Plugin Value Lab v1.0.0-rc7

Plugin Value Lab v1.0.0-rc7

本次候选发布整理了首个外部实验室 L1 试用所需的现有路径。版本标识：Git tag / Release v1.0.0-rc7，插件 1.0.0-rc7，Python 包 1.0.0rc7。

本次变化

• pvl screen 接收插件清单与 3–5 个任务，冻结方案、导入配对记录并生成使用卡；结果为停止、证据不足或进入先导。缺失、失败、未知费用和安全早停均保留。
• 实验室在冻结方案中选择模型、逐模型费率和预算，使用自己的账户或明确配置的本地/内网端点。$0.10 为默认单次预算，可由实验室明确设定；不会改变任何外部账户或 AgentMuxer 的授权上限。
• 机构内数据治理说明明确原始数据、受试者数据与未发表材料的存储边界；只有经过人工批准的最小汇总和完成证明可导出。软件本身不是网络隔离或自动脱敏工具。
• 首个实验室试用包提供最短离线安装步骤、4 个任务模板、治理说明、完成验收清单，以及不调用模型的安装演练。真实 L1 使用机构内已有宿主执行并保留原始证据。
• 先前已实现的本地协作闭环纳入本次发布：保留记录复算、绑定结果审阅、版本纠正、双向交接和检查点恢复。MCP 写入入口需显式开启。

下载与选择

• 首次 L1 试用：PVL-first-lab-L1-v1.0.0-rc7.zip，解压后阅读 README.md，先运行 check_package.py。
• Codex/Claude 本地市场：plugin-value-lab-marketplace-1.0.0-rc7.zip，选择包含市场清单的解压根目录。
• 通用插件：plugin-value-lab-agent-plugins-1.0.0-rc7.zip。
• Python CLI/库：plugin_value_lab-1.0.0rc7-py3-none-any.whl；源码分发：plugin_value_lab-1.0.0rc7.tar.gz。
• SHA256SUMS.txt 用于核对附件；release-verification.json 记录本次实际工程检查与证据边界。

证据边界与下一步

本版是 prerelease。正式 v1.0 门槛为 1 个非作者外部实验室完成完整 L1，当前为 0/1。停止、证据不足和进入先导均可成为完整 L1 的有效结果；不完整尝试和作者合成演练不计入。工程测试、包校验与 CI 不证明实际收益、真实宿主验收或科学有效性。

功能继续冻结，仅处理既有路径的阻塞问题和试用支持。本次发布不包含向实验室发送邀请或上传其原始数据。

English

This release candidate packages one-command L0/L1 screening, lab-owned model/rate/budget configuration, institutional data governance, and a first-lab L1 trial kit with four task templates and an offline installation rehearsal. It also includes the previously implemented opt-in local collaboration workflow.

Use the first-lab ZIP for the guided trial, the marketplace ZIP for plugin installation, or the wheel for the CLI/library. Verify assets with SHA256SUMS.txt; actual engineering checks are recorded in release-verification.json. External L1 completion remains 0/1. Engineering checks and synthetic rehearsals do not establish real-world benefit or native-host acceptance. Feature development remains frozen until a separate owner decision.

Plugin Value Lab v1.0.0rc5

Plugin Value Lab v1.0.0rc5

Release candidate for the latest plugin evaluation, decision, and distribution changes.
The Git tag is v1.0.0rc5; plugin manifests use 1.0.0-rc.5 and Python packages use 1.0.0rc5.

Changes

• Add the scientific balance icon to the plugin listing and composer, and include it in marketplace and portable packages.
• Respect the frozen power plan when judging independent task-family coverage. Repeated runs cannot bypass the family requirement; underpowered positive evidence remains a trial-only signal.
• Add bounded action recommendations and 3–32 candidate comparisons with explicit researcher losses, scope, expiry, alternatives, and next evidence. Shrinkage displays cannot bypass independent confirmation gates.
• Add a literature task-family intake and custody-signature workflow, with 30 synthetic candidate placeholders. These are not expert-validated real task families.
• Add request-budget reservations to the parent model gateway, retaining unresolved reservations and rejecting requests beyond the allowed budget before dispatch.
• Render structured Markdown values as readable tables and keep HTML escaping at the HTML layer.
• Consolidate current guidance into the README and retain previous guides and evidence under the history index. Repair installation links in generated marketplace packages.
• Separate fast and full test tiers; release validation rejects skipped tests.

Install

In Add plugin marketplace, use source HERRY423/Plugin-Value-Lab, Git reference v1.0.0rc5, and leave sparse paths empty. Use main instead to follow subsequent repository changes.

