency、formal break-even、target certification;旧 producer 不能替代 b94 runtime evidence。Historical / evidence limitation; target-bound ;coordinate/count recompute ;median/raw-replay
docs/evidence/prepared-data-phase7-closeout-v1.json:closeout source target db1824feb57bff7c182617451d7c0fe40d923f07/tree 08306efa86c9a8d83de472ea649972b838a8720e;reviewed target 17dec479dcf08853f21617578ee0f93968942ea6/tree 11c624ed7ff3b0ccb71bd6ebea59b15f457afe8c;verdict not_certifiable;P8/productionized false。closeout inventory 包含 phase0–6 evidencecloseout 自己保留的 blockers:no object-bound Host materialization join、CPython preparation retains WASI clock/random、Phase6 critical paths projected not observed;validation 中 phase6 aggregates 54/coordinates 162。支持“prepared data novelty 仍需严格 join/authority/observability gate”;禁止用 closeout 的 full_go_test/race/vet=pass 反推 target b94 已 certification。Current closeout record / Historical verdict; target-bound ;raw-replay

6.3 明确的 projected-vs-observed 判定

research/prepareddatasetcampaign/campaign.gorecordsFor 直接以 overlap(read+decode,gap)overlap(cow.DatasetPrepareNanos,gap) 等公式构造 CriticalPathNanosoverlap 仅做 prep-gap 的 schedule projection。它没有 per-run read-start/read-end/decode/seal/claim event interval。故:

  • phase6 的 54 aggregates 是 projected schedule / canonical derived output
  • phase3-v2 的单次 positive probe虽然记录了 read_started_nanos/read_completed_nanos/decode_completed_nanos/final_source_released_nanos,可支持该 probe 的 causal order,但不能替代 phase6 的 162-record observed campaign;
  • 没有 raw report 时,任何从 median_critical_path_nanos 反推 interval overlap 的句子都应删掉。

6.4 Prepared-data invariants 与 rejected alternatives

  • syntax fact 不能 mint Host contract;contract 必须绑定 source span、canonical args、host Plan、profile/import closure、freshness、budget、privacy 和 file/body identity。
  • staged object 的 ID 不是任意 non-empty token;Guest claim 必须消费同一 body/file/source/contract-bound object。phase3-v2 新增的 branch-not-taken/earlier-exception negative control 正是防止 scalar token tautology。
  • body-free evidence 只证明 typed receipt/claim join,不证明 external-world truth;Host owns immutable body,Guest 只看到 typed metadata/result。
  • 不采用 generic allocator、public fanout、single-flight、durable cross-run cache、pickle、pointer/FD transfer;private-COW 是 substrate,不是 prepared-data novelty 本身。

paper placement:contract、positive/negative claim lifecycle 可进主文;phase6 schedule projection、phase7 closeout blockers、raw missing 说明放 appendix/limitations;不要把 phase6 数字写成 observed interval trace。


7. Native/backend boundary

7.1 numpy-core static-native profile

问题:能否把一个 source-locked、profile-owned 的 NumPy subset 作为 bounded native backend qualification,而不把 native imports 误写成通用 purity/DBI 或 arbitrary Python support?

  • source/artifactdocs/evidence/numpy-core-phase2-mechanism-evidence-v1.json 绑定 source commit 47070b1e2ae09e6292f2a5ce0dd9f8c4ee9f6a6c、tree 7fa3d93dbae011207a82f4998925d23033aed8c9;artifact 1b759febddb9dd0d65760c74b21b07c4f04fa992ce761639fbb9baef155926aa、manifest eea4e8ead6be395c02612f3dc803ab61a0b409caf78189c1bbf4596cc4f2d055;builder gpu31,WASM/WASI,CPython 3.14.0/WASI SDK 33。
  • harness/checkerscripts/review-numpy-core-phase2.py 验证 RESULT.READY、SHA256SUMS、manifest/inventory/qualification/build cache、profile identity 和 negative controls;它需要外部 artifact_root,target tree 没有 artifact bundle。
  • 样本/数字:native modules 19、support archives 2;fresh process runs 2,results equal,capability calls [0,0];NumPy 1.26.0b1 C-contiguous oracle qualified;filesystem write 返回 FileNotFoundError 且 no result/host file;base profile substitution 在 Guest 前拒绝。
  • boundary:no preopened Host FS、workspace、Broker、runtime package install、dynamic native loading;actual enforcement 是 WASM/WASI capability absence + explicit Host-call + source/import admission + fresh Guest。instruction_level_dbi_claimed=false
  • 状态Current bounded profile / Historical artifact;target-bound ;canonical structural-recompute ;raw/artifact replay 。支持 profile-bound qualification,不支持 generic NumPy、native correctness outside profile、DBI 或 arbitrary Python。