For a local installation, download plugin-value-lab-marketplace-1.0.0-rc.5.zip, extract the complete archive, and select its marketplace root folder. The portable Agent Plugins archive and Python wheel are separate installation options. Verify downloads with SHA256SUMS.txt.

Evidence and limits

Validation outcomes and package hashes are recorded in the attached release-evidence.json. This is a prerelease, not general-availability or native-host certification.

The evidence ladder remains a design proposal. Expert review, independent holdout custody, real plugin benefit, and clinical/scientific validity are not established by the bundled synthetic cases or engineering tests. The gateway reservation mechanism does not establish a provider-side spending limit; the currently authorized $0.10 request ceiling still blocks the configured model's conservative cost envelope.

Release validation

• Full local suite: 1,605 tests passed, zero failures, errors or skips.
• Isolated wheel installation: all 14 checks passed, including CLI and notebook execution.
• GitHub CI: all 12 jobs passed for the release commit, covering Linux/Windows and Python 3.11/3.13 test matrices and the separate acceptance jobs.
• Marketplace and portable packages match the source inventory; icon paths and installation links resolve. Uploaded package downloads matched local SHA-256 checksums.
• Release commit: b9744f727edf7c806fe34a53e321985c2f743913.

Plugin Value Lab 1.0.0-rc.1

Plugin Value Lab 1.0.0-rc.1 Candidate Release

概要 / Overview

Plugin Value Lab（PVL）发布首个正式候选版本 v1.0.0-rc.1。本候选版本冻结了候选提交代码与全套安装分发包，通过了离线全量测试（1567 项测试无跳过、842 项子测试通过）、功能范围冻结检查、社区策略校验、插件规范验证以及全新隔离虚拟环境下的端到端工程可用性验收。

此外，本版本正式确立并规范了科学对照评测的四大严谨性支柱（Four Rigor Pillars）。

---

四大严谨性支柱 / Four Rigor Pillars

1. 真实两臂隔离（True Two-Arm Isolation）：

   • 每次评估尝试严格绑定全新独立会话（fresh_session: true，prior_messages: 0，parent_session_id: null，memory_enabled: false，shared_artifacts: false）。
   • 会话与请求 ID（session_id, request_id）全局去重，严防跨尝试状态泄漏。
   • WITH 臂严格仅暴露指定候选插件及声明的公共基础工具；WITHOUT 臂严格零插件暴露（plugins: []）。
   • 运行上下文密码学摘要（context_sha256）与事前冻结方案完全一致。

2. 完整工具清单（Complete Effective Tool Inventory）：

   • 清单凭证必须源自宿主真实生效列表：source_kind: "host_effective_inventory"，显式声明 inventory_complete: true。
   • 拒绝未消费的分页或截断（has_more: false, truncated: false, next_cursor: null）。
   • 每个暴露的工具均计算并比对规范 Schema 定义的 SHA-256 摘要（definition_sha256），严禁任何隐形工具注入或未声明的功能扩展。

3. 原始产物封存与重新评分（Raw Artifact Preservation & Local Re-scoring）：

   • 封存大模型输出的 UTF-8 原始原生文本（raw_output），禁止在评分前进行中间润色或摘要损耗。
   • 强制绑定哈希指纹：output_sha256 = sha256(raw_output.encode("utf-8"))。
   • 评分器在本地基于客观规则（正则、AST 语法树、JSON Schema、科学校验器）对原始产物重新评分，零信任外部宿主上报标记或模型自称的“任务已完成”声明。

4. 断线／取消／重试的原尝试对账（Attempt Journal Reconciliation）：

   • 基于不可篡改触发器的 SQLite 本地持久化日志（AttemptJournal），覆盖 prepared → dispatched → running → 终态全生命周期。
   • 网络断线或宿主失联（unknown）必须通过外部凭证（source_ref）向原始尝试进行回溯对账。
   • 分母守恒原则：取消（cancelled）与失败（failed）尝试一律保留在有效样本量分母中，杜绝幸存者偏差。
   • 首次尝试首要性：诊断重试必须显式绑定 retry_of = primary_attempt_id，主要效应分析永久以首次尝试为准，重试结果不得篡改或掩盖首次失败。
   • 全量结算对账：所有发出的尝试（含失败、重试、取消）产生的 Token 消耗和供应商成本必须全额纳入累计结算账本。

---

冻结安装包资产 / Frozen Distribution Packages

文件名 / Asset	文件大小 / Bytes	SHA-256 校验和 / Digest
plugin_value_lab-1.0.0rc1-py3-none-any.whl	587,926	95423b27370889c47e5bb5c340a5c179649f4d4b42ad2168c7fc049a7c826bf4
plugin-value-lab-marketplace-1.0.0-rc.1.zip	2,048,301	da7651c7556dbb4ea68efc86a4ccd5bc8ccfbd07d7015fc4e048cb12ef8c2825
plugin-value-lab-agent-plugins-1.0.0-rc.1.zip	1,997,259	b938d13b7d285284696eaad96604b567d4c03421fc6c067824c0706d24d62559
release-evidence.json	1,267	411ef15f09c5e409e26d4e94c7b9e5da5429a56962ca808ef1cd1a1e548742f0
SHA256SUMS.txt	419	864e9f3ebf3bae84e1de0d1716af96dd6b448727db60d76a13d97ca297cf08dc

快速安装与离线校验 / Installation & Verification

# 校验全部资产
sha256sum -c SHA256SUMS.txt

# 在独立虚拟环境中安装 Wheel
python -m pip install plugin_value_lab-1.0.0rc1-py3-none-any.whl

# 验证 CLI
pvl --version
# 输出: 1.0.0-rc.1

边界声明 / Limitations：本候选发布仅代表工程代码、测试门禁与打包资产的自洽验收；不代表已获得第三方独立机构采用，亦不代表已取得任何特定插件普遍优越性的外部科学定论。

Plugin Value Lab 0.9.0

Plugin Value Lab 0.9.0 — minimum effective combinations, burden view, and native context P0 telemetry

Version 0.9.0 introduces bounded multi-plugin combination evaluation, independent burden view (context pressure, cost, time), progressive evidence disclosure, native Codex host context telemetry capture, and local evidence dependency rechecking.

Key Changes

• Plugin Combinations: Evaluate bounded plugin pairs (baseline, A, B, A→B, optional B→A) under matched tasks to determine the minimum sufficient plugin set without automatic install/uninstall.
• Decision-Time Evidence View: Lead combination results with actionable issues, tradeoffs, and missing records; export interactive offline EVIDENCE.html with progressive disclosure.
• Independent Burden View & Architecture Decoupling: Dedicated view comparing context window pressure, monetary costs, and execution duration among feasible candidates, with clean separation between collector, analyzer, presentation layers, and immutable contracts.
• Native Host Telemetry (Codex Desktop 0.159.2 P0): Read-only adapter and standalone collector binding responses, models, sessions, and events with exact-prefix replay and unmetered unknown preservation.
• Local Evidence Dependencies: Conservative dependency graphs and selective invalidation rechecks beside immutable usage cards, distinguishing historical artifact re-scoring from current execution requirements.

Installation

Use GitHub tag �0.9.0 to pin this release. Choose the marketplace ZIP for the Codex/Claude-compatible repository layout, the Agent Plugins ZIP for the portable standard manifest, or the Python wheel:

\\sh
python -m pip install "plugin_value_lab-0.9.0-py3-none-any.whl[mcp,science,pseudobulk]"
\\

Verify downloads using the attached \SHA256SUMS.txt\ and
elease-evidence.json.

Maintenance commitment and paid-execution boundary / 维护承诺与付费边界

Effective 2026-10-03 (America/Los_Angeles). This is a dated policy addendum for v0.9.0; its original tag and five installation/evidence assets are unchanged.

• Maintainer: @HERRY423; no appointed backup. The latest non-prerelease minor line at its newest patch is maintained, currently v0.9.0. Ownership and review obligations · compatibility scope.
• Serious issues: P0/P1 acknowledgement within 1/2 working days; assessment, mitigation or stop-use advice within 2/5 working days after acknowledgement; updates daily/every 5 working days respectively. Working days are Monday-Friday in America/Los_Angeles. No guaranteed patch date or 24/7 SLA. Handling and escalation.
• Private vulnerabilities: Report privately (GitHub sign-in required). Rechecked enabled = true. No test report was submitted; notification delivery and human response remain untested. Original channel check. Its earlier publication flag remains false as a historical observation; this addendum is the later publication.
• 付费边界： 付费原生执行仍以 BUDGET_BOUNDARY_UNAVAILABLE 阻断。后续源码的 request_budget 仍为 provider_cap_verified: false / paid_launch_authorized: false。本地记账、预算估算、用户授权均不等于供应商强制限额；断线、取消、重试先对账原尝试，未结算费用保持 UNKNOWN，不得用新 key 或新任务重置额度。模型、评审、工具及收费轮询都必须纳入经核验的单次上限和共享任务预算，付费授权与数据发送授权分别确认。
• Artifact boundary: the original v0.9.0 tag points to b0ae821d7a7d9a5e2fc962c234206c38e80a0032 and already blocks paid native launches. request_budget is later source at b4f1596968c54e9039c7c3996009a807d48bc1bb, not part of the original v0.9.0 installation assets. This documentation publication does not certify paid execution, live Codex acceptance or scientific benefit.