7.2 Native sandbox lifecycle 与 natural placement control

  • 目标 source:runtime/engine/native/backend.goBackend.ExecuteWithEvidencevalidateConfigimageIdentityBoundEvidence.Lifecyclerootfs.go:VerifyOCIImageConfig/RootFSIdentitybackend_evidence_test.go 的 image identity、fresh physical identifier、resource projection。它绑定 image config/artifact/decision identity,限制 memory/pids/output,执行后要求 delete reconciliation、cgroup/root unmount、scratch/workspace lease release。
  • docs/evidence/natural-placement-open-swe-v1.json:Open-SWE revision ad4805a5aa7de70d99cab0bb8f99b15304c76de0,source digest sha256:4b91b39f54849bac8323a468b67fbed6065535dba004731c99ea21cb6345de1e,one frozen Python trajectory (95 messages, record SHA sha256:607d9fb4f9ebfe90ba654f85a5d012ba1ef3f845cf2d12ed53c022686d795a4d);required features shell/subprocess,selected backend native_sandboxnative_backend_calls=1、Pysolate guest 0、effects/workspace started before placement false,promotion false。
  • 该 control 只支持“one frozen trajectory conservatively placed before Guest start”;禁止 coding success、native backend correctness、model success、general placement optimality。它是 Observed/Framing,dataset producer historical,target-bound ,target-tree raw body ,只能做 body-safe structural check。

trade-off:native sandbox 提供 shell/subprocess 等 Pysolate Guest 不具备的 feature,但要求更强的 image/rootfs/resource lifecycle evidence;不能把“选择 native”说成自动透明 fallback,也不能把 numpy-core static native profile 与 native sandbox 当同一 backend。


8. Natural cohort、证据缺口与 no-go

8.1 Natural corpus importer / probe

问题:公共 agent dataset 是否能在不收集私有 corpus、不调用 LLM labeler 的情况下提供可审计的 real Python slice?

路径scripts/natural-corpus.py 的 bounded AST/import screening、manifest identity recomputation、opportunity census;scripts/tests/test_natural_corpus.pyintegration/e2e/natural_placement_canary_test.go;body-safe manifest 只保留 dataset/source-record/action/trajectory digest、bytes、oracle/privacy/authority/backend class。

artifact / producer / checker / raw sample平台、workload、样本量核心数字与独立复核支持与禁止推论状态
docs/evidence/natural-corpus-pilot-v1.jsonnatural-corpus-manifest-v1.json;CodeAct revision afba34367a8609a1d0044eded531548ab71a58cf,Open-SWE revision ad4805a5aa7de70d99cab0bb8f99b15304c76de0;manifest identity sha256:8ffde0e8882097320e61a0ec8606c1e2c9ee60d71c357f143ce546b720c65dcf;checker scripts/natural-corpus.py + scripts/tests/test_natural_corpus.pypublic mixed cohort:50 CodeAct records + 10 Open-SWE trajectories;manifest denominator 147 = 137 CodeAct actions + 10 trajectories;body/private raw remains ~/.hermes/evidence/pysolate/<private raw>,不在 target treeCodeAct 137 actions、20 static compatible、99 environment-dependent、12 third-party、6 boundary-crossing;8 deterministic no-import real top-level actions smoke;baseline 8/8、Guest 8/8、matched completion 8/8。manifest states 22 included/125 rejected/0 unclassifiable/0 truncated,probe 8 completed/139 not run。支持:body-safe importer 找到 8 个可在 named Host-profile-bound Guest 完成的 actions。禁止:original task correctness/output equivalence、latency、sharing frequency、production readiness、Open-SWE repository replay;single-run timings 只是 smoke。Observed bounded pilot / Current contract; public canonical counts可重算;raw body/artifact replay ;target-bound
docs/evidence/natural-corpus-opportunity-v1.json;checker scripts/natural-corpus.py opportunity validator;canonical identity sha256:dd1621c88dc07e6721711aab9f6599e9f54cac65a52178ab1f85f4d3eea6ced4CodeAct 137 actions,Open-SWE 400 execute_bash calls;duplicates 保留在 denominatorCodeAct exact unique 130、duplicate instances 7、cross-record duplicate groups 0、sequential duplicate records 4;Open-SWE unique 338、duplicate instances 62、cross-trajectory groups 1、parallel exact duplicate 0;authority/workspace equivalence not_recorded、overlap not_observed;verdict insufficient_evidence / do_not_implement_sharing_pass支持“当前 corpus 没有足够 Host-owned equivalence + overlap 证据来开 sharing pass”。禁止:sequential retry=coalescing、one repeated command=physical sharing opportunity、absence outside cohort。Observed no-go / Framing; target-bound ;canonical-recompute ;raw-replay
docs/evidence/natural-corpus-probe-v1.json:artifact sha256:664077c1d63445ec267b1b30e30ce31c72e7038d62a08fe1682c675a64cff257;runner sha256:edff71e7dc127d3db25bca7f240fa215f61c2c643e94abc05ea89e0b72836d8a;source commit 31b58661e26b8a1f826af5f2bc99cf7918ac32b48 selected CodeAct actions,baseline/Guest one run each8/8 completion only;每行 oracle class completion_only;timing non-comparable,source/import profile-bound。只支持 bounded compatibility smoke;禁止 original task correctness、performance、model quality。Observed smoke / Historical; target-bound ;canonical structural check ;raw-replay

8.2 tau2 remediation 与 source-bound evidence gap

docs/evidence/tau2-t2-remediation-v1.json(parent prereg sha256:64a66141c1a3382743e7867c087c6062d2222165698c1f616471eb272cddf082;treatment prereg sha256:0185482fae6882056630bc375a9878ef5e4882a30577b85c91b57b5044e6af4c;source tau2 revision c3398666e6559e3a063da3fc04b5acf7f941464e;checker scripts/tau2-t2-remediation-report.pyscripts/tau2-t2-oracle.py):

  • 16/16 planned/recorded cells completed,post_hoc_dropped=015 reconstructed,1 duplicate receipt identity partially reconstructed;source_joins=35、unique receipt-bound physical calls 35、known provider calls 241、unknown provider cells 0;private bodies excluded。
  • 支持:remediation treatment 的 source/receipt/oracle join 结果,以及 duplicate identity 被 downgrade;禁止 leaderboard、matched-surface performance、model WRITE ability、production external effects。
  • parent tau2-t2-cohort-v1.json 则是 32 cells,23 completed/1 not_recorded/8 unscorable,shared raw path collisions 12;它明确说明只能把 reconstructed rows 当 source-bound treatment evidence,不能把 collided rows 当 causal evidence。
  • tau2-t2-runtime-replay-pilot-v1.json 只有一个 validation anchor,classification HARNESS_VALIDATION_ONLY_NOT_PERFORMANCE_EVIDENCE;不得升级成 runtime replay campaign。

8.3 Effect-aware census 与 natural source-prefix no-go 的共同教训

docs/evidence/effect-aware-opportunity-census.json / docs/research/effect-aware-opportunity-census.md 是 19 个 checked-in program 的 structural/overlay census:10 opaque/unknown、9 no-barrier、2 whole-Run reusable、16 WASM placement、3 native placement、11 structural exact pre-dispatch annotations 只收窄为 4 exact calls、其中 1 necessarily reached;CanPreissue legal 1、rejected 3。它证明的是 fail-closed narrowing,不是 population claim。

与 source-prefix natural census 的 36/36 final-or-only region、timing_not_recorded=36 一样,这些 no-go 都保留 denominator 和 absent timing,而不是补 synthetic samples。zero-provider/zero-raw/body-free 是 evidence limitation 的一部分,不是 performance zero。

paper placement:natural corpus denominator、opportunity gate、tau2 duplicate/source-join gap、semantic/source-prefix no-go 放 appendix 与 limitations;主文只保留“当前 natural evidence 不足以授权 sharing/overlap pass”的一句边界。


9. 验证结果、不可重算项与论文证据优先级

9.1 已做的只读验证

在 target archive/clean target worktree 上完成:

  1. git cat-file -tgit rev-parse:commit/tree 精确匹配;git verify-commit 通过。
  2. git show --stat:b94 只修改 docs/evidence/prepared-data-phase3-v2.json 与 prepared-data plan;没有 runtime/benchmark source 改动可让旧数字自动升级为 target benchmark。
  3. Python JSON decode:target docs/evidence/*.json 全部可解析;按 schema/字段读取关键 catalog files。
  4. 独立 canonical recompute:
    • semantic pre-dispatch 两份 5+5 rows:3214550/2196151/1018399 µs3149103/2132259/1016844 µs
    • source-prefix 两份 3+3 rows:v1 2950.046792/1533.922917 ms,day-trip 2942.787333/1537.471584 ms
    • authority-transparent 5+5 runs:19/17 physical median、每对 reduction 2、wall median 25728.820333/22004.745333 ms
    • NumPy reuse Phase 7 40 economics rows:四个 platform/profile median ratio 与 best ratio、40/40 negative break-even;
    • prepared Phase 6:6×3×3=54 coordinate、162 record count;仅验证结构/count,不把 aggregate median 重算成 observed trace。
  5. 未执行任何 campaign、provider、CI、full benchmark 或 external dataset replay。