Read the full maintenance and paid-execution policy or download maintenance-and-paid-boundary-v0.9.0-20261003.md below. Policy commit: e9e0200. Local validation: community metadata and feature-freeze checks passed; 6 community tests and 13 request-budget tests passed. No paid directory offering or model evaluation was invoked for this publication; unrelated account/session costs are not asserted to be zero.

Plugin Value Lab 0.8.0

Plugin Value Lab 0.8.0 — data scale execution, trace diagnostics, and scientific engineering usability

Version 0.8.0 introduces large-scale data execution pipelines, trace diagnostics review, preregistration protocols, and end-to-end workflow integrations.

Changes

• Data Scale Execution: scalable science runtime, storage limit controls, and backed pseudobulk designs for high-throughput single-cell datasets.
• Trace Diagnostics & Review: interactive trace diagnosis, diagnostic review workflows, and blind review protocols with claim-scope boundary verifications.
• Workflow & Kernel Acceptance: Nextflow, Snakemake, and Jupyter notebook workflow integration examples, along with automated kernel acceptance checks.
• SDK & Usability: unified easy CLI and Python SDK for programmatic plugin value evaluation, component localization, and evidence ledger generation.

Install

Use GitHub tag v0.8.0 to pin this release. Choose the marketplace ZIP for the Codex/Claude-compatible repository layout, the Agent Plugins ZIP for the portable standard manifest, or the Python wheel.

python -m pip install "plugin_value_lab-0.8.0-py3-none-any.whl[mcp,science,pseudobulk]"

Verify downloads using the attached SHA256SUMS.txt.

Plugin Value Lab 0.7.0

Plugin Value Lab 0.7.0 — detection boundaries validation and clean release

Version 0.7.0 introduces detection boundaries validation, frozen contract controls, and release artifact cleanup.

Changes

• Detection Boundaries: comprehensive validation across eight categories including reference leakage, label semantic inversion, over-refusal, silent fallback, pseudo-duplicates, ID misalignment, and trigger verification.
• Controlled Observation: newline transport normalization with hash retention, direct use normal control retention, and access attempt monitoring in reviewed Python runners.
• Clean Packaging: removed temporary test fixtures from distribution, synchronized Agent Plugins and marketplace manifests, and updated all manifests to version 0.7.0.

Install

Use GitHub tag v0.7.0 to pin this release. Choose the marketplace ZIP for the Codex/Claude-compatible repository layout, the Agent Plugins ZIP for the portable standard manifest, or the Python wheel.

python -m pip install "plugin_value_lab-0.7.0-py3-none-any.whl[mcp,science]"

Verify downloads using the attached SHA256SUMS.txt.

Plugin Value Lab 0.6.0

Plugin Value Lab 0.6.0 — native workflows and reusable scientific evidence

Version 0.6.0 brings the latest native execution, scientific artifact checks, repair comparisons and evidence handoffs into one versioned release.

Changes

• Native workflow: reuse frozen official cases, probe Linux/WSL isolation and executable startup, authorize a single host attempt, retain failures and partial results, and recover collection without relaunching the model.
• Analysis and repair: bind foreground script calls to retained workspaces and frozen inputs; recompute scientific artifacts offline; compare complete before/after repair packages while retaining regressions and unresolved observations.
• Falsifiable diagnosis: connect six diagnostic stages to observed evidence and explicit predictions/falsifiers. Explicit-invocation probes remain separate from natural-use value evidence.
• Scientific evaluation: add prospectively frozen estimands, explicit failure classes, planned denominators and descriptive uncertainty; add donor-aware raw-count aggregation, a pinned PyDESeq2 backend and controlled error/valid-variation checks.
• Reuse and guidance: add registry-free pinned handoffs, separate scorer delivery, non-executing offline replay, scoped host/model/plugin/replicate comparisons and conditional guidance under frozen quality, cost and failure limits.
• Windows compatibility: normalize short directory names when binding retained native traces and workspaces; reject traversal and linked artifacts.
• Packaging: synchronize Python, runtime and all plugin manifests at 0.6.0. The marketplace and portable Agent Plugins packages include the latest documentation, examples and tests. Research/team extensions remain opt-in.