9.2 目标 tree 中不能独立重算/重放的项目

以下项目只在 canonical JSON 中保留 digest、summary 或 derived aggregate;raw body/JSONL、Guest artifact、private workspace 或运行环境不在目标 tree,因此要明确写“不可独立重算/重放”:

  • prepared-data Phase 4–6 的 raw report/JSONL 与 phase6 per-event interval;phase6 median critical path 只能视为 projected schedule。
  • semantic Phase 4 的 360 原始 trial records、Phase 5r external private artifact/cost raw;只能校验 canonical gate/count,不能重建 raw median 或 Guest execution。
  • NumPy reuse Phase 7 的 macOS/Linux JSONL;canonical economics rows 可重算 ratio/negative break-even,但不能 raw-replay consumer lifecycle。
  • numpy-core Phase 2 的 RESULT.READY、manifest、WASM artifact 和 import inventory/qualification bundle;脚本 validator 需要外部 artifact root。
  • natural CodeAct/Open-SWE/tau2 原始 bodies、conversation/tool output、provider traces、workspace roots;body-safe digest/receipt rows不等于可 replay。
  • authority-transparent walkthrough 的完整 Guest/Host raw process trace;10 summary runs 和 148 canonical events可做有限 recompute,不可重放 arbitrary case。

9.3 主文/appendix 建议与优先证据清单

主文优先(最多三条核心链)

  1. prepared-data-phase3-v2 positive + branch/exception negative:一张 authority/data/claim lifecycle 图,明确 Host physical 与 Guest logical join;同时标注 producer historical、非 target certification。
  2. source-prefix-overlap-v1 authored 3-pair observed intervals:展示 one early read + source tail 的 overlap;旁边放 natural census 0/36 作为不外推的 boundary。
  3. semantic-predispatch G1/v0 exact live-only shape:展示 unchanged Python boundary、single claim、no duplicate physical call;把 timing 作为 bounded fixture,不写成 production latency。

Appendix 必须保留

  • authority-transparent 20-program matrix、P01–P20 rejection/control table;
  • NumPy reuse Phase 7 全 40 economic rows、四 platform/profile aggregates、0/40 break-even、raw JSONL digests;
  • prepared Phase 6 54 projected coordinates 与 overlap(prep,gap) 公式,明确“非 observed interval”;
  • semantic Phase 4 120-cell/360-record aggregate、Phase 5 parent no-go、Phase5r 11-case mechanism pass + zero economics records;
  • native profile 19 modules/2 archives/2 fresh runs 和 native sandbox lifecycle invariants;
  • natural 147-item manifest、8/8 smoke、137/400 opportunity census、tau2 16-cell remediation 与 duplicate/raw-path gap;
  • 所有 source commit/tree/artifact/harness/checker/raw digest 与 target-bound 结论。

最优先补强的未来证据(不是本次执行)

  1. 用 b94/6b 构建并 hash-bind target-owned Guest artifact、harness、checker、manifest,再做一次 preregistered complete campaign;旧 producer 不得替代。
  2. prepared-data 要记录 per-run read start/end → decode → source finalization → seal → exact claim → retire observed intervals,并把 raw rows 与 artifact 一起 append-only 保存。
  3. semantic pre-dispatch 只在 G1 exact live-only subset 上扩展独立 matched campaign;derived-AST 需新的 preregistration,不能从 Phase5r mechanism pass 直接晋级。
  4. natural cohort 先找有 trailing source regions、实际 overlap 和 Host-owned equivalence 的新冻结 corpus;zero-opportunity cohort 不应被 trace replay 强行补成 timing study。
  5. native/backend 若要主张通用性,必须新增明确 ABI/profile/negative corpus;numpy-corenative_sandbox 不可合并成 generic native claim。

10. 防御式一句话结论

在精确 target b94c989… / 6b8bb6… 中,Pysolate 的 authority/fresh-Guest、exact object claim、窄 semantic pre-dispatch、source-prefix authored overlap、typed NumPy/native boundary 和 natural no-go 都有可追溯的 checked-in contract/evidence surface;但 benchmark 数字的 producer 全部是历史 commit,prepared Phase 6 是 projected schedule 而非 observed interval trace,多个 raw sample/artifact 只以 digest 或 private path 存在。因此本 target 可以作为evidence catalog 与 claim-boundary review,不能作为旧 benchmark producer 的 target certification、自然 workload performance 证明或通用 Python/native 优化定理。