Install

Use GitHub tag v0.6.0 to pin this release. Choose the marketplace ZIP for the Codex/Claude-compatible repository layout, the Agent Plugins ZIP for the portable standard manifest, or the Python wheel. Extract ZIP files before loading them.

Python 3.11+ is required. After downloading the wheel:

python -m pip install "plugin_value_lab-0.6.0-py3-none-any.whl[mcp,science]"

Add the pseudobulk extra when running the optional PyDESeq2 reference backend. This GitHub release does not imply PyPI publication. Verify downloads using the attached SHA256SUMS.txt.

Validation and evidence boundaries

Local Windows / Python 3.13 acceptance: 518 tests passed, zero failures, errors or skips. Manifest validation and byte-identical marketplace checks passed. Current release acceptance is recorded in VERIFICATION.json; dated P0/P1/P2 and native-workflow receipts preserve their original versions and observations. The release packages contain public examples and evidence summaries, not private local studies or account configuration.

Prior native canaries and cached public-data computations are bounded observations. Missing Skill calls, uncertain sealing, failures and unknown settled costs remain visible. Offline replay creates no new independent observations. External independent reuse, expert scientific validation, cross-host benefit and researcher time savings remain unestablished. Software publication is not scientific certification or a claim of general availability.

Hosted CI: all four Windows/Linux and Python 3.11/3.13 jobs passed for commit 94ae305b716e8934fd9e0e0ffe9d88fe77aff386. View exact-commit CI.

Plugin Value Lab 0.5.0

Plugin Value Lab 0.5.0 — executable scientific evidence

Scientific artifacts can contain plausible numbers while using the wrong independent unit, omitting tests or reporting an unsupported conclusion. Version 0.5.0 adds a bounded replicate_effect verifier that recomputes scientific results from frozen independent-unit measurements.

• Independent units and design: reject repeated donors in independent designs, incomplete pairs, completely confounded batches and declared training/evaluation overlap. Permutations preserve blocks or pairs.
• Recomputed outcomes: compare treatment-control effects, exact two-sided p values, complete-family BH values, sample counts, frozen-threshold decisions and leave-one-unit effect ranges. Correct non-significant findings are valid outcomes.
• Portable evidence: design/data commitments join replay dependency preflight. Bundle identity, environment checks and field-level differences distinguish recomputation from complete or scientifically valid evidence.
• Cross-host and methodology: unified Codex/Claude archive verification; requested settings are not observations; process checks do not add outcome points; unsupported acceptance and over-refusal have separate limits; unknown costs remain unknown.
• Value measures: success uplift, rescued and harmed pairs, family-level gains, time and cost per successful task; prospective task-family sample planning remains distinct from a confirmatory test.
• Release consistency: Python/runtime/plugin versions are all 0.5.0; Git preserves frozen bytes across operating systems. Research/team extensions remain opt-in. Epistemic Plugin Arena remains future outlook.

Install

Choose the marketplace ZIP for the Codex/Claude-compatible repository layout, the Agent Plugins ZIP for the portable standard manifest, or the Python wheel. Extract ZIP files before loading them. Python 3.11+ is required; install the mcp extra for MCP and science for h5ad/SciPy support. Core deterministic verification uses the standard library.

python -m pip install "plugin_value_lab-0.5.0-py3-none-any.whl[mcp,science]"

This command installs a downloaded wheel; this release does not imply PyPI publication. See attached checksums and the repository installation guide.

Evidence boundaries

Tests include adversarial fixtures and independent SciPy numerical checks. Public examples are synthetic teaching material, not a hidden benchmark. The verifier accepts already prepared unit-level measurements; it does not validate normalization, raw-count models, biological identity, exchangeability or hidden training exposure. Exact enumeration has fixed resource limits and no silent approximate fallback.

No new real model trial, cross-host benefit study, external expert review or biological validation is claimed. The historical native pilot remains evidence-limited and is preserved separately in VERIFICATION.json. Computational agreement does not establish scientific truth or plugin efficacy.

Release verification

403 tests passed with no skips in each Windows/Linux and Python 3.11/3.13 CI job. Published assets were built by the successful Ubuntu/Python 3.13 job, checked member-by-member against commit 95321ac20d83038796318bec4c369d6d2e333363, and the hosted wheel was independently installed and executed outside the source directory.

Exact-commit CI. See release-evidence.json and SHA256SUMS.txt for asset provenance and hashes